MULTI-FLOOR INDOOR LOCATION ESTIMATION SYSTEM BASED ON
WIRELESS LOCAL AREA NETWORK
CHUA TIEN HAN
UNIVERSITI TEKNOLOGI MALAYSIA
iii
To my beloved parents and brothers.
iv
ACKNOWLEDGEMENT
I am greatly thankful to my supervisor, Professor Dr Tharek Bin Abdul
Rahman for his continuous guidance, supports and encouragements. Without his
patient, valuable suggestions and comments, I would not be working on this exciting
research work.
I also wish to thanks all my fellow friends at Wireless Communication
Centre, Universiti Teknologi Malaysia. Thanks for their continuous helps, supports,
cheers, friendship and nice working atmosphere they provided.
I gratefully acknowledge the kindness from Mr. Luke Klein-Berndt and Mr.
Camillo Gentile from National Institute of Standards and Technology, U.S. for
sharing their works with me.
Simply I could not have reached where I am today without my pa, ma and
brothers. It has been a long journey for me and they always give me love and
encouragement. To them I dedicate this work.
v
ABSTRACT
The proliferation of high speed wireless technologies and mobile computing
infrastructures has fostered a rapid development in location based services. The key
to the success of location based services is the estimation of user’s location. Indoor
location estimation system using various wireless technologies such as infrared and
ultrasound are available. However, these systems require specialized infrastructures
and incur high costs. This study focuses on design and development of a softwarebased multi-floor indoor location estimation system using Wireless Local Area
Network (WLAN). Location fingerprinting technique is employed to estimate
Mobile Terminal’s (MT) location. WLAN Received Signal Strength (RSS) measured
by MT is used as location fingerprint. Before location estimation, database of
location fingerprint is constructed by collecting histograms of RSS at predefined
reference locations. During location estimation, current histogram of RSS at
unknown location will be compared to the database. The most probable match is
selected and returned as estimated location based on Bayesian filtering algorithm.
Estimated location is reported as physical location and symbolic location. Before
developing the system, study on characteristics of RSS is conducted to help the
design, development and implementation of the proposed system. The proposed
system is then designed and developed using Java programming language. The
performance of the proposed system is evaluated in a two-floor building using offthe-shelf WLAN access points and client device. Finally, various factors which affect
the performance of the proposed system are investigated. From the evaluations in the
two-floor building, the proposed system achieved best accuracy of 4.56 meters
during stationary tests and 4.54 meters during mobile tests with 90% precision. The
best percentage of correct floor estimation is 100% for both tests.
vi
ABSTRAK
Perkembangan pesat teknologi wayarles berkelajuan tinggi dan infrastruktur
komputer bergerak telah menggalakkan pembangunan cepat dalam perkhidmatan
berdasarkan lokasi. Kunci kejayaan perkhidmatan berdasarkan lokasi ialah
penganggaran lokasi pengguna. Sistem penganggaran lokasi dalam bangunan yang
menggunakan pelbagai jenis teknologi wayarles seperti inframerah dan ultrabunyi
boleh didapati. Namun, sistem-sistem ini memerlukan infrastruktur khas dan
mendatangkan kos yang tinggi. Kajian ini memfokus pada rekabentuk dan
pembangunan perisian sistem penganggaran lokasi dalam bangunan bertingkat
dengan menggunakan Rangkaian Kawasan Setempat Wayarles (WLAN). Teknik
pencap-jarian lokasi diguna untuk menganggarkan lokasi Terminal Bergerak (MT).
Kekuatan Isyarat Diterima (RSS) WLAN yang diukur oleh MT digunakan sebagai
cap jari lokasi. Sebelum penganggaran lokasi, pangkalan data untuk cap jari lokasi
dibina dengan mengumpulkan histogram RSS di kawasan rujukan yang ditakrifkan
awal. Semasa penganggaran lokasi, histogram RSS semasa di lokasi yang tidak
diketahui dibanding dengan pangkalan data. Padanan yang paling hampir dipilih dan
dikembalikan sebagai lokasi anggaran berasaskan algoritma penapisan Bayesian.
Lokasi anggaran dilaporkan sebagai lokasi fizikal dan lokasi simbol. Sebelum
membangunkan sistem ini, kajian atas ciri-ciri RSS dijalankan untuk membantu
rekabentuk, pembangunan dan perlaksanaan sistem yang dicadangkan. Sistem yang
dicadangkan kemudiannya direkabentuk dan dibangunkan dengan menggunakan
bahasa pengaturcaraan Java. Perlaksanaan sistem yang dicadangkan dinilai dalam
bangunan dua tingkat dengan menggunakan titik capaian dan alat pelanggan WLAN.
Akhirnya, pelbagai faktor yang menpengaruhi perlaksanaan sistem yang dicadangkan
telah disiasat. Daripada penilaian dalam bangunan dua tingkat, sistem yang
dicadangkan mencapai kejituan terbaik 4.56 meter semasa ujian pegun dan 4.54
meter semasa ujian bergerak dengan kepersisan 90%. Peratusan terbaik
penganggaran tingkat yang betul adalah 100% untuk kedua-dua ujian.
vii
TABLE OF CONTENTS
CHAPTER
1
TITLE
PAGE
DECLARATION
ii
DEDICATION
iii
ACKNOWLEDGEMENT
iv
ABSTRACT
v
ABSTRAK
vi
TABLE OF CONTENTS
vii
LIST OF TABLES
xv
LIST OF FIGURES
xvii
LIST OF SYMBOLS
xxiv
LIST OF ABBREVIATIONS
xxvi
LIST OF APPENDICES
xxviii
INTRODUCTION
1.1
Background
1
1.2
WLAN-Based Multi-Floor Indoor Location
2
Estimation System
1.3
Problem Statement
3
1.4
Objectives
4
1.5
Research Scope
4
1.6
Contributions
5
1.7
Thesis Organization
6
viii
2
WIRELESS INDOOR LOCATION ESTIMATION
SYSTEM
2.1
Basic Architecture of Wireless Indoor
8
Location Estimation System
2.2
2.3
Location Sensor
9
2.2.1
Classification of Location Sensor
9
2.2.2
Wireless Sensing Technology
10
2.2.2.1
Infrared
10
2.2.2.2
Ultrasound
11
2.2.2.3
Radio Frequency
11
Location Estimation Technique
12
2.3.1
Proximity Sensing
12
2.3.2
Triangulation
13
2.3.2.1
Lateration
13
2.3.2.2
Angulation
14
2.3.3
2.4
Location Fingerprinting
Properties of Wireless Indoor Location
15
16
Estimation System
2.4.1
Network-Based System and
16
Terminal-Based System
2.4.2
Specialized Infrastructure and
17
Existing Infrastructure
2.4.3
Physical Location and Symbolic
17
Location
2.5
2.4.4
Accuracy and Precision
18
2.4.5
Scalability
18
2.4.6
Cost
19
Previous Researches on Wireless Indoor
19
Location Estimation System
2.6
Summary
21
ix
3
WLAN-BASED INDOOR LOCATION
ESTIMATION SYSTEM
3.1
Overview of IEEE 802.11 WLAN
22
3.1.1
23
Basic Components of IEEE802.11
WLAN
3.1.2
Topologies of IEEE 802.11 WLAN
23
3.1.3
Basic Operation of IEEE 802.11
25
WLAN
3.2
Overview of WLAN-Based Indoor Location
26
Estimation System
3.3
WLAN-Based Indoor Location Estimation
27
Techniques
3.3.1
WLAN-Based Proximity Sensing
27
3.3.2
WLAN-Based Triangulation
29
3.3.3
WLAN-Based Location
29
Fingerprinting
3.4
Overview of WLAN-Based RSS Location
30
Fingerprinting
3.4.1
Basic Operation
30
3.4.2
Reference Location
32
3.4.3
Location Fingerprint
33
3.4.3.1
Location Information
33
3.4.3.2
Fingerprint Information
34
3.4.4
Location Estimation Algorithm
36
3.4.4.1
36
Nearest Neighbor in Signal
Space
3.4.4.2
3.5
Probabilistic Algorithm
38
Bayesian Filtering for Location Estimation
40
3.5.1
Bayesian Filter
40
3.5.2
Implementation of Bayesian Filter
45
3.6
Related Researches
46
3.7
Summary
50
x
4
5
METHODOLOGY
4.1
Overview of the Proposed System
51
4.2
Design and Development Processes
54
4.3
Proposed System Specifications
56
4.4
Experimental Test Bed
57
4.5
Experimental Hardware
58
4.6
Experimental Software
60
4.7
Summary
62
CHARACTERISTICS OF WLAN RECEIVED
SIGNAL STRENGTH
5.1
Introduction
63
5.2
Location Dependency of WLAN RSS
65
5.3
WLAN RSS Distribution at a Stationary
67
Indoor Location
5.4
Statistical Representation of WLAN RSS
70
Distribution at a Stationary Indoor Location
5.5
Variation of Numbers of APs Detected at a
74
Stationary Indoor Location
6
5.6
Effects of User Proximity
77
5.7
Effects of User’s Orientation
79
5.8
Effects of Different Environments
82
5.9
Effects of Types of WLAN Client Device
87
5.10
Effects of Multi-Floor Environment
90
5.11
Summary
94
SYSTEM DESIGN AND DEVELOPMENT
6.1
Proposed Multi-Floor Indoor Location
96
Estimation Framework
6.1.1
Topological Map Representation
98
6.1.2
Environmental Model
100
xi
6.1.3
Bayesian Filtering
104
6.1.3.1
State Space Representation
106
6.1.3.2
Topological Markov
109
Localization
6.1.3.3
Perceptual Model
113
6.1.3.4
Motion Model
114
6.1.3.5
Proposed Location
117
Estimation Algorithm
6.1.4
Continuous Space Location
119
Estimation
6.1.5
6.2
6.1.4.1
Centre of Mass
119
6.1.4.2
Time Averaging
122
Logical Area Estimation
123
System Development
124
6.2.1
Software Architecture
124
6.2.2
System Interface Component
126
6.2.2.1
Main GUI
126
6.2.2.2
System Configuration
127
6.2.3
6.2.4
Data Collection Component
128
6.2.3.1 RSS Reader
130
6.2.3.2
133
Access Point Filter
Database Component
134
6.2.4.1
Histogram Builder
135
6.2.4.2
Histogram Reliability
138
Checker
6.2.4.3
6.2.5
Topological Map Builder
140
Location Estimation Component
142
6.2.5.1
142
Discrete-Space Location
Estimator
6.3
6.4
System Implementation
152
6.3.1
Off-Line Phase
152
6.3.2
On-Line Phase
152
Summary
157
xii
7
RESULTS AND DISCUSSION
7.1
Introduction
158
7.2
Performance Metrics
158
7.2.1
Accuracy and Precision
159
7.2.2
Percentage of Correct Logical Area
159
Estimation
7.2.3
Percentage of Correct Floor
160
Estimation
7.3
Types of Experiment
160
7.4
Stationary and Mobile Multi-Floor Indoor
161
Location Estimation
7.4.1
Experimental Setup
162
7.4.1.1
Test-bed
162
7.4.1.2
Off-line Phase
165
7.4.1.3
On-line Phase for
165
Stationary Multi-floor
Indoor Location Estimation
7.4.1.4
On-line Phase for Mobile
168
Multi-floor Indoor
Location Estimation
7.4.2
Experiment 1a: Stationary Multi-
169
floor Indoor Location Estimation
7.4.2.1
Accuracy and Precision
169
7.4.2.2
Percentage of Correct
170
Logical Area Estimation
and Percentage of Correct
Floor Estimation
7.4.2.3
7.4.3
Discussion
Experiment 1b: Mobile Multi-floor
171
173
Indoor Location Estimation
7.4.3.1
Accuracy and Precision
173
xiii
7.4.3.2
Percentage of Correct
174
Logical Area Estimation
and Percentage of Correct
Floor Estimation
7.4.3.3
7.5
Discussion
Factors Affecting the Performance of the
175
177
Proposed System
7.5.1
Experimental Setup
177
7.5.2
Experiment 2a: Effects of number of
180
access points
7.5.2.1
Accuracy
180
7.5.2.2
Percentage of Correct
182
Logical Area Estimation
7.5.2.3
Percentage of Correct
183
Floor Estimation
7.5.2.4
7.5.3
Discussion
Experiment 2b: Effects of number of
184
185
topological nodes
7.5.3.1
Accuracy
185
7.5.3.2
Percentage of Correct
186
Logical Area Estimation
7.5.3.3
Percentage of Correct
188
Floor Estimation
7.5.3.4
7.5.4
Discussion
Experiment 2c: Effects of number of
188
189
RSS samples per fingerprint
7.5.4.1
Accuracy
189
7.5.4.2
Percentage of Correct
190
Logical Area Estimation
7.5.4.3
7.5.5
Discussion
Experiment 2d: Effects of off-line
191
192
phase sampling interval
7.5.5.1
Accuracy
192
xiv
7.5.5.2
Percentage of Correct
193
Logical Area Estimation
7.5.5.3
7.6
8
Summary
Discussion
193
194
CONCLUSION
8.1
Conclusion
196
8.2
Future Works
198
REFERENCES
200
Appendix A
205
Appendix B
209
Appendix C
210
xv
LIST OF TABLES
TABLE NO.
TITLE
PAGE
3.1
Example of radio map
35
4.1
Specifications of the proposed system
56
4.2
Channel allocation for access points installed in
WCC, UTM
60
5.1
Objectives and types of studies on WLAN RSS
characteristics
64
5.2
Average RSS tuple from three access points at
different measurement locations
67
5.3
Summary of statistics for RSS distribution from
access point A7 at location L1 with different numbers
of samples
72
5.4
Numbers of RSS samples collected from access
points at location L1 over 60 RSS measurements
75
5.5
Average RSS at location L3 and L4 with forward and
backward orientations
80
5.6
Average and standard deviation of the RSS
distributions for access point A7, A11 and B3 at
location L1 during busy office and empty office
environments
85
5.7
List of IEEE 802.11b/g WLAN client device
88
5.8
Summary on WLAN RSS characteristics
94
6.1
Paths and tracks represented by the topological map
in Figure 6.3
107
6.2
Allowed and restricted state representations
109
xvi
7.1
Objectives and types of experiment carried out to
study the performance of the proposed system and
various factors affecting the performance
161
7.2
Performance of three location modes during
stationary multi-floor location estimation
172
7.3
Performance of three location modes during mobile
multi-floor location estimation
176
7.4
Access points used to investigate the effects of
number of access points on location estimation
performance
180
xvii
LIST OF FIGURES
FIGURE NO.
TITLE
PAGE
2.1
Basic architecture of wireless indoor location
estimation system
9
2.2
Concept of proximity sensing
12
2.3
Concept of triangulation via lateration
14
2.4
Concept of triangulation via angulation
15
2.5
Concept of location fingerprinting
16
3.1
Basic components of IEEE 802.11 WLAN
23
3.2
WLAN topologies. (a) Independent basic service
set (b) Basic service set (c) Extended service set
24
3.3
Passive scanning
25
3.4
Active scanning
26
3.5
Basic architecture of a WLAN-based indoor
location estimation system
27
3.6
Basic operation of WLAN-based RSS location
fingerprinting. (a) Off-line phase (b) On-line phase
31
3.7
Reference locations (a) Regular grid (b) Irregular
32
3.8
Nearest neighbor in signal space (NNSS) and kNNSS (k = 3)
38
3.9
One-dimensional illustration of Bayesian Filtering
technique
44
4.1
Basic architecture of the proposed system
51
4.2
Flow chart of off-line phase
52
xviii
4.3
Flow chart of on-line phase
3
4.4
Design and development process of the proposed
system
55
4.5
Floor 1 of WCC, UTM and the locations of the
access points
57
4.6
Floor 2 of WCC, UTM and the locations of the
access points
58
4.7
Hardware used in the proposed system (a) DLink
DWL-2000AP+ WLAN access point (b) HP
Compaq tc1100 Tablet PC
59
4.8
Snapshot of Netstumbler software
61
5.1
Average RSS from three access points at different
locations as MT traveled along hallway Path_1a
65
5.2
RSS measured from access point (a) A7 (b) A11 (c)
B3 for duration of 12 hours at location L1
68
5.3
Histograms of RSS distribution for access point (a)
A7 (b) A11 and (c) B3 measured for duration of 12
hours at location L1
71
5.4
Histograms of RSS distribution for access point A7
at location L1 calculated with (a) 30 (b) 50 (c) 100
(d) 200 (e) 300 RSS samples
73
5.5
Number of access point s detected by WLAN client
device at location L1 over 60 RSS measurements
75
5.6
RSSs measured from access point A7, A8, A10,
A12 and B3 at location L1 over 60 RSS
measurements
76
5.7
RSS Measured from access point (a) A6 (b) B3 (c)
B4 (d) A5 at location L2 with user absence in the
first two hours and presence for the following two
hours
78
5.8
Average RSS Measured from access point B3, B4,
A5 and A6 at location L2 with user absence in the
first two hours and presence for the following two
hours
79
xix
5.9
RSS measured from access point A7 at location L1
during (a) Busy office environment (b) Empty
office environment
83
5.10
RSS measured from access point A11 at location
L1 during (a) Busy office environment and (b)
Empty office environment
84
5.11
RSS measured from access point B3 at location L1
during (a) Busy office environment and (b) Empty
office environment
85
5.12
Average RSS measured from access point A3, A8
and B3 at location L5 with five different WLAN
client devices
89
5.13
RSS measured at different floors from access point
(a) A7 and (b) B4 at location L6 and L7
respectively
91
5.14
RSS measured at different floors from access point
(a) A4 and (b) B3 at location L8 and L9
respectively
92
5.15
Average and difference of RSS measured from
access point A7, B4, A4 and B3 at different floors
93
6.1
Proposed multi-floor indoor location estimation
framework
97
6.2
Topological map representation of a two-floor
indoor environment
99
6.3
Calibration nodes
101
6.4
An example of normalized RSS histogram collected
from one AP at one calibration node
102
6.5
Structure of the proposed environmental model
103
6.6
Structure of proposed state representation
108
6.7
Basic operation of topological Markov localization
111
6.8
Motion Model
115
6.9
State transitional probability
116
6.10
Procedure to recursively update the proposed
topological Markov localization algorithm
118
xx
6.11
Concept of centre of mass
120
6.12
Continuous space location estimation using centre
of mass technique
121
6.13
Representation of indoor area using logical area
123
6.14
Software architecture of the proposed system
125
6.15
The Main GUI of the proposed system
127
6.16
The Location Options submenu. (a) Discrete or
continuous space location estimation. (b) The time
averaging interval for time averaging technique
128
6.17
Flow chart of Data Collection Component
129
6.18
Flow chart of RSS reader module
131
6.19
Source code for retrieving RSS tuple using
WiFiSpotter API
132
6.20
Example of text file which defines the SSID of the
trusted WLANs
133
6.21
Flow chart of access point filter module
134
6.22
Flow chart of Database component
135
6.23
Flow chart of Histogram Builder Module
136
6.24
Flow chart of Histogram Reliability Checker
module
139
6.25
(a) Simple graph (b) Directed graph
140
6.26
Source Code for constructing topological map with
simple graph and directed graph using JGraphT
API
141
6.27(a)
Flow chart of the Discrete-Space Location
Estimator Module
143
6.27(b)
Sub-Function: Generate all Possible States
144
6.27(c)
Sub-Function: Generate all Allowed States
145
6.27(d)
Sub-Function: Generate all Possible State
Transitions
146
xxi
6.27(e)
Sub-Function: Initialize all state probabilities with
global probability
147
6.27(f)
Sub-Function: Calculate time-invariant state
transition probabilities
148
6.27(g)
Sub-Function: Calculate RSS observation
Probabilities
149
6.27(h)
Sub-Function: Predict Current State
150
6.27(i)
Sub-Function: Correct and Normalize the Predicted
State Probabilities
151
6.28
Choose and insert floor map
153
6.29
Floor maps for multi-floor environment loaded
153
6.30
Define topological map
154
6.31
Define logical area
154
6.32
Calibration Process. Go to calibration node and
face the direction of arrow shown. Press start
collect RSS button and wait until progress bar
completed
155
6.33
Save the model as radio map
155
6.34
The histograms in radio map can be observed
through analyzer. This example shows a part of the
histograms collected at calibration node at location
(420.0, 476.0).
156
6.35
Load radio map model created in off-line phase
156
6.36
Textual and graphical location report, <x, y, floor,
logical area>
157
7.1
Distribution of topological nodes and logical areas
on floor 1 of WCC, UTM
163
7.2
Distribution of topological nodes and logical areas
on floor 2 of WCC, UTM
164
7.3
Testing locations and testing paths used for
performance evaluation on floor 1 of WCC, UTM
166
xxii
7.4
Testing locations and testing paths used for
performance evaluation on floor 2 of WCC, UTM
167
7.5
CDF of location error for stationary multi-floor
location estimation
169
7.6
Comparison of average location error and time
duration for 750 location estimates in stationary
multi-floor location estimation
170
7.7
Percentage of correct logical area estimation and
percentage of correct floor estimation for stationary
multi-floor location estimation
171
7.8
CDF of location error for mobile multi-floor
location estimation
173
7.9
Comparison of average location error and average
MT’s walking speed during on-line phase for
mobile multi-floor location estimation
174
7.10
Percentage of correct logical area estimation and
percentage of correct floor estimation for mobile
multi-floor location estimation
175
7.11
Rectangular hallway test-bed with (a) 4 (b) 10 (c)
22 topological nodes
178
7.12
Staircase area test-bed with (a) 4 (b) 7 (c) 9
topological nodes
179
7.13
Effects of number of access points on average
location error during stationary location estimation
181
7.14
Effects of number of access points on average
location error during mobile location estimation
181
7.15
Effects of number of access points on percentage of
correct logical area estimation during stationary
location estimation
182
7.16
Effects of number of access points on percentage of
correct logical area estimation during mobile
location estimation
183
7.17
Effects of number of access points on percentage of
correct floor estimation during stationary location
estimation
184
xxiii
7.18
Effects of number of topological nodes on average
location error during stationary location estimation
185
7.19
Effects of number of topological nodes on average
location error during mobile location estimation
186
7.20
Effects of number of topological nodes on
percentage of correct logical area estimation during
stationary location estimation
187
7.21
Effects of number of topological nodes on
percentage of correct logical area estimation during
mobile location estimation
187
7.22
Effects of number topological nodes on percentage
of correct floor estimation during stationary
location estimation
188
7.23
Effects of number of RSS samples per fingerprint
on average location error during stationary location
estimation
190
7.24
Effects of number of RSS samples per fingerprint
on percentage of correct logical area estimation
during stationary location estimation
191
7.25
Effects of off-line phase sampling interval on
average location error during stationary location
estimation
192
7.26
Effects of off-line phase sampling interval on
percentage of correct logical area estimation during
stationary location estimation
193
xxiv
LIST OF SYMBOLS
a
-
Angle
A
-
Set of access points installed in indoor area
Bel (i)
-
Current belief / Corrected belief
Bel − (i)
-
Predicted belief
b
-
Number of topological nodes, V
c
-
Number of edges in topological map, E
d
-
Direction
D
-
Euclidean distance in signal space
E
-
Set of edges in topological map
e
-
Edge in topological map
F
-
Fingerprint information (RSS Values)
f
-
Floor number
G
-
Graph
i
-
Arbitrary index
j
-
Arbitrary index
l
-
Calibration node
L
-
Location Information
L
-
Set of calibration nodes
meanRSS
-
Mean received signal strength
M
-
Set of single RSS measurement
Na
-
Number of access points installed in indoor area
Nb
-
Number of access points detected at a given calibration node
Nc
-
Number of calibration nodes
Nm
-
Number of RSS samples in histogram
xxv
Np
-
Number of different RSS values in histogram
Ns
-
Number of state
O
-
Location sensor observation / Observed RSS value
r
-
Radius
RSS
-
Value of received signal strength
S
-
State space
S
-
State
st
-
State at time t
t
-
Time
V
-
Set of topological nodes in topological map
v
-
Topological node
x
-
x-axis coordinate
y
-
y-axis coordinate
z
-
z-axis coordinate
%
-
Percentage
α
-
Normalizing constant
θ
-
Orientation (Forward or backward)
xxvi
LIST OF ABBREVIATIONS
2D
-
Two-Dimensional
3D
-
Three-Dimensional
a.m.
-
Ante Meridiem
AOA
-
Angle of Arrival
AP
-
Access Point
API
-
Application Programming Interface
BER
-
Bit Error Rate
BSS
-
Basic Service Set
cdf
-
Cumulative Distribution Function
dB
-
Decibel
dBm
-
mili-Decibel
DOA
-
Direction of Arrival
DSSS
-
Direct Sequence Spread Spectrum
ESS
-
Extended Service Set
FHSS
-
Frequency Hopping Spread Spectrum
GHz
-
GigaHertz
GPS
-
Global Positioning System
GUI
-
Graphical User Interface
HP
-
Hewlett-Packard
IBSS
-
Independent Basic Service Set
IEEE
-
Institute of Electrical and Electronics Engineers
ISM
-
Industrial, Scientific and Medical
kHz
-
KiloHertz
k-NNSS
-
k- Nearest Neighbor in Signal Space
LANDMARC -
Location Identification based on Dynamic Active RFID
LBS
-
Location Based Services
LES
-
Location Estimation System
xxvii
LOS
-
Line of Sight
MAC
-
Medium Access Control
Mbps
-
Mega bits per second
MILES
-
Multi-floor Indoor Location Estimation System
MiniPCI
-
Mini Peripheral Component Interconnect
MT
-
Mobile Terminal
NNSS
-
Nearest Neighbor in Signal Space
NLOS
-
No Line of Sight
OFDM
-
Orthogonal Frequency Division Multiplexing
OS
-
Operating System
PAL
-
Precision Asset Location
PC
-
Personal Computer
PCMCIA
-
Personal Computer Memory Card International Association
PDA
-
Personal Digital Assistant
PHY
-
Physical
p.m.
-
Post Meridiem
POA
-
Phase of Arrival
RF
-
Radio Frequency
RFID
-
Radio Frequency Identification
RL
-
Reference Location
RSS
-
Received Signal Strength
RSSI
-
Received Signal Strength Indicator
SNR
-
Signal to Noise Ratio
SSID
-
Service Set Identifier
TDOA
-
Time Difference of Arrival
TOA
-
Time of Arrival
UNII
-
Unlicensed National Information Infrastructure
USB
-
Universal Serial Bus
UTM
-
Universiti Teknologi Malaysia
UWB
-
Ultra-wideband
WCC
-
Wireless Communication Centre
WLAN
-
Wireless Local Area Network
xxviii
LIST OF APPENDICES
APPENDIX
TITLE
PAGE
A
Graphical user interface of the proposed system
205
B
Award
209
C
Bayes’ Theorem
210
CHAPTER 1
INTRODUCTION
1.1
Background
The rapid development of mobile computing technology and high speed
wireless communication systems has fostered tremendous growth in location based
services (LBS). Through LBS, various applications and services are delivered to the
user based on their current physical location. Indoor environments present
opportunities for a rich set of LBS such as navigational tools for humans and robots,
interactive virtual games, resource discovery and asset tracking.
The key to the success of LBS is the estimation of user’s location. Location
estimation is a process of estimating physical location of a mobile terminal (MT)
with respect to a set of reference locations within a predefined space. This research
focuses on design and development of multi-floor indoor location estimation system
(MILES) using wireless local area network (WLAN). This system is proposed to
determine the MT’s location in a multi-floor indoor environment.
Currently, there are many location estimation systems (LES) available.
Global Positioning System (GPS) is the most common and universally used LES for
outdoor area. Unfortunately, GPS is not suitable for indoor applications due to the
absence of line of sight (LOS) from the MT to the GPS satellites [1].
2
Various alternatives are proposed to provide indoor location estimation.
Some researches use specialized hardware for indoor location estimation. These
hardware are designed specifically and solely for location estimation purpose only.
Ultrasonic, infrared, optical and radio frequency (RF) are major technologies used
for this type of system. Although these systems able to estimate indoor location
accurately, they are usually expensive in terms of investment and maintenance costs.
In order to overcome the disadvantages mentioned above, indoor LES can be
developed using existing infrastructures. WLAN, Bluetooth and cellular network are
major infrastructures used for this type of systems. These infrastructures are usually
developed for other purposes such as data networking and communication. By
developing a software layer on top of these infrastructures, a lower cost LES can be
achieved. In this study, the proposed MILES is built on top of off-the-shelf WLAN
infrastructure.
1.2
WLAN-Based Multi-Floor Indoor Location Estimation System
WLAN is widely deployed in various indoor areas such as homes, offices,
schools and museums. Besides using the WLAN infrastructure for wireless data
networking, the RF signals transmitted or received by WLAN devices can be used to
estimate the MT’s location. Because signal strength measurement is part of the
standard operating mode of WLAN devices, no other hardware infrastructure is
required. WLAN-based LES can be developed using proximity sensing, triangulation
or location fingerprinting technique.
Location fingerprinting technique is the most popular solution for WLANbased indoor location estimation [2]. The basic idea behind location fingerprinting is
that RF signal has different characteristics in different indoor locations. Location
dependent RF signal elements such as WLAN received signal strength (RSS), signal
to noise ratio (SNR) or bit error rate (BER) are used as the “fingerprint” of a
particular location.
3
Location fingerprinting usually works in two phases [3]. First, in off-line
phase, the LES is calibrated by collecting location fingerprints at finite predetermined reference locations (RL) within the targeted multi-floor indoor area and
stored in a database called radio map. Second, location is estimated in on-line phase.
The current observed location fingerprint is measured and the LES will determine the
best match between the on-line observations and the off-line fingerprints in the radio
map. The RL with the closest match is then reported as the estimated MT’s location.
This matching can be done according to deterministic or probabilistic algorithms.
In this study, the proposed system is developed using location fingerprinting
technique. WLAN RSS is used as the location fingerprint. Probabilistic algorithm is
adopted to estimate the MT’s location.
1.3
Problem Statement
Currently, conventional GPS system does not work well in indoor area.
Indoor LES based on specialized hardware usually require high investment and
maintenance costs. Therefore, an economic LES is needed.
WLAN-based indoor LES is one of the economic alternatives. Many
solutions are proposed for this type of system in previous researches and their
performances are very encouraging [4, 5, 6, 7]. However, most of the proposed
systems are only tested for single-floor indoor location estimation [4, 5, 6, 7]. In
reality, an indoor LES is usually used in a multi-floor environment.
In this study, location fingerprinting uses the WLAN RSS to estimate indoor
location. The RF propagation channel in indoor environment is complex due to
multi-path fading phenomenon. Therefore, a basic understanding of the indoor
WLAN RSS characteristics is crucial before the design, development and
deployment of the proposed system.
4
In addition, performance of WLAN-based location fingerprinting system is
affected by various factors such as number of WLAN access points (AP) installed. A
basic understanding on these factors will help to achieve and improve the targeted
LES performance level.
1.4
Objectives
The following objectives are determined in order to solve the problems
mentioned above.
(i)
To study the characteristics of WLAN RSS in indoor environment through
measurement for WLAN-based indoor location estimation application.
(ii)
To design and develop a WLAN-based indoor location estimation system for
multi-floor environment using RSS location fingerprinting technique.
(iii)
To evaluate the performance of the proposed WLAN-based multi-floor
indoor location estimation system.
(iv)
To study factors which affect the performance of the proposed WLAN-based
multi-floor indoor location estimation system.
1.5
Research Scope
In this study, a software-based MILES is designed and developed. The
proposed system is implemented and evaluated over the off-the-shelf IEEE 802.11g
WLAN infrastructures deployed in Wireless Communication Centre (WCC),
Universiti Teknologi Malaysia (UTM).
The study is divided into four major phases. In the first phase, literatures on
current indoor location estimation technologies and previous researches in the field
are reviewed. Strengths and weaknesses of available systems are compared.
5
In the second phase, measurements are conducted to study the characteristics
of WLAN RSS in multi-floor indoor environment. The goal of the measurement is to
understand the characteristics of WLAN RSS for indoor location estimation
application. Results obtained here are used to design and develop the proposed
system.
Based on the literature review and results from the RSS characteristics study,
the MILES is designed and developed. In this third phase, location fingerprinting
technique using WLAN RSS is proposed. Java programming language is used to
develop the proposed system.
In the final phase, the performance of the proposed MILES is evaluated via
real-time stationary and mobile multi-floor indoor location estimation experiments.
This is followed by a series of experiments on effects of four factors on the
performance of the proposed system. The factors investigated are number of WLAN
APs installed, number of topological nodes, number of RSS samples collected per
location fingerprint and off-line phase sampling interval.
1.6
Contributions
Contributions of this study are listed below:
(i)
A test-bed for WLAN-based multi-floor indoor location estimation is built at
WCC, UTM. This test-bed can be used as a general platform for future
researches related to WLAN-based indoor location estimation and WLAN
technology.
(ii)
Characteristics of WLAN RSS as location fingerprints are identified through
measurements. The results obtained can be used in future for design,
development and deployment of LES based on WLAN RSS.
6
(iii)
An economical software-based MILES using off-the-shelf WLAN
infrastructures is proposed and developed. The prototype can be further
developed into commercial product.
(iv)
Practicality and reliability of MILES using off-the-shelf WLAN
infrastructures is validated via real-time experiments.
(v)
Various factors affecting the performance of WLAN-based MILES is
identified. Results obtained will help to achieve and improve targeted
performance of a given MILES in real-time implementation.
1.7
Thesis Organization
The thesis consists of eight chapters. Chapter 1 introduces the background of
this study. In addition, the problem statement, research objectives, research scope,
research contributions and thesis organization are included.
Chapter 2 contains the literature reviews on general indoor LES technology.
Basic architecture of LES, location estimation technique, classification and
properties of LES are reviewed. Previous studies on LES using various technologies
are also reviewed. Advantages and disadvantages of these systems are identified.
Chapter 3 presents the overview on WLAN-based indoor LES. WLAN
technology is reviewed. A detailed description on WLAN-based indoor LES,
location fingerprinting technique and previous related studies are given. Strengths
and weaknesses of available systems are compared.
Chapter 4 discusses the methodology of the study. Approach taken to achieve
the research objectives is presented here. The test-bed, hardware and software
components used in the study are also listed.
7
Chapter 5 presents the study on WLAN RSS characteristics. Measurements
are conducted to study the characteristics of WLAN RSS for location fingerprinting
application.
Chapter 6 presents the design, development and implementation of the
proposed WLAN-based MILES.
Chapter 7 contains the real-time evaluation results of the proposed system.
Type of experiments, experimental setups and results are analyzed and discussed. In
addition, real-time experimental results on various factors affecting the performance
of the proposed system are presented and discussed.
Finally, Chapter 8 concludes the thesis with conclusion and gives directions
for future works.
CHAPTER 2
WIRELESS INDOOR LOCATION ESTIMATION SYSTEM
2.1
Basic Architecture of Wireless Indoor Location Estimation System
Basic architecture of a wireless indoor LES is illustrated in Figure 2.1 [8].
Generally, a wireless indoor LES consists of three main functional blocks: location
sensors, location estimation algorithm and location information display system.
During estimation of MT’s location, the location sensors will detect and
measure the wireless signals transmitted by or received at known RL in the operating
environment. The wireless signals detected by the location sensors can be RF signal,
ultrasound, infrared or other types of wireless signals. These wireless signals will be
converted into location dependent parameters called location metrics. Then,
deterministic or probabilistic location estimation algorithm will process the location
metrics and estimate the MT’s location. Finally, the location information display
system will display the estimated location information to the user in form of textual
coordinates or illustration on digital map.
The location metric relates the MT’s current unknown location with respect
to a known RL. The location metric may indicate the approximate arrival direction of
the signal or the approximate distance between the MT and RL. Commonly used
location metrics are time of arrival (TOA), angle of arrival (AOA) and phase of
arrival (POA). Alternatively, RF signal characteristics such as RSS, SNR and BER at
known RL can be used as fingerprint to represent that particular location uniquely.
9
Wireless signals:
RF, Infrared, ultrasound, ...
Location metrics:
ToA, AoA, RSS, …
Location
information:
(x, y, z), ...
Location
Sensor
Location
Estimation
Algorithm
Operating
Environment
Display
System
Location
Sensor
Figure 2.1: Basic architecture of wireless indoor location estimation system [8]
2.2
Location Sensor
2.2.1 Classification of Location Sensor
Location sensors can be classified into passive and active sensors [9]. These
two types of sensors use different techniques to detect and measure the wireless
signals from the MT’s operating environment.
Active sensors emit wireless signal into the operating environment and then
measure the reactions from the environment. The reactions measured will be
converted into location metrics which directly or indirectly indicate the MT’s
location. Active sensors are capable to interact with the environment in a more
controlled manner. Therefore, generally the performance of LES using active sensors
is better compared to passive sensors [9]. However, active sensing introduces several
risks. The outbound emitted signal may affect the reactions that the active sensor is
attempting to measure. Active sensor may experience interferences between its signal
and signals emitted by other nearby sensors. These may bring inaccuracy in the
measurement of reactions and consequently degrade the performance of location
estimation. Examples of active sensors are ultrasonic sensor and laser rangefinder.
10
In contrast, passive sensors detect and measure the wireless signals entering
the sensors. The signals measured will be converted into location metrics. The
performance of LES using passive sensors is usually lower than active sensors [9].
This is mainly due to the environment and signals measured are not under the control
of the sensors. However, implementation of passive sensing is simpler and easier
compared to active sensing. It is more power saving since no signal emission is
involved. Examples of passive sensors include WLAN client adapter and Bluetooth
client adapter.
2.2.2
Wireless Sensing Technology
2.2.2.1 Infrared
Infrared sensing uses short-range transmission of modulated infrared light to
transmit the identity of a MT to a fixed receiver at a particular known RL. Typically,
receivers are installed at every possible indoor location where the MT might be
found. When a particular receiver detects the identity of the MT, the location of that
receiver is assumed to be the current location of the MT.
Infrared sensor is cheaper compared to other technologies. However, infrared
signal has limited transmission range and cannot penetrate obstructions such as walls.
Consequently, many infrared sensors need to be installed around the indoor
environment and thus incur significant installation, reconfiguration and maintenance
costs. Furthermore, infrared sensing does not work well in indoor areas with
fluorescent lighting and direct sunlight because of the spurious infrared emissions
generated by these light sources. This reduces the infrared sensing accuracy. Due to
these limitations, the infrared sensing-based LESs are only capable to provide
coarse-grained location estimation [3].
11
2.2.2.2 Ultrasound
The basic principle of ultrasound sensing is to transmit an ultrasonic wave
and measure the time it takes for this wave to reach the receiver. Ultrasonic wave
typically has a frequency range between 40 to 180 kHz [9]. The measured time is
used to calculate the distance between the ultrasonic transmitter and receiver. The
MT’s location is then estimated based on a set of these distances. Ultrasonic
transmitter will be carried by the MT and ultrasonic receivers will be installed at
known RLs, or vice versa.
Ultrasound sensor is simple and cheap. However, similar to infrared,
ultrasound sensors have limited and short effective coverage range [9]. Thus many
ultrasound sensors are needed to cover a given indoor area, which incurred
significant installation and maintenance costs. In spite of these limitations,
ultrasound sensor-based LESs provide finest-grained location estimation compared
to infrared and RF sensing technologies [3].
2.2.2.3 Radio Frequency
RF technology such as WLAN, Bluetooth, Ultra-wideband (UWB) and Radio
Frequency Identification (RFID) can be used to estimate MT’s location.
Infrastructures of these technologies are typically designed for data communication
purposes but can be adopted as RF sensors. These RF technologies usually operate in
license-free frequency bands such as Industrial, Scientific and Medical (ISM)
spectrum and Unlicensed National Information Infrastructure (UNII) spectrum.
Since RF signal penetrates most indoor building materials and has a longer
transmission range in indoor environment. Thus in order to cover a given indoor area,
the number of RF sensors needed is less compared to infrared and ultrasound. In
addition, usually RF sensors come from existing hardware. Therefore, the purchasing,
installation and maintenance costs can be significantly reduced.
12
2.3
Location Estimation Technique
There are three basic techniques available to estimate MT’s indoor location,
which are proximity sensing, triangulation and location fingerprinting [10]. The basic
concepts of these techniques are described in the following sections.
2.3.1
Proximity sensing
Proximity sensing estimates the MT’s location by detecting the closeness of
the MT to a set of known RLs. The location of the closest RL is used as the MT’s
current location. The concept of proximity sensing is illustrated in Figure 2.2.
Referring to Figure 2.2, three RLs: RL1, RL2 and RL3, with known coordinates:
(X1, Y1), (X2, Y2) and (X3, Y3) are defined in the operating environment. Each RL has
a defined limited coverage area. Since the MT is currently closest to RL3, the MT’s
current location is estimated as (X3, Y3).
Proximity sensing is the easiest location estimation technique, because it
requires no or minor modifications on existing infrastructure and causes less
overhead [11]. However, proximity sensing only capable to provide coarse-grained
location estimation accuracy.
RL1
(X1, Y1)
Mobile Terminal (MT)
RL2
(X2, Y2)
RL3
(X3, Y3)
Reference Location (RL)
Coverage Area of RL
Figure 2.2: Concept of proximity sensing [11]
13
2.3.2 Triangulation
Triangulation technique estimates the MT’s location through the geometric
properties of triangle. Triangulation can be implemented via lateration or angulation.
2.3.2.1 Lateration
Lateration estimates MT’s location by measuring the distances between the
MT to a set of RLs. The distance can be measured using TOA, Time Difference of
Arrival (TDOA) or other techniques. In order to estimate two-dimensional (2D)
location, distance measurements from three non-collinear RLs are required.
Estimation of three-dimensional (3D) location will require distance measurements
from four non-coplanar RLs [11]. The concept of estimating MT’s 2D location using
lateration is shown in Figure 2.3.
Referring to Figure 2.3, assumed that the distances between MT and RLs are
given by r1, r2 and r3, respectively. In Figure 2.3(a), with knowledge of distance
between MT to RL1, the MT’s location can be estimated as any locations on the
circle with radius r1 centered at RL1. In Figure 2.3(b), with another distance
measurement from RL2, the uncertainty of MT’s location is reduced to two locations
where both circles intersect. Finally, in Figure 2.3(c), with knowledge of distances
between MT to RL3, the MT’s location can be pinpointed to a single location.
14
RL2
(X2, Y2)
RL2
(X2, Y2)
RL1
(X1, Y1)
RL1
(X1, Y1)
r1
r2
RL1
(X1, Y1)
r1
r2
r3
r1
RL3
(X3, Y3)
(a)
(b)
(c)
Mobile Terminal (MT)
Coverage Area of RL
Reference Location (RL)
Radius (r)
Figure 2.3: Concept of triangulation via lateration [11]
2.3.2.2 Angulation
Angulation estimates MT’s location by measuring the MT’s angle or bearing
relative to a set of RLs. The angle can be measured using AOA or Direction of
Arrival (DOA) [11]. In order to estimate 2D location, two angle measurements are
required as shown in Figure 2.4.
Referring to Figure 2.4(a), angle of signal transmitted by the MT is measured
at one RL, represented as a1. Thus the MT’s location can be estimated at any location
along the line connecting the MT and RL1. In Figure 2.4(b), with addition of angle
measurement from RL2, represented as a2, another line is created and the intersection
between these two lines represents the estimated MT’s location.
15
a1
RL1
(X1, Y1)
Mobile Terminal (MT)
a1
a2
RL1
(X1, Y1)
(a)
Reference Location (RL)
RL2
(X2, Y2)
(b)
Figure 2.4: Concept of triangulation via angulation [11]
2.3.3
Location Fingerprinting
Location fingerprinting technique observes the operating environment and
estimate MT’s current location from theses observations. It is also known as scene
analysis or pattern matching technique. This technique works under the assumption
that every physical location has a unique characteristic or fingerprint, as analogy to
every human being has a unique fingerprint [11].
Location fingerprinting involves two phases known as off-line phase and online phase. Firstly in off-line phase, fingerprints or observations are taken at predetermined finite RLs around the environment to construct the fingerprint database.
Then in on-line phase, the current fingerprint or observation at the unknown MT’s
location is measured and compared to the fingerprint database. Generally, the
location of the RL which has the closest match with the MT’s current fingerprint is
reported as the estimated current MT’s location.
Concept of location fingerprinting is illustrated in Figure 2.5. Referring to
Figure 2.5, fingerprints are collected at three pre-determined RLs during off-line
phase: RL1, RL2 and RL3. These location fingerprints are unique and are stored in a
database. During on-line phase, the current fingerprint at unknown location is
16
measured and compared with the database. Location of RL3, (X3, Y3), is reported as
the estimated current MT’s location, since its’ fingerprint is the closest match with
highest likelihood to the current fingerprint.
Fingerprint 1
RL1
(X1, Y1)
Current
Fingerprint
Room 1
RL3
(X3, Y3)
RL2
(X2, Y2)
Room 2
Fingerprint 3
Open Area
Fingerprint 2
Mobile Terminal (MT)
Reference Location (RL)
Figure 2.5: Concept of location fingerprinting
2.4
Properties of Wireless Indoor Location Estimation System
2.4.1
Network-Based System and Terminal-Based System
Main difference between network-based and terminal-based LES is the
platform that performs measurements and estimation of the location. In a networkbased LES, the MT’s location is measured and estimated by external infrastructure in
the network such as a location estimation server. This approach reduces the
computational burden and power consumption of MT.
17
On the other hand, in a terminal-based LES, the MT’s location is measured
and estimated by the MT itself. Since the MT’s location is only known to itself, its
privacy is protected. Other entities can access the location information only if the
MT publishes that information.
2.4.2
Specialized Infrastructure and Existing Infrastructure
LES based on specialized infrastructure uses hardware which are designed
exclusively for location estimation application. Ultrasound and infrared-based
systems are examples of LES based on specialized infrastructure. Advantages of this
type of system are that the straightforward infrastructure design and the performance
of location estimation can be controlled by the designer. Typically, these systems are
capable to provide accuracy in the millimeter range. However, proprietary
infrastructures will usually incur high investment and maintenance costs [3, 10].
LES based on existing infrastructures uses existing hardware such as WLAN,
Bluetooth and cellular network for location estimation. These infrastructures are
designed for other application such communication and data networking. With
installation of extension software, they can be reused for location estimation. This
type of system only needs low investment and maintenance costs and the LES does
not need to be built from scratch. However, it may bring additional burdens and
traffic loads to existing infrastructure. The accuracy achieved is usually in the range
of a few meters [3, 10].
2.4.3
Physical Location and Symbolic Location
LES can provide two types of location information: physical location and
symbolic location [10]. Physical location represents location as a single point in the
Euclidean space. It is usually expressed by means of 2D or 3D coordinates such as
(x, y, z) which represents the latitude, longitude and altitude.
18
On the other hand, symbolic location represents location through descriptions
related to geographical objects such as country, cities, roads, buildings and rooms. It
is usually expressed in terms of names, numbers or identifiers such as “Pantry”,
“Corridor on Floor 2” and “Room 123”. Symbolic location is more meaningful to
human because location is reported as real world abstraction.
2.4.4
Accuracy and Precision
The performance of a LES is usually expressed in terms of accuracy and
precision. Accuracy refers to the closeness of several estimated locations to the true
but unknown MT’s location. The closer the estimated location to the true location,
the higher is the degree of accuracy and vice versa.
Precision is the percentage of correct location estimations at a given accuracy
level. For example, when a given location estimation system is said to be capable to
locate a MT to within 2 meters for 90% of location estimations, the accuracy
mentioned is 2 meters and the precision is 90%.
2.4.5
Scalability
The scalability of a LES is limited by various factors such as operating space,
time and frequency of the system. Each LES has a defined operating space where it
can perform with acceptable accuracy and precision. The operating space covered
may be within a given multi-floor building or within a given room only.
Besides that, LES may have limited number of MTs that it can locate with a
certain number of infrastructures or over a given time. For instance, in a RF sensingbased LES, only a limited number of MTs can be located at a given time due to the
limited communication channels.
19
2.4.6
Cost
A practical LES should have a reasonable and acceptable cost in terms of
capital, time and space. Capital costs include the investments on infrastructures,
maintenance and the salaries for supporting personnel and other direct or indirect
financial costs. Time costs take into consideration the time required for installation,
maintenance frequency and administrative needs while space costs cover the amount
of infrastructures needed to cover a given area, hardware’s physical size and software
capacity.
2.5
Previous Researches on Wireless Indoor Location Estimation System
The Active Badge system uses infrared active sensing technology to estimate
MT’s location [12]. In this system, the MT carries a specially designed badge which
periodically emits a unique infrared signal for approximately a tenth of a second
every fifteen seconds. These periodic infrared signals will be detected by infrared
sensors installed at known RLs around the indoor environment such as ceiling tiles or
walls, rooms and entrances or exits of corridors. The detected signals will be sent to
the central server which will then estimate the MT’s location based on the proximity
of infrared badge to sensor. Active Badge system is only capable to provide symbolic
location information and coarse-grained accuracy. The main disadvantages of this
system are line of sight (LOS) requirement, limited infrared signal transmission
range and degraded performance under direct sunlight or fluorescent lights.
The Bats system uses RF and ultrasound active sensing technology [13]. In
this system, the MT carries a specially designed tag called Bat. During location
estimation, a central server will send request to the Bat via short-range RF signal.
The Bat will emit an ultrasonic pulse to a grid of ceiling mounted ultrasonic sensors
in response to the RF signal. When the central server send the RF request signal, a
synchronization reset signal is sent to the ceiling mounted sensors simultaneously via
wired serial network. Each sensors measures the ultrasound time of flight from the
Bat by calculating time interval from reset to ultrasonic pulse arrival. The ultrasonic
20
time of flight measurements will then be converted into distance measurements and
MT’s location will be estimated through distance lateration by the central server.
This system can provide accuracy of 3 centimeters with 95% of precision [14]. The
disadvantages of this system are high infrastructure cost and installation and
maintenance difficulties. Large amount of ceiling mounted ultrasonic sensors and
precise placement of these sensors are required to provide such high accuracy.
The Cricket Location Support System is also based on RF and ultrasound
technologies [15]. Complementary to Bats system, Cricket is a terminal-based LES.
In this system, wall- and ceiling-mounted beacons are installed around the building
and MT is attached with ultrasound and RF listener. Each beacon will periodically
transmit its’ location identity via RF signal together with an ultrasonic pulse. The
listener will correlate these RF and ultrasonic signals and estimate distances to
different beacons based on difference in RF and ultrasonic signal propagation times.
Then the current MT’s location is estimated based on distance lateration. Room-sized
accuracy is reported for this system [15]. Similar to Bats system, Cricket system
requires high infrastructures and installation costs.
Precision Asset Location (PAL) system [16, 17] used UWB technology. The
system consists of a central server, a number of passive UWB receivers and a UWB
reference tag which are installed at known RLs. MT is attached with an UWB active
tag which periodically transmit a short RF synchronization preamble signal followed
by tag identification information once every few seconds. This periodical UWB
signal will be detected by the passive receivers and its’ TOA will be measured. The
MT’s location will be estimated by a central server based on the TDOA data among
all passive receivers. System calibration is conducted at system startup by monitoring
data from the reference tag installed at known location. Results in [17] shows that
PAL is capable to achieve accuracy less than 1 foot. UWB is a potential solution for
high accuracy location estimation since it is capable to provide sub-foot resolutions
even in the presence of severe multi-path such as in indoor environment [17].
However, it will incur higher cost since this type of system is based on specialized
infrastructures
21
Location Identification based on Dynamic Active RFID (LANDMARC)
system is a LES based on RFID technology [18]. LANDMARC consists of a number
of RFID readers and reference RFID tags installed at known RLs. MT carries an
active RFID tag which will periodically transmit tag identification information
through RF signal. The RF signal will be detected by the RFID readers and later sent
to a central server. MT’s location will be estimated based on proximity of the tag to
the RFID readers and a number of reference tags. The proximity of the tag is
estimated based on the RSS information. The accuracy of this system is very
dependent on the density and location of the RFID readers and reference tags.
2.6
Summary
In this chapter, the wireless indoor location estimation technology is
overviewed. In order to design a wireless indoor LES, there are various types of
wireless technologies and location estimation techniques available. Every approach
has its own advantages and disadvantages.
Based on the literature review, there are many types of wireless indoor LES
available [12, 13, 14, 15, 16, 17, 18]. Most of these systems have common
limitations when operate in indoor environment. Wireless sensing signals such as
infrared cannot penetrate through walls and floors which are common in multi-floor
indoor environment. Systems based on specialized hardware usually incur higher
cost in term of investment, hardware or software installation and maintenance.
These disadvantages can be overcome by using existing WLAN
infrastructures for indoor location estimation. The following chapter will review the
technology and researches on WLAN-based indoor location estimation.
CHAPTER 3
WLAN-BASED INDOOR LOCATION ESTIMATION SYSTEM
3.1
Overview of IEEE 802.11 WLAN
The 802.11 standard is the internationally sanctioned WLAN standard
approved by Institute of Electrical and Electronics Engineers (IEEE). The base IEEE
802.11 specification includes the 802.11 medium access control (MAC) layer and
two physical layers (PHY): Direct Sequence Spread Spectrum (DSSS) layer and
Frequency Hopping Spread Spectrum (FHSS) layer [19].
Currently, IEEE 802.11 WLAN infrastructures are available in 2.4 GHz ISM
bands and 5 GHz UNII bands. Three versions of IEEE 802.11 are available in market,
namely the IEEE 802.11b, IEEE 802.11a and IEEE 802.11g. These standards operate
in different frequencies, bandwidth, modulation techniques, data rates, coverage and
other characteristics.
In this study, the proposed system is implemented on top of IEEE 802.11g
WLAN infrastructures. IEEE 802.11g is sanctioned in year 2003 and it uses the
Orthogonal Frequency Division Multiplexing (OFDM) technology. It operates in 2.4
GHz ISM bands, provides data rates of up to 54 Mbps and achieves coverage ranges
between several tens of meters and a few hundreds of meters, strongly depending on
the operating environment. It is backward compatible with the IEEE 802.11b
infrastructures.
23
3.1.1 Basic Components of IEEE802.11 WLAN
A basic IEEE 802.11 WLAN consists of four components which are AP, MT,
wireless medium and distribution system as shown in Figure 3.1 [19]. The MT is a
WLAN-enabled computing device such as laptop, Personal Digital Assistant (PDA)
or desktop. AP acts as the bridge between the distribution system and the wireless
medium. The distribution system can be implemented either as wires, radio links or
optical links.
In this study, the MT used is a WLAN-enabled tablet personal computer (PC).
The APs are linked together with wired Ethernet.
Wireless
Medium
Access
Point
Distribution
System
Mobile
Terminal
Figure 3.1: Basic components of IEEE 802.11 WLAN [19]
3.1.2 Topologies of IEEE 802.11 WLAN
WLAN can be constructed in ad hoc or infrastructure mode [19]. In ad hoc
mode, the MTs form a peer-to-peer wireless network where MTs communicate
directly with each other as illustrated in Figure 3.2(a). It is also known as
Independent Basic Service Set (IBSS). It is usually used over short distances with
limited number of participants and without integration into a larger network structure
such as wired networks.
24
In infrastructure mode, the MTs communicate with each other via AP as
illustrated in Figure 3.2(b). This type of network is called Basic Service Set (BSS). If
more than one BSSs are operating in the same WLAN via the distribution system, the
network is known as Extended Service Set (ESS) as illustrated in Figure 3.2(c).
Both ad hoc and infrastructure mode WLANs can be used for location
estimation purposes. However, the location estimation approaches used for these
modes are different. In this study, only infrastructure mode WLAN-based LES will
be studied. The WLAN is configured in infrastructure mode with ESS topology.
Access
Point
Mobile Terminal
IBSS
BSS
(a)
(b)
Distribution System
Access
Point
Access
Point
BSS1
BSS2
ESS
(c)
Figure 3.2: WLAN topologies. (a) Independent basic service set (b) Basic service set
(c) Extended service set [19]
25
3.1.3 Basic Operation of IEEE 802.11 WLAN
Communication in WLAN is divided into three basic operations: scanning,
authentication and association [19]. When a MT wanted to join a WLAN, firstly it
needs to scans for existence of WLAN within range. There are two types of scanning:
passive and active scanning [19]. Basic processes for passive and active scanning are
illustrated in Figure 3.3 and Figure 3.4 respectively. In passive scanning, the MT
listens to beacons on each channel. When a beacon is received, related WLAN
information will be extracted. In contrast, in active scanning, the MT sends a probe
request frame into each channel and waits for response from APs within range. When
a probe response frame is received, related WLAN information will be extracted.
After scanning, the best AP will be chosen and the MT will sent an association
request.
When the WLAN detects the association request, the authentication process
will be performed. The identity of the MT will be verified either through open
system or shared key authentication service. If the authentication process is
successful, then the MT will be allowed to associate with the WLAN.
Measurement of RSS of the received packets is a part of these standard
operations. In this study, this RSS information will be extracted for indoor location
estimation application. The RSS information is obtained via passive scanning process.
Distribution System
Access
Point
Access
Point
Beacons
Mobile Terminal
Figure 3.3: Passive scanning [19]
26
Distribution System
Access
Point
1. Probe request
Access
Point
2. Probe response
Mobile Terminal
Figure 3.4: Active scanning [19]
3.2
Overview of WLAN-Based Indoor Location Estimation System
The basic architecture of a WLAN-based indoor LES is depicted in Figure
3.5. The system is build on top of existing WLAN infrastructures. The hardware
constitutes of a number of APs and a MT such as laptop or PDA. The MT must be
WLAN-enabled in order to communicate with the APs.
The MT’s location will be measured and calculated by the location estimation
software. If the LES is terminal-based system, the software will be installed in the
MT. Otherwise, if the LES is network-based system, then it will be installed in a
server connected to the WLAN. The location estimation software will estimate the
MT’s location based on the WLAN related location metrics such as currently
associated AP’s MAC address, RSS or SNR. The estimated location will be reported
to the user in textual coordinates or illustration on digital map through the graphical
user interface (GUI).
27
Indoor Location
Estimation System
Graphical User
Interface
WLAN Signal
Access Points
Mobile Terminal
(WLAN-enabled)
Indoor Environment
Location
Metric
Estimated
Location
Location Estimation
Software
Figure 3.5: Basic architecture of a WLAN-based indoor location estimation system
3.3
WLAN-Based Indoor Location Estimation Techniques
WLAN-based indoor LES can be implemented using the three techniques
mentioned in Section 2.3: proximity sensing, triangulation and location
fingerprinting. These techniques will be discussed and compared in the following
sections.
3.3.1
WLAN-Based Proximity Sensing
WLAN-based proximity sensing system is the simplest technique, but suffers
from degraded accuracy. In this technique, the location of the AP closest to the MT is
reported as the current MT’s location. There are two approaches to estimate the AP
which is currently closest to the MT.
28
The first approach is to adopt the location of the AP that the MT has scanned
with the best signal quality as the closest AP. This approach is based on the
assumption that the closer AP will give the better signal quality. However in reality,
due to the various indoor propagation effects and the building structure, the AP with
best signal quality may not be the MT’s closest AP.
The second approach is to adopt the location of the AP that the MT is
currently associated with as the closest AP. This approach is based on the assumption
that MT always associate with the closest AP. However, the weakness of this
approach is similar to the above. The MT may be associating with AP that is not the
closest due to propagation effects, especially when the MT is moving and roaming.
The advantage of proximity sensing is that location metric required for
location estimation, that is AP scanned with best signal quality or currently
associated AP, can be easily obtain from the WLAN client device. This technique
does not require any parameter estimation and does not involve complex estimation
algorithms.
The drawback of proximity sensing is that it is only capable to provide
accuracy in the range of several tens to hundreds of meters, depending on the
coverage radius of AP and density of APs in the indoor area. For instance, if an AP
has a 100 meters x 100 meters coverage area, then proximity sensing only can
estimate that the MT is located within this 10000 meters2 area. It is also difficult for
proximity sensing to discriminate between different floors due to propagation effects.
In addition, to implement this system, the physical locations of all APs in the indoor
area must be known and this information is not always accessible to the system
designer [11].
29
3.3.2
WLAN-Based Triangulation
WLAN-based triangulation system is not very attractive and often yields
highly erroneous results [20]. In this technique, the location of the MT is determined
through WLAN signal triangulation, either using lateration or angulation. The
distances or angles such as TOA and AOA between APs and MT are derived by
using mathematical RF propagation models.
According to [21], TOA is the most popular approach among other
triangulation techniques. TOA-based system uses the TOA of the first path to
estimate distance. However, it is a challenging task to estimate TOA of the first path
accurately in indoor propagation environment [21]. In addition, derivation of
distances from timing measurements are not practical, because it is difficult to obtain
precise timing synchronization using off-the-shelf WLAN infrastructures and time
differences would be difficult to measure owing to the short ranges the signals travel
in indoor and local environment [11, 22].
In order to obtain accurate distance or angle measurements, all the APs in the
indoor area must be accurately positioned and aligned. Due to the severe multi-path
effects in indoor area and limited mathematical propagation model, it is difficult to
obtain accurate distance or angle measurements using off-the-shelf WLAN
infrastructures. In order to implement this technique effectively, more complex
mathematical propagation model and location estimation algorithms are needed.
3.3.3
WLAN-Based Location Fingerprinting
Due to the multi-path fading in indoor areas, WLAN-based proximity sensing
and triangulation location estimation techniques are not capable to provide high
accuracy. Owing to the limitations and difficulties of the two techniques mentioned
above, indoor location estimation based on WLAN location fingerprinting is more
popular [11].
30
Compared to proximity sensing and triangulation, WLAN location
fingerprinting technique provides higher accuracy in indoor area. Rather than
negatively impacted by the unpredictable indoor propagation effects, WLAN
fingerprinting technique turns the propagation phenomenon into good use. It takes
into consideration the real-time behaviors of WLAN signals and propagation factors
such as reflection, attenuation and multi-path effects when estimating the MT’s
location [11].
3.4 Overview of WLAN-Based RSS Location Fingerprinting
3.4.1
Basic Operation
Indoor location estimation using WLAN RSS location fingerprinting is based
on the basic concept that RSS value from a given AP is different at every physical
location in an indoor area. This is generally true due to the complex indoor
propagation phenomenon especially multi-path fading. The unique RSS value
collected at a given RL is used as the “fingerprint” for that particular RL. Indoor
location estimation based on WLAN RSS location fingerprinting consists of two
main phases: off-line and on-line phases (refer to Figure 3.6).
In off-line phase, the RSS location fingerprints will be collected at a set of
finite pre-determined RLs. Firstly, the locations and numbers of RLs are decided.
Then, RSS values from APs detected at each RL will be collected. These RSS values
will be combined with the RL’s location information and known as location
fingerprint. The location fingerprint will then be stored in a database called radio
map. On-line phase is the location estimation phase. When the MT’s current location
is required, the current RSS values observed by the MT will be recorded and
compared with the location fingerprints in the radio map. The MT’s current location
will be calculated through the location estimation algorithm. Location of RL with the
highest likelihood will be returned as the estimated MT’s location.
31
Access Points
Set of RLs
Radio Map
WLAN
Signals
Location
Information
RSS Values
+
Location
Fingerprint
Location
RSS
Information Values
Location 1 Fingerprint 1
Location 2 Fingerprint 2
Mobile Terminal
(WLAN-enabled)
(a)
Radio Map
Access Points
Location
RSS
Information Values
Location 1 Fingerprint 1
Location 2 Fingerprint 2
WLAN
Signals
Location
Fingerprint
RSS Values Location Estimation
Estimated
Location
Algorithm
MT’s current
location
Mobile Terminal
(WLAN-enabled)
(b)
Figure 3.6: Basic operation of WLAN-based RSS location fingerprinting. (a) Offline phase (b) On-line phase [3]
32
3.4.2
Reference Location
There are two approaches in selecting and locating RLs in the indoor area.
The RLs can be tabulated equally on a regular grid or located irregularly, depending
on the type of application and accuracy required. Both of these scenarios are
illustrated in Figure 3.7.
The total number of RLs in the indoor area will affect the accuracy of the
LES. For a given indoor area, if the number of RLs is too little, the system will
provide coarse-grained accuracy since the radio map is small. On the other hand, if
the number of RLs is too many, it will increase the consumption of resources
including off-line phase location fingerprint collection time and efforts, memory
space for radio map and also computation power during on-line phase.
Besides that, the physical distance between RLs is also important. If the
distance between RLs is too close, the location fingerprints collected at two
consecutive RLs will be similar. This similarity will not help in improving the
accuracy of the system, but instead increase the off-line phase location fingerprint
collection and on-line phase computation burden.
Reference Location (RL)
Indoor Area
(a)
Indoor Area
(b)
Figure 3.7: Reference locations. (a) Regular grid (b) Irregular
33
3.4.3
Location Fingerprint
Generally a single location fingerprint in the radio map can be represented as
(L, F ), where L denotes the location information of a given RL and F denotes the
fingerprint information or RSS values collected at that particular RL.
3.4.3.1 Location Information
The location information, L for indoor environment can be represented in two
forms: as a tuple of real physical coordinates or as an indicator variable. The real
physical coordinates can be represented with different dimensions varied from one to
five [23]. Representation in one dimension indicates that the MT only move along a
fixed line or axis, while five dimensions will include the 3D physical coordinates and
two orientation variables expressed in spherical coordinates. For instance, location
information represented in four dimensions can be expressed as:
L = { (x, y, z), d }
(3.1)
where (x, y, z) are 3D physical coordinates and d represent either North, South, East
or West orientation of the MT.
On the other hand, the location information can be represented as an indicator
variable or known as logical area. Instead of using real coordinates, RLs are
represented with variable according to the function of the location such as room
number or name. For instance, location information can be expressed as:
L = { Location_ID }
(3.2)
where Location_ID represents the logical area such as “Meeting Room”, “Lobby” and
“Floor 2”.
34
3.4.3.2 Fingerprint Information
The fingerprint information, F or RSS values in the location fingerprints can
be obtained via two approaches: empirical or mathematical modeling approach [4].
In empirical approach, the real-time RSS reported by the WLAN client device will
be used as the fingerprint. In off-line phase, user has to physically go to each RL to
collect the real-time location fingerprint. The main advantage of this approach is that
the RSS information collected is more realistic and accurate. The drawback of
empirical approach is that it is time and energy consuming if the indoor area is large.
In addition, the accuracy of the LES will be reduced if the indoor environment during
on-line phase is different compared to the condition when the radio map is
constructed.
The mathematical modeling approach is more flexible compared to empirical
approach. Instead of manually record the real-time RSS values at each RL,
mathematical RF propagation models are used to simulate the RSS observed at each
RL. The main advantage of this approach is that it is time and energy saving since
the off-line phase can be done automatically through software in a computer. If the
indoor environment is changed, only little efforts are needed to rebuild the radio map.
The drawback of this approach is the reduced accuracy compared to empirical
approach. This is mainly due to the inaccuracy of the RF propagation model used.
Typically, the RSS values, F, are collected more than one sample at each RL.
This is mainly because WLAN RF signal fluctuates over time. Therefore, instead of
representing F as a single RSS value, this randomness is usually captured by
collecting a set of RSS samples. Then descriptive statistics parameter or statistical
distribution is derived from the samples collected for each AP at each RL to
represent F.
35
For instance, RF characteristic at a given RL is represented with mean RSS
value [4, 5]. A set of RSS values from N b APs is captured at a given RL. The mean
RSS value for each AP is then calculated. Finally, the RSS value, F at the given RL
is represented as:
F = (MeanRSS0 , MeanRSS1 , &, MeanRSS Nb-1 )
(3.3)
where MeanRSSi is the mean RSS value for the ith-AP at the given RL. N b
represents the total number of APs detected at the given RL.
Table 3.1 shows an example of radio map which consists of location
fingerprints collected at three RLs. The location information L for each RL is
represented with 3D physical coordinates and one orientation information. At each
RL, a total of three APs can be detected, N b = 3. At each RL, the RSS from each AP
is taken for a given sampling duration. Then mean RSS value is calculated for each
AP at each RL. These are used as the RSS value, F at each RL.
Table 3.1: Example of radio map
Location Information, L
3D Location
(x, y, z)
RL1
(2, 2, 2)
RL2
(20, 35, 32)
RL3
(10, 3, 20)
Orientation
North
South
East
West
North
South
East
West
North
South
East
West
RSS Value, F (dBm)
Mean RSS
from
AP A
-69
-65
-62
-63
-39
-35
-46
-36
-56
-55
-51
-52
Mean RSS
from
AP B
-50
-55
-60
-53
-45
-55
-50
-49
-46
-50
-55
-45
Mean RSS
from
AP C
-56
-60
-52
-50
-52
-58
-58
-50
-66
-57
-59
-63
36
3.4.4
Location Estimation Algorithm
There are two main categories of location estimation algorithm which are
deterministic and probabilistic [3]. These two categories differ in the approach taken
to model the relationship between location information, L and RSS value, F.
In deterministic approach, the RSS values from each AP at a given RL are
represented by a scalar value such as the mean RSS value. Then, non-probabilistic
algorithms are used to estimate the MT’s location. Nearest neighbor algorithm and
neural network algorithm are examples of deterministic algorithm.
In probabilistic approach, the RSS values from each AP at a given RL are
represented with RSS probability distributions. Probabilistic algorithms such as
Bayesian filtering algorithm and support vector machines are used to estimate the
MT’s location.
Examples of deterministic and probabilistic algorithms will be presented in
the following section.
3.4.4.1 Nearest Neighbor in Signal Space
Nearest neighbor in signal space (NNSS) algorithm is the simplest
deterministic algorithm [4, 5]. In the off-line phase, mean RSS of each AP at every
RL is recorded in the radio map. During on-line phase, the NNSS algorithm will
compute the Euclidean distance in signal space between the currently observed RSS
and each location fingerprint in the radio map.
37
The Euclidean distance in signal space, D j is given by the following equation:
Dj =
Nb −1
∑ ( RSS
i =o
i
− RSSij′ ) 2
(3.4)
where, Nb is the total number of APs detected at each RL, RSSi is the current
observed RSS value from the i-th AP and RSSij′ is the mean RSS value from the i-th
AP of the j-th location fingerprint entry in the radio map.
The RL with the minimum Euclidean distance, minimum[ D j ] is the NNSS to
the MT and is chosen as the estimated location of the MT. Location estimated using
NNSS algorithm are discrete. Since the estimated location is chosen from the set of
RLs in the radio map, radio map with higher density or more location fingerprints
will increase the accuracy.
Although NNSS algorithm is simple, it suffers from the following weakness.
Due to indoor RF propagation phenomenon, two different locations may have the
same RSS value from a given AP. In other words, two physically distant RLs may be
near in signal space. NNSS algorithm may return the physically distant RL as the
estimated location instead of the nearer RL. This will reduce the location estimation
accuracy.
The weakness of NNSS algorithm can be improved by using average NNSS
algorithm known as k-NNSS algorithm [4]. Instead of using a single nearest neighbor
as the estimated location, k-NNSS uses the average of k number of nearest
neighbors’ locations.
The concept of NNSS algorithm and k-NNSS algorithm are illustrated in
Figure 3.8. From Figure 3.8, the location estimated using k-NNSS algorithm is closer
to the true location compared to the NNSS algorithm.
38
Neighbor 2
3-NNSS
NNSS
True Location of MT
Reference Location (RL)
Neighbor 1
Estimated MT’s Location with NNSS
Estimated MT’s Location with 3-NNSS
Neighbor 3
Figure 3.8: Nearest neighbor in signal space (NNSS) and k-NNSS (k = 3)
3.4.4.2 Probabilistic Algorithm
The probabilistic algorithm represents the RSS at a given location as a
probability distribution [24]. By using probability distribution, the RSS information
can be described more completely compared to deterministic approach.
One of the popular probabilistic algorithms is Bayesian filtering [25]. The
location fingerprints are modeled with conditional probability and Bayes’ rule is
applied to estimate the location.
The general Bayes’ rule is expresses as:
p (l | o) =
p(o | l ) p(l )
p (o)
(3.5)
where p (l | o) denotes the conditional probability of MT is currently at location l
given the currently observed RSS value o . It is also known as a posteriori
probability.
39
p (o | l ) denotes the conditional probability of observing RSS value o given the MT
is currently at location l . It is also known as the likelihood function or a priori
probability.
p (l ) denotes the prior probability that MT is located at location l .
p (o) denotes the probability of the observed RSS value o .
The prior probability, p (o | l ) can be extracted from the MT’s movement
history. For example, previously the MT is in location li , then the probability of MT
is now located to places near location li is higher compared to elsewhere. Besides
that, this prior probability can also be obtained based on MT’s behavior or
experience, for example, usually the MT will be at location li most of the time, then
the prior probability of being at location li is higher compared to elsewhere. If this
prior knowledge is not available, it can be assumed to be uniformly distributed. The
probability of the observed RSS value, p (o) is independent of the location l ,
therefore it is treated as a normalizing constant. The likelihood function or a priori
probability, p (o | l ) can be generated from the location fingerprints in the radio map,
since the location fingerprints represent the pre-collected RSS values at each RL.
Bayes’ rule is used to calculate the a posteriori distribution of possible MT’s
location based on the a priori probability, p (o | l ) . From the a posteriori distribution,
location with the maximum a posteriori probability, maximum[ p (l | o) ] will be
selected as the estimated MT’s current location. For example, if the a posteriori
probability of being at location li , is higher than the a posteriori probability of being
at location l j , p (li | o) > p (l j | o) , then location li will be selected as the MT’s current
location. Since the probabilistic algorithm captures more information on the RSS
distribution, the accuracy of location estimation based on probabilistic algorithm is
usually better than deterministic approach such as NNSS algorithm.
40
3.5
Bayesian Filtering for Location Estimation
Location estimation is the process of estimating MT’s current location from a
stream of location metrics observed or measured by the location sensors. Most of the
location sensors do not provide perfect observations and do not work well in all
environments [25]. This will directly reduce the location estimation performance
such as accuracy and precision. Therefore, it is crucial to represent the uncertainty in
the location metrics provided by the location sensors.
Bayesian filtering is a powerful technique to represent the uncertainty and
ambiguity in location sensor. Due to its’ probabilistic statistical nature, Bayesian
filter is applicable to arbitrary types of location sensor and representation of
environments [26]. It has been successfully adopted in various estimation
applications such as robot tracking, speech recognition and vision recognition. The
following section will review the Bayesian filtering framework for location
estimation application.
3.5.1
Bayesian Filter
Bayesian filter estimate the state of a dynamic system from a sequence of
location sensor’s observations probabilistically. The state indicates the MT’s location
and the location sensor provides the location metric observed at a given state. The
state can be a simple 2D location, a node in a topological map or a vector including
3D physical coordinates, orientation and velocity [25]. In WLAN-based RSS
location fingerprinting approach, the location metric is the RSSs detected by the MT.
All possible MT’s locations in the indoor area are represented as state space
Si . The state of the MT at time t is represented as random variables st . The
probability distribution over state st , Bel ( st ) or known as belief, represents the MT’s
location uncertainty at each point of time. Given the location sensor’s observations,
41
Bayesian filter will sequentially estimate the belief Bel ( st ) over the entire state space
Si . Let O = {o1 , o2 ,..., ot } denotes a sequence of time indexed location sensor’s
observations. Then, the belief Bel ( st ) can be defined by the posteriori density over
the random variables st conditioned on all location sensor observations available at
time t :
Bel ( st ) = p ( st | o1 , o2 ,..., ot )
(3.6)
The term p( st | o1 , o2 ,..., ot ) can be interpreted as the probability of MT at location or
state st if the history of location sensor’s observation is given by ( o1 , o2 ,..., ot ) [25].
The complexity of computing (3.6) will grow exponentially over time as the
number of location sensor’s observations increases over time. In order to reduce the
complexity, Bayesian filter assumes that the dynamic system is Markov. Under
Markov assumption, the location sensor’s observations depend only on MT’s current
location and MT’s state at time t, st depends only on previous state, st −1 [25].
The belief Bel ( st ) will be updated recursively from time to time. The update
of the Bayesian filter is performed in two steps which are the prediction stage and
correction stage. When the location sensor provides a new observation, the Bayesian
filter will firstly predict the state of MT and then correct the predicted estimation
using the observation.
(i)
Prediction stage
The Bayesian filter will predict the state according to the following equation:
Bel − ( st ) ← ∫ p ( st | st −1 ) Bel ( st −1 ) dst −1
(3.7)
where Bel − ( st ) denotes the predicted belief at time t , Bel ( st −1 ) denotes the
previous belief at time (t-1) and p ( st | st −1 ) is the motion model which
42
describes the system dynamics. This motion model represents how the state
of the system changes over time by telling where the MT might be at time t
given that it was previously at location or state st −1 [25].
(ii)
Correction stage
Then the predicted belief Bel − ( st ) obtained from the prediction stage will be
corrected using the new location sensor’s observation ot . Correction is
performed according to the following equation:
Bel ( st ) ← α t p (ot | st ) Bel − ( st )
(3.8)
where Bel ( st ) denotes the corrected or current belief. Bel − ( st ) denotes the
predicted belief obtained from prediction stage. The term p (ot | st ) is the
perceptual model which describes the likelihood of making observation ot
given the MT is at location or state st . The term α t is a normalizing constant
which ensures that the posteriori over the entire state space Si sums up to one
[25].
The updating process of Bayesian filter can be represented by combining equation
(3.7) and equation (3.8) to become:
Bel ( st ) ← α t p (ot | st ) ∫ p ( st | st −1 ) Bel ( st −1 )dst −1
(3.9)
Equation (3.9) shows the updating equation for a continuous state space. If the state
space is discrete, the updating process is given by:
Bel ( st ) ← α t p(ot | st )∑ p( st | st −1 ) Bel ( st −1 )
st −1
(3.10)
43
Figure 3.9 illustrate the basic concept and steps in recursively updating the
Bayesian filters for location estimation application [25]. In this simple example, the
MT is restricted to one-dimensional movement along a horizontal hallway. The MT
is equipped with location sensor which can detect the existence of doors along the
hallway such as camera.
Figure 3.9(a) indicates the initial state of the system, where t = 0. Here, the
prior knowledge about MT’s initial location is assumed unknown. Therefore, the
initial predicted belief Bel − ( s0 ) is initialized with a uniform distribution over all
possible locations. Next, the location sensor observed the environment and sends
“door found” information to the LES. With this location sensor’s observation, the
belief is corrected as Bel ( so ) resulting in higher probability at locations with doors
and lower probability elsewhere, as shown in Figure 3.9(b). Since there are three
doors along the hallway, the probability distribution possesses three peaks.
Probability for locations without door is non-zero in order to account for inherent
uncertainty in location sensor’s observation.
Then the MT moves to the right along the hallway as indicated in Figure
3.9(c). Due to this motion, Bayesian filter predict and shifts the belief Bel − ( s1 ) in the
direction of motion. In addition, the belief Bel − ( s1 ) is smoothed out to account for
inherent uncertainty in motion estimates. Finally, the process repeats where the
location sensor observed another door in Figure 3.9(d) followed by a MT’s
movement again in Figure 3.9(e). It can be observed from Figure 3.9(e) that, now the
probability distribution is higher at MT’s current location compared to previous steps.
In the example above, Figure 3.9(c) and Figure 3.9(e) show the prediction
stage where the Bayesian filter predicts the motion of the MT through motion model.
The prediction stage produces the predictive belief, Bel − ( st ) . The prediction stage is
followed by correction stage where the predictive belief is corrected using the
perceptual model, as shown in Figure 3.9(b) and Figure 3.9(d). The correction stage
produces the corrected belief, Bel ( st ) .
44
t=0
Initialization
Bel − ( s0 )
s
Bel − ( s0 ) =
Uniform distribution
(a)
t=0
Correction stage
Bel ( s0 ) =
p (o | s )
s
α 0 p (o0 | s0 ) Bel − ( s0 )
Bel ( s0 )
s
(b)
t=1
Prediction stage
Bel − ( s1 ) =
Bel − ( s1 )
s
∫ p(s
1
| s0 ) Bel ( s0 )ds0
(c)
t=1
Correction stage
Bel ( s1 ) =
p (o | s )
s
α1 p (o1 | s1 ) Bel − ( s1 )
Bel ( s1 )
s
(d)
t=2
Prediction stage
Bel − ( s2 ) =
Bel − ( s2 )
s
∫ p(s
2
| s1 ) Bel ( s1 )ds1
(e)
Figure 3.9: One-dimensional illustration of Bayesian Filtering technique [25]
45
3.5.2 Implementation of Bayesian Filter
Bayesian filter is an abstract concept which only provides the probabilistic
framework for recursive state estimation [25]. In order to implement Bayesian filter,
the following components need to be specified: belief Bel ( st ) representation,
perceptual model p (ot | st ) and the motion model or system dynamic p( st | st −1 ) .
There are many different types of representation for the belief function
Bel ( st ) of Bayesian filter. There are two main approaches to represent the Bayesian
function’s belief. First approach is to assume the sensor observations and process
have a particular continuous distribution such as Gaussian. Kalman filter, extended
Kalman filter and Multi-hypothesis tracking are examples of continuous
representation of Bayesian filter’s belief. Second approach is to approximate the
location probability distributions through discrete distribution. Topological, grid and
particle filter are examples of discrete representation of Bayesian filter’s belief. An
overview on the various representations mentioned above can be found in [25, 27].
The representation for perceptual model p (ot | st ) will depend on the types of
location sensors. For instance, in WLAN RSS location fingerprinting system, the
perceptual model p (ot | st ) can be interpreted as the likelihood of measuring a certain
RSS value ot given that the MT is at location or state st .
The motion model or system dynamics p( st | st −1 ) represents how the system
changes over time. The motion model is dependent on the location related
information available during the location estimation process. For example, the MT’s
movement velocity can be used to predict the next possible location or the building
structure information such as MT must go through staircase in order to reach the next
floor.
46
3.6
Related Researches
In this section, some of the previous studies and works on WLAN RSS
location fingerprinting system will be reviewed and discussed.
Firstly, WLAN RSS location fingerprinting system based on deterministic
approach will be reviewed. Microsoft Research RADAR location system is one of
the pioneering researches in WLAN-based indoor LES using RSS location
fingerprinting technique [4]. Proprietary 2.4 GHz WLAN infrastructures were
applied to locate and track MT within the second floor of a three-floor building. Both
empirical and mathematical modeling approaches were used to construct the radio
map. In empirical approach, real-time RSS values from three APs were collected
together with the RL’s locations and orientations. The NNSS algorithm gave a
median accuracy of 2.94 meters. In modeling approach, Wall Attenuation Factor
model was applied to simulate the RSS values in the location fingerprints [4]. This
approach decreased the median accuracy to 4.3 meters. These encouraging results
show that, despite the hostile nature of the indoor propagation environment, WLAN
RSS location fingerprinting technique is capable to locate and track MT with high
accuracy. The results also show that empirical approach is superior in terms of
accuracy and the main advantage of modeling approach is easier system deployment.
The basic RADAR system was later enhanced and improved [5]. The
enhanced version introduced AP-based environment profiling scheme, where
different radio map will be used for different environment such as “busy hour” and
“non-busy hour” environment. The accuracy of the system was improved more than
three times with this scheme. Viterbi-like algorithm was also included to improve the
continuous MT tracking performance. This algorithm successfully improved the
accuracy by over 33%.
Research in [28] introduced a system similar to RADAR except that RSS
values from other APs installed at other floors were also included in the radio map.
With more APs, the dimension of the location fingerprint was increased and thus the
accuracy was also increased more than three times compared to RADAR.
47
Work in [22] developed a WLAN RSS location fingerprint indoor LES based
on IEEE 802.11b WLAN. Empirical approach was taken to construct the radio map.
They compared the performance of three algorithms namely the NNSS, back
propagation neural network and histogram matching algorithms. From the results
obtained, the back propagation neural network is better than the others in terms of
accuracy and precision since it was more complex. The neural network based system
gave an accuracy of less than 1 meter with 72% of precision.
Research in [29] proposed a WLAN-based indoor LES for a library
environment. The system was designed to locate a static person carrying a PDA
within a floor of a multi-floor building. In this work, performance of four different
location estimation algorithms were studied and compared. The algorithms studied
were RSS location fingerprinting, weighted center-of-gravity, triangulation and
smallest M-vertex polygon algorithms. From the results obtained, the RSS location
fingerprinting technique outperformed the other algorithms. By using 3-NNSS
algorithms, RSS location fingerprinting is capable to provide accuracy of 3.15 meters
averagely.
Work in [30] also studied the feasibility of using WLAN RSS location
fingerprinting technique to estimate MT’s location. Instead of using NNSS algorithm,
this system performed the location estimation by comparing the location fingerprint
using normal curve function. The results shown that, in three out of five tests the
MT’s location were estimated correctly. Others location estimations were not correct
but still in the correct room or area.
The following section will review works related to probabilistic approach.
Nibble [31] used Bayesian network to estimate the MT’s location probabilistically.
Instead using RSS values, Nibble used SNR as location fingerprint. The strength of
Nibble is that its modular structure provides flexibility when APs are added, removed
or spatially relocated. In addition, contextual information such as the usual
movement behavior of the MT, MT’s preferences for locations and likelihood of the
MT to appear in a particular location can be easily incorporated in the a priori
probability of the Bayesian network. These advantages improved the accuracy of
48
location estimation. In a test on twelve adjacent locations along a hallway, Nibble
achieved an accuracy of 97% on correct estimated locations.
Work in [6] is one of the key researches in probabilistic RSS location
fingerprinting system. This research applied the knowledge from robotics field to
solve the problems in WLAN-based LES such as non-Gaussian signal, noise, strong
correlations due to multi-path effects and interferences. By applying probabilistic
Bayesian inference of location and hidden Markov Model for tracking purposes, the
proposed system proved that off-the-shelf WLAN infrastructures are reliable to
perform indoor location estimation with high accuracy, provided after careful system
deployment and adequate off-line calibration. The proposed system achieved 2D
location estimation accuracy of 1.5 meters over 80% of the time for stationary MT
and 1 meter over 50% of the time for mobile MT.
Horus [32] is another successful probabilistic RSS location fingerprinting
system. Horus was designed to achieve high accuracy and low computational
requirements. Besides basic probabilistic location estimation algorithm, this work
studied causes for wireless channel variations and addressed the problems with
various approaches to achieve high accuracy. Correlation among consecutive RSS
samples from AP is introduced to enhance the system performance [33]. Small-scale
compensation using perturbation technique was introduced to handle the small-scale
variation problem in wireless channel [34]. Indoor locations are pre-clustered into
small groups to reduce the computational cost of searching the radio map via location
clustering technique [35]. With these enhancements, Horus is compared with
RADAR system in two test-beds. Results shown the average accuracy of Horus was
better than RADAR by more than 89% and 82% respectively in both test-bed.
Research in [7] proposed an improvement to the system proposed by work in
[6]. Although proposed system in [6] was capable to provide high accuracy of 1.5
meters with the probability of 77% in hallways, this system often experiences logical
errors such as MT estimated at the wrong side of a cubicles or a wall. Therefore,
proposed system in [7] extended the same Markov localization framework in [6] and
incorporated system dynamics and motion constraint which indirectly encode the
49
physical characteristics of the indoor environment into the basic system. With these
enhancements, the proposed system was capable to maintain the accuracy of [6] but
with minimum logical errors when tested on a singe-floor environment.
Research in [36] proposed a model-based RSS distribution training scheme to
trade-off the accuracy of RSS distribution and off-line workload. A tracking-assistant
location estimation algorithm which uses topological knowledge was incorporated to
improve the accuracy.
Most of the previous works are only evaluated for 2D location estimation.
Some previous researches [4, 5, 7] predicted that their systems should work well in
3D location estimation, but no actual results for multi-floor environment were shown.
Efforts in [37] proposed and tested a system for multi-floor environment using IEEE
802.11b WLAN. However, this system requires extra specialized infrastructures such
as stationary emitters and sniffers and it was a network-based system. Work in [38]
proposed a multi-floor indoor LES based on IEEE 802.11b WLAN. This networkbased system applied deterministic approach using NNSS algorithm to estimate user
in a three-floor building. The results obtained show that WLAN-based location
estimation in multi-floor environment is feasible.
50
3.7
Summary
IEEE 802.11 WLAN and WLAN-based indoor location estimation
technology are overviewed in this chapter. Based on the literature review on WLANbased indoor LES and related researches, the following general observations can be
made:
(i)
WLAN-based indoor LES is feasible and reliable to provide indoor location
estimation with accuracy of few meters or the size of a small room.
(ii)
WLAN RSS location fingerprinting is more practical than proximity sensing
and triangulation techniques.
(iii)
Radio map constructed with empirical approach is more accurate than
mathematical modeling approach.
(iv)
Location estimation using probabilistic algorithm is more accurate than
deterministic algorithm.
Based on these positive observations, the proposed MILES is designed and
developed using WLAN RSS location fingerprinting technique. This study differs
from previous studies in few aspects. The proposed system concentrates on multifloor indoor location estimation while majority of previous studies only evaluated
their systems on single-floor indoor location estimation. Besides providing physical
location information, the proposed system also provides symbolic location
information. The proposed system is developed on top of IEEE 802.11g
infrastructures compared to IEEE 802.11b or proprietary infrastructures in most of
the previous studies. The study not only proposed a MILES, but also conducted a
study on RSS characteristics and also study on various factors affecting the
performance of WLAN-based MILES.
The following chapter will present the methodology on design and
development of the proposed system.
CHAPTER 4
METHODOLOGY
4.1
Overview of the Proposed System
Based on literature review and studies on previous researches, a WLANbased MILES is proposed. The proposed system is capable to locate a MT in a multifloor building using only off-the-shelf WLAN infrastructures. The general system
architecture and operation of the proposed system is depicted in Figure 4.1. Standard
off-the-shelf IEEE 802.11g WLAN APs and client device are used as the wireless
technology to enable the location estimation. The MT is a WLAN-enabled tablet PC
carried by a human user.
Off-Line Phase
RSS
IEEE 802.11g
Access Points
Tablet PC
(IEEE 802.11g
enabled)
Radio
Map
Empirical
Approach
Probabilistic
Bayesian
Filtering
Location
Estimated
On-Line Phase
Figure 4.1: Basic architecture of the proposed system
52
The proposed system uses WLAN RSS location fingerprinting technique to
estimate 3D location within a two-floor building. Operation of the proposed system
consists of two phases: off-line phase and on-line phase. During off-line phase,
empirical approach is adopted to construct the radio map. The MT moves to a finite
pre-determined RLs throughout the indoor area. At each RL, RSS from the APs in
the vicinity are collected and stored in radio map.
During on-line phase, the MT measures the current RSSs from APs in the
vicinity. These current RSSs are compared to the radio map using probabilistic
Bayesian filtering algorithm. The brief flow charts on these two phases are shown in
Figure 4.2 and Figure 4.3 respectively.
START
Start the off-line calibration application
Move to first reference location
Collect RSS values (location fingerprint) from
access points in the vicinity
Store the RSS values (location fingerprint) in
radio map together with current reference
location’s physical coordinate
No
Last reference
location ?
Move to next reference
location
Figure 4.2: Flow chart of off-line phase
Yes
END
53
START
Start the on-line location estimation application
Collect current RSS values from access points
in the vicinity
Compare current RSS values with location
fingerprints in the radio map
Estimate current location through probabilistic
Bayesian filtering algorithm
Display estimated current location
Yes
Continue?
No
END
Figure 4.3: Flow chart of on-line phase
54
4.2
Design and Development Processes
In order to achieve the objectives of this study, the design and development
processes are divided into eight inter-related stages as shown in Figure 4.4.
The design process started with identification of technical specifications for
the proposed system. Next, suitable test-bed, hardware and software components are
identified. Then, the test-bed is setup together with the required hardware and
software.
Since the proposed system is using WLAN RSS as location fingerprints, a
basic understanding of the characteristics of WLAN RSS in indoor environment is
important. Therefore, before design and developing the proposed system, the
characteristics of WLAN RSS for indoor location estimation application are studied.
A series of experiments on WLAN RSS are carried out. Results obtained from this
study are used in design, development and deployment stages of the proposed system.
Next, the proposed system architecture and the location estimation algorithm
are designed. Then, the proposed system is developed to fulfill the design
specifications identified in the initial stage.
Finally, the performance of the proposed system developed is validated at the
test bed through real-time experiments. The performance of the proposed system is
compared and discussed referring to the design specifications.
Besides that, a series of experiments are conducted to investigate various
factors that affect the performance of the proposed system. The four factors
investigated are number of WLAN APs installed, number of topological nodes,
number of RSS samples collected per location fingerprint and off-line phase
sampling interval.
55
Objective (i)
Determine technical specifications for
the proposed system
To study the
characteristics of
WLAN RSS in indoor
environment through
measurement for
WLAN-based indoor
location estimation
application
Identify test-bed, hardware and software
Objective (ii)
Setup test bed
Study on WLAN RSS characteristics in
indoor environment for
location estimation application
To design and develop a
WLAN-based indoor
location estimation
system for multi-floor
environment using RSS
location fingerprinting
technique
Design the proposed system architecture
and location estimation algorithm
Objective (iii)
Develop the proposed system
Evaluate the performance of the proposed
system through real-time experiments
Investigate various factors affecting the
performance of the proposed system
through experiments
To evaluate the
performance of the
proposed WLAN-based
multi-floor indoor
location estimation
system
Objective (iv)
To study factors which
affect the performance
of the proposed WLANbased multi-floor indoor
location estimation
system
Figure 4.4: Design and development process of the proposed system
56
4.3
Proposed System Specifications
The proposed system is aimed to support LBSs for multi-floor home and
office indoor environments. It is terminal-based software where the MT will
determine its own location. The location estimation technique chosen is location
fingerprinting. IEEE 802.11g WLAN RSS will be used as the location fingerprints.
Location estimation algorithm adopted is probabilistic Bayesian filtering algorithm.
Estimated location will be reported in terms of physical and symbolic
locations. The physical location will be represented with 3D coordinates (x, y, floor).
The symbolic location will be given as logical area using descriptive names.
The proposed system is capable to determine the location of the MT while it
is stationary and also mobile. For both stationary and mobile location estimation, the
targeted accuracy is less than 5 meters for 90% of the location estimation. Besides
that, it is aimed to provide correct floor estimation for 90% of the floor estimations.
Table 4.1: Specifications of the proposed system
Characteristic
Proposed Specification
Location estimation technique
Single or multi-floor indoor
environment (Home and office)
Location Fingerprinting
Location Fingerprint
WLAN RSS
Location sensing party
Mobile Terminal
Location calculation party
Mobile Terminal
Packet beacons scanning
Passive scanning
Off-line calibration approach
Empirical
Location estimation algorithm
Bayesian filtering (Probabilistic)
Type of location report
Physical and symbolic location
Target Movement
Stationary and mobile
Accuracy and Precision
90% less than 5 meters
Percentage of correct floor estimation
90%
Operating environment
57
4.4
Experimental Test Bed
All experiments in this study were conducted inside the two-floor building of
WCC, UTM. The floor plans of the test bed are shown in Figure 4.5 and Figure 4.6
respectively. The dimension of floor 1 is approximately 40 meters x 40 meters and
dimension of floor 2 is approximately 30 meters x 30 meters. There is only one
staircase connecting between floor 1 and floor 2. Majority of the building’s walls are
concrete wall. Some of the walls are made of plaster partition board. The test bed is a
typical office environment consisting of rooms, laboratories, open area and corridors.
L3
A1
A12
A2
A10
A11
A3
L4
A5
A4
L5
A6
L8
L1
A8
Legend:
L6
A7
A9
Access Point
Path_1a
Figure 4.5: Floor 1 of WCC, UTM and the locations of the access points
58
B1
B2
B3 L9
Legend:
L2
Access Point
B4
L7
Figure 4.6: Floor 2 of WCC, UTM and the locations of the access points
4.5
Experimental Hardware
The main hardware used in this study are WLAN APs, WLAN client device
and a mobile computer. The type of AP used for this study is DLink DWL-2000AP+
as show in Figure 4.7(a). This AP supports IEEE 802.11b and IEEE 802.11g
standards. The operating frequency range is from 2.4 GHz to 2.4835 GHz. The
theoretical wireless coverage range is 100 meters in indoor environment. Each AP is
equipped with one off-the-shelf external antenna.
59
(a)
(b)
Figure 4.7: Hardware used in the proposed system. (a) DLink DWL-2000AP+
WLAN access point (b) HP Compaq tc1100 Tablet PC
A total of sixteen APs are installed at the test bed. The installed location and
name of all APs are illustrated in Figure 4.5 and Figure 4.6. A total of twelve APs:
A1 to A12, are located at floor 1 and four APs: B1 to B4, are located at floor 2. All
the APs are installed against the walls and near to the ceiling. The APs’ antennas are
positioned vertically pointing to the ceiling. The variation in AP and its’ antenna
position may change the wireless coverage. Therefore, during this study, the position
of the APs and its’ antenna are fixed.
The operating channels for the sixteen APs are listed in Table 4.2. The
operating channels are selected to minimize the co-channel interference among APs.
During the experiments, besides the APs mentioned above, there are other
surrounding active APs detected. RSS from these surrounding APs are not used in
the system.
The MT used in this study is HP Compaq tc1100 Tablet PC as shown in
Figure 4.7(b). It is embedded with Intel PRO/Wireless 2200BG MiniPCI WLAN
client device. The operating system for the tablet PC is Microsoft Windows XP
Tablet PC edition Service Pack 2. The WLAN client device support IEEE 802.11b
and IEEE 802.11g standards.
60
Table 4.2: Channel allocation for access points installed in WCC, UTM
Location
Name of Access Point
Channel Number
A1
A2
A3
A4
A5
A6
A7
A8
A9
A10
A11
1
6
11
6
1
11
1
6
11
11
1
A12
6
B1
B2
B3
1
1
6
B4
11
Floor 1
Floor 2
4.6
Experimental Software
The proposed system is a software-based MILES. The software is written
with Java programming language. Java is designed by Sun Microsystem and freely
downloadable from the Internet. It is an object-oriented programming language that
is becoming increasingly useful in the world of software design. The main reason
Java is used as the programming language in this study is the simplicity of the
language and also Java is a cross-platform programming language. Software
application written with Java can works on multiple platform such as Microsoft
Windows, Linux and Mac OS X. In this study, the software is developed in
Microsoft Windows platform. With the cross-platform feature, the proposed system
should be able to run on other platforms such as Linux. However, in this study, the
proposed system only tested under Microsoft Windows platform.
61
In order to study the characteristics of WLAN RSS in indoor environment,
NetStumbler (version 0.4.0 Build 554) is used as the software to capture the WLAN
packet beacons from the APs and to provide WLAN RSS reading. NetStumbler is a
WLAN site survey software freely downloadable from the Internet. Snapshot of
Netstumbler GUI is illustrated in Figure 4.8. Netstumbler provides WLAN network
information such as service set identifier (SSID), RSS, noise level, signal to noise
ratio and security scheme. During the experiments, the data captured in NetStumbler
are exported into text files for post-processing in Microsoft Excel.
Microsoft Excel is used as the statistical analysis tool to analyze the results
obtained from all the experiments in this study. It is used to calculate the statistics
and plotting the statistical figures.
Figure 4.8: Snapshot of Netstumbler software GUI.
62
4.7
Summary
Based on the literature reviews, the basic architecture and operation of the
proposed WLAN-based MILES are presented in this chapter. WLAN RSS location
fingerprinting technique is proposed to estimate MT’s location in a multi-floor
indoor environment. Off-line phase is conducted through empirical approach.
Bayesian filtering is chosen as the probabilistic location estimation algorithm.
The design and development processes involved in this study are described
here. These processes are conducted in sequence in order to achieve the objectives of
this study. The targeted technical specifications for the proposed system are
determined. Experimental test-bed, hardware and software components needed are
identified and setup.
The next process is to study the characteristics of WLAN RSS in multi-floor
indoor environment for location estimation application. Therefore, this issue is
discussed in the next chapter.
CHAPTER 5
CHARACTERISTICS OF WLAN RECEIVED SIGNAL STRENGTH
5.1
Introduction
Basic understanding on the characteristics of WLAN RSS will assist in the
design, development and deployment of the WLAN MILES. Although there is an
extensive knowledge available regarding the RF phenomenon and characteristics of
RSS in indoor environment, such knowledge is aimed towards other applications,
such as transceiver design and determining communication capabilities.
Therefore, in this chapter, characteristics of IEEE 802.11g WLAN RSS in
indoor environment for location fingerprinting application will be studied through
measurements. Results obtained here are used as guideline in design and deployment
of the proposed MILES later.
The goals of the measurements in this chapter are:
(i)
To show the feasibility of using WLAN RSS as location fingerprint to
estimate indoor location.
(ii)
To study the characteristics of WLAN RSS in indoor environment. Results
obtained are used in design and development of the proposed system.
(iii)
To investigate the environmental factors affecting the characteristics of
WLAN RSS in indoor environment. Results obtained are used in deployment
of the proposed system.
64
In order to achieve these objectives, a series of measurements are conducted. The
objectives and related studies are listed in Table 5.1. The characteristics studied here
are by no means exhaustive. Other factors such as types of antenna are not
considered here. The details and results from these studies will be discussed further
in the following sections.
Table 5.1: Objectives and types of studies on WLAN RSS characteristics
Objective
Objective (i)
To show the feasibility of
using WLAN RSS as
location fingerprint to
estimate indoor location.
Type of study
Location dependency of WLAN RSS
WLAN RSS distribution at a stationary indoor location
Objective (ii)
To study the characteristics
of WLAN RSS in indoor
environment.
Statistical representation of WLAN RSS distribution at
a stationary indoor location
Variation of numbers of APs detected at a stationary
indoor location
Effects of user proximity
Objective (iii)
To investigate the
environmental factors
affecting the characteristics
of WLAN RSS in indoor
environment.
Effects of user’s orientation
Effects of different environments
Effects of types of WLAN client device
Effects of multi-floor environment
65
5.2
Location Dependency of WLAN RSS
The feasibility of using WLAN RSS as location fingerprint to estimate indoor
location is investigated here. RSSs from multiple APs are measured at different
locations as the MT traveled clockwise around hallway Path_1a in the test-bed as
shown in Figure 4.5. Measurement is taken at 46 locations along Path_1a and
distance between consecutive locations is averagely 1 meter. At each measurement
location, 100 RSS samples from each detected AP are recorded with sampling
interval of 1 second. Average RSS from each AP is calculated. Measurement is
conducted at midnight where movement of people in the test-bed is minimal. During
the measurement, RSS from fifteen to sixteen APs are recorded at each location. For
clarity, only RSS results from three APs (A3, A7 and A11) are presented here. Figure
5.1 shows how the average RSS from these three APs varies as the MT traveled
along hallway Path_1a.
Distance Traveled (m)
0
5
10
15
20
25
30
35
40
45
0
-10
A7
A3
A11
Average RSS (dBm)
-20
-30
-40
-50
-60
-70
-80
-90
-100
Location of
Access Point
A7
Location of
Access Point
A3
Location of
Access Point
A11
Figure 5.1: Average RSS from three access points at different locations as MT
traveled along hallway Path_1a
66
From Figure 5.1, there is a trend in the average RSS measured from the APs
as the MT traveled along Path_1a. The average RSS from a given AP is strongest
when the MT is close to it and weakest when the MT is far away.
The combination of average RSS from different APs at one location is called
RSS tuple. From Figure 5.1, the RSS tuple at each measurement location is unique.
The RSS tuples at selected distance traveled in Figure 5.1 are listed in Table 5.2.
From Table 5.2, no measurement locations shared the same RSS tuple. This indicates
that RSS from multiple APs can be used as a fingerprint to infer location uniquely.
Observations from Figure 5.1 and Table 5.2 also show that values of RSS
from APs varied slowly against the distance traveled. The RSSs measured from APs
remain approximately same over short distance. For instance, average RSS measured
from AP A7, A3 and A11 at 5 meters and 6 meters only varied in 1.7 dB, 1.1 dB and
2.5 dB respectively. In contrast, RSSs measured from APs over longer distance are
varied greatly. Average RSS measured from AP A7, A3 and A11 at 0 meter and 10
meters varied in 38.5 dB, 22.0 dB and 14.2 dB respectively. This suggests that the
spatial separation between RLs to collect RSS fingerprints should not be too close to
each other.
67
Table 5.2: Average RSS tuple from three access points at different measurement
locations
Average RSS from AP (dBm)
Distance Traveled
(m)
A7
A3
A11
0
-36.3
-82.0
-86.6
1
-50.2
-82.2
-85.0
2
-44.1
-77.9
-81.1
3
-46.6
-80.1
-83.4
4
-51.1
-81.0
-79.6
5
-55.5
-71.8
-79.4
6
-57.2
-72.9
-81.9
7
-70.4
-70.8
-79.1
8
-65.2
-67.6
-83.2
9
-66.7
-75.7
-72.7
10
-74.8
-60.0
-72.4
20
-81.3
-46.6
-62.9
30
-85.0
-67.1
-48.2
40
-76.1
-76.1
-57.3
5.3
WLAN RSS Distribution at a Stationary Indoor Location
The WLAN RSS distribution at a stationary location in indoor environment is
investigated here. RSS measurement is taken at location L1 as shown in Figure 4.5.
Measurement is taken for 12 hours (10:10 a.m. to 10.10 p.m.) with sampling interval
of 2 seconds. This measurement period include office hour where movement of
people and opening or closing of doors happened frequently and also off-office hour
where the test-bed is empty. This will show WLAN RSS distribution in a typical
office environment. During the measurement, a total of sixteen APs are detected
simultaneously. In the following section, only results from AP A7, A11 from floor 1
and AP B3 from floor 2 will be discussed. At location L1, the WLAN client device’s
antenna has a LOS path to AP A7 and No Line of Sight (NLOS) to AP A11 and B3.
The RSS measurement results for these three APs are shown in Figure 5.2.
RSS (dBm)
12 hours at location L1
-75
21
20
19
18
17
15
14
13
12
11
10
:3
3
:2
5
:1
7
:0
8
:0
0
:5
1
:4
3
:3
4
:2
6
:1
8
:1
0
:4
0
:2
6
:0
0
:2
4
:0
4
:3
6
:1
4
:5
8
:5
4
:2
0
:0
0
RSS (dBm)
10
:1
0:
00
11
:1
8:
16
12
:2
6:
26
13
:3
4:
32
14
:4
3:
08
15
:5
1:
40
17
:0
0:
12
18
:0
8:
24
19
:1
6:
50
20
:2
5:
20
21
:3
3:
34
RSS (dBm)
:1
0:
00
11
:1
8:
10
12
:2
6:
42
13
:3
4:
48
14
:4
3:
04
15
:5
1:
26
16
:5
9:
54
18
:0
8:
14
19
:1
6:
48
20
:2
5:
16
21
:3
3:
30
10
68
-30
-35
-40
-45
-50
-55
-60
-65
-70
(a)
Time (hh:mm:ss)
-60
-65
-70
-75
(b)
-80
-85
-90
Time (hh:mm:ss)
-60
-65
-70
(c)
-80
-85
-90
Time (hh:mm:ss)
Figure 5.2: RSS measured from access point (a) A7 (b) A11 (c) B3 for duration of
69
From Figure 5.2, the RSSs measured over time from the three APs at
stationary location L1 are not stable all the time, but fluctuate around a mean value.
Due to reflection, refraction, diffraction, absorption and scattering, WLAN signal
propagation in indoor environment suffers from severe multi-path fading effects [39].
Changes in indoor environment such as movement of people, doors opening and
closing contribute to multi-path fading. Besides that, in the 2.4 GHz band,
microwave ovens, Bluetooth devices and cordless phones can be sources of
interference. As a result, transmitted signal usually reach the receiver through
different paths with each having its own amplitude and phase. Multi-path fading
causes the signal received fluctuate over its mean value at a given location [40].
The standard deviations for the RSS measurements in Figure 5.2 are
calculated. Standard deviation for RSS distributions measured from AP A7, A1 and
B3 are 4.2 dB, 2.0 dB, 2.7 dB respectively. AP A7 which has a LOS path to the
receiver at location L1 has higher standard variation value compared to AP A11 and
B3. This suggests that, at a given stationary location, value of RSS from AP with
LOS path to the receiver fluctuate more compared to AP with NLOS path. Similar
result is reported by works in [40].
Measurements in this section suggest that a single RSS value is insufficient to
represent the RSS distribution at a given stationary indoor location. A number of
RSS values should be collected in order to enable representation of RSS distribution
at a given stationary indoor location through statistical distribution.
70
5.4
Statistical Representation of WLAN RSS Distribution at a Stationary
Indoor Location
RSS distribution at a stationary location in an indoor environment is site
specific and difficult to predict. A general statistical model which models the
distribution accurately is not available. Although many studies have been made to
model radio signal distribution in an indoor environment, different studies suggested
different distribution models [6].
The statistical distribution of RSSs measured from AP A7, A11 and B3 in
section 5.3 are plotted in Figure 5.3. From histograms in Figure 5.3, the RSS
distributions at a stationary location for AP A7, A11 and B3 are obviously different
from each other. In general, the distributions are asymmetric and have multiple
modes. For each distribution, the dominant mode is different from the mean value. In
addition, the value of mean, median and mode for each distribution are different.
These observations suggest that RSS distribution in indoor environment measured
with off-the-shelf WLAN client devices are essentially non-Gaussian. Similar
observations are reported in [6, 40].
Since the RSS distribution at a stationary location is non-Gaussian, using
mean or average of RSS values as location fingerprint will not be accurate.
Therefore, using the histogram of RSS distribution collected directly rather than
post-process the data into mean value as location fingerprint for a given indoor
location is more reliable and accurate.
71
Probability
0.20
Mean = -44.9 dBm
Median = -45.0 dBm
Mode = -40.0 dBm
0.15
(a)
0.10
0.05
0.00
-70
-65
-60
-55
-50
-45
-40
-35
RSS (dBm)
Percentage
0.50
0.40
Mean = -81.8 dBm
Median = -83.0 dBm
Mode = -83.0 dBm
0.30
(b)
0.20
0.10
0.00
-90
-85
-80
-75
-70
-65
-6
RSS (dBm)
Percentage
0.20
Mean = -74.5 dBm
Median = -74.0 dBm
Mode = -73.0 dBm
0.15
0.10
(c)
0.05
0.00
-85
-80
-75
-70
-65
RSS (dBm)
Figure 5.3: Histograms of RSS distribution for access point (a) A7 (b) A11 and (c)
B3 measured for duration of 12 hours at location L1
72
In order to construct a reliable histogram that represents the real RSS
distribution for a given location, the number of RSS samples collected is important.
Here, the number of samples required to construct a reliable histogram which
represent the RSS distribution at a given stationary location is studied. The RSS
measurement collected in Section 5.3 will be used as the real RSS distribution.
Histograms and statistics of RSS received from AP A7 are calculated with different
numbers of samples ranging from 30, 50, 100, 200 and 300. By assuming the RSSs
collected in Section 5.3 as the accurate RSS distribution of AP A7 at location L1, the
RSS histograms and statistics are compared. The results are shown in Table 5.3 and
Figure 5.4.
Table 5.3: Summary of statistics for RSS distribution from access point A7 at
location L1 with different numbers of samples
Number of RSS Samples
Statistic
30
50
-48.1
-48.4
-48.2
-48.1
-48.2
2.9
2.9
2.8
3.1
3.2
Median (dBm)
-47.0
-47.0
-47.0
-47.0
-47.0
Mode (dBm)
-46.0
-46.0
-46.0
-46.0
-46.0
60
100
200
400
600
Mean (dBm)
Standard Deviation (dB)
100
200
300
Total Sampling Time (s)
[Number of Sample ÷ 0.5
sample/s]
73
Probability
0.40
0.30
(a)
0.20
0.10
0.00
-70
-65
-60
-50
-45
-40
-35
RSS (dBm)
0.40
Probability
-55
0.30
(b)
0.20
0.10
0.00
-70
-65
-60
-50
-45
-40
-35
RSS (dBm)
0.40
Probability
-55
0.30
(c)
0.20
0.10
0.00
-70
-65
-60
-50
-45
-40
-35
RSS (dBm)
0.40
Probability
-55
0.30
(d)
0.20
0.10
0.00
-70
-65
-60
-50
-45
-40
-35
RSS (dBm)
0.40
Probability
-55
0.30
(e)
0.20
0.10
0.00
-70
-65
-60
-55
-50
-45
-40
-35
RSS (dBm)
Figure 5.4: Histograms of RSS distribution for access point A7 at location L1
calculated with (a) 30 (b) 50 (c) 100 (d) 200 (e) 300 RSS samples
74
From Table 5.3 and Figure 5.4, the summary of statistics and distribution in
RSS histograms are almost similar for all the numbers of samples studied. This
indicates that a small numbers of RSS samples, 30 and 50 samples in this case, are
sufficient to construct the histogram to represent the RSS distribution at a given
location. Smaller numbers of RSS samples is an advantage for indoor LES based on
empirical location fingerprinting. This is because with smaller number of samples
required, the time needed to collect location fingerprints will reduce. From Table 5.3,
the time spend to collect 30 samples is 1 minute compared to 10 minutes for 300
samples.
5.5
Variation of Numbers of APs Detected at a Stationary Indoor Location
Passive scanning can be used to collect the RSS tuples at RLs. The weakness
of passive scanning is that the detection of beacon frame from APs is not guaranteed.
Possible reasons for failure of detecting beacon frame from a given AP are the
beacon frame lost in propagation path or the signal strength or the beacon frames are
too low to be detected and decoded by the WLAN client device.
The number of APs detected by WLAN client device against measurement
time using passive scanning will be investigated here. RSS measurement is taken at
location L1 for a period of 2 minutes with sampling interval of 2 seconds. Figure 5.5
shows the number of APs detected by the WLAN client device for the 2 minutes RSS
measurement. Observation from Figure 5.5 shows that the number of APs detected
by the WLAN client device is not fixed but varied with time.
The numbers of RSS samples collected from all APs during the 2 minutes
measurement are listed in Table 5.4. Using 2 minutes measurement period with
sampling interval of 2 seconds, the total RSS samples collected should be 60.
However, referring to Table 5.4, some of the beacon frames were not detected due to
unknown reasons.
Number of AP Detected
75
18
16
14
12
10
8
6
4
2
0
0
5
10
15
20
25
30
35
40
45
50
55
Measurement
Figure 5.5: Number of access points detected by WLAN client device at location L1
over 60 RSS measurements
Table 5.4: Numbers of RSS samples collected from access points at location L1 over
60 RSS measurements
AP
Number of Samples
AP
Number of Samples
A1
59
A9
59
A2
59
A10
35
A3
59
A11
60
A4
59
A12
51
A5
59
B1
60
A6
57
B2
59
A7
60
B3
59
A8
60
B4
59
76
1
11
21
31
41
51
-30
RSS (dBm)
-40
-50
-60
-70
-80
-90
Measurement
A7
A10
A12
B3
A8
Figure 5.6: RSSs measured from access point A7, A8, A10, A12 and B3 at location
L1 over 60 RSS measurements
Figure 5.6 gives a clearer picture of APs detected during the 2 minutes RSS
measurement. Only RSSs measured from AP A7, A8, A10, A12 and B3 are shown in
Figure 5.6. From Figure 5.6, not every AP in the range will be detected by the
wireless client device at every measurement. For instance, referring to Figure 5.6, AP
A10 was always undetected. This may be due to the multi-path phenomena and also
the structure o the building which caused the beacon frames from the AP undetected.
Measurements in this section suggest that RSS measurements from some
AP’s at a given location are not always reliable. The numbers of RSS samples
collected from those APs may not be sufficient to be used as location fingerprints.
For instance, the total number of RSS samples from AP A10 measured in this
experiment is only approximately half of the samples needed, about 35 out of 60
samples. The RSS samples from AP A10 in this case may not be reliable to be used
as location fingerprint. Therefore, in the design of indoor LES based on RSS location
fingerprinting, some mechanism or filter may be required to filter out unreliable AP’s
RSS measurements during both off-line and on-line phase.
77
5.6
Effects of User Proximity
In some LBSs using the WLAN-based indoor LES, the MT is usually carried
by a human user. The proximity of human user to the WLAN client device will affect
the RSS in a measurable way.
The effect of human user’s body on WLAN RSS will be investigated here.
Measurement is performed at location L2 inside a room as shown in Figure 4.6. At
location L2, the MT doesn’t have LOS to any APs in the test-bed. RSS from multiple
APs are recorded for 4 hours with sampling interval of 2 seconds. During the first 2
hours, measurement is taken without presence of the user. In the following 2 hours,
the user is present in the room sitting in front of the MT. The door of the room is
closed for both measurement periods. The experiment is conducted at night where
people movement is minimal. During the measurement, RSS from ten APs are
detected at location L2. For clarity, only RSS results from four APs: B3, B4, A5 and
A6, are presented in Figure 5.7.
Figure 5.7 depicts the difference between RSS distributions at the location L2
with and without user proximity. The average value for RSS distribution in Figure
5.7 is summarized in Figure 5.8.
From the observations, the presence of user’s body influenced the RSS
distribution in different ways. It may cause the RSS to decrease, increase or no
changes. At location L2, the presence of user caused average RSS value to decrease
for AP B3 and A6, increase for A5 and nearly no changes for B4. With presence of
user, the average RSS for AP B3 decreased 3.2 dB from -60.4 dBm to -63.6 dBm,
while the average RSS for AP A6 decreased 7.6 dB from -68.6 dBm to -76.2 dBm.
The average RSS for AP A5 increased 3.3 dB from -80.7 dBm to -77.4 dBm when
user is present. The average RSS for AP B3 remain nearly no changes as user
presence only decreased the value by 0.3 dB. A possible explanation of these
changes is the scattering and shadowing effects of human body in indoor
environment [41]. The user acts as an obstacle which obstructs the propagation path
between AP and MT and thus causes changes in the RSS value.
78
Absence of user
RSS (dBm)
0
RSS (dBm)
14400
(a)
7200
-50
-55
-60
-65
-70
14400
(b)
0
RSS (dBm)
7200
-60
-65
-70
-75
-80
0
7200
-42
-47
-52
-57
-62
14400
(c)
0
RSS (dBm)
Presence of user
7200
-70
-75
-80
-85
-90
14400
(d)
Time of Measurement (s)
Figure 5.7: RSS Measured from access point (a) A6 (b) B3 (c) B4 (d) A5 at location
L2 with user absence in the first two hours and presence for the following two hours
79
B3
Access Point
B4
A5
A6
Average RSS (dBm)
0.0
-20.0
-40.0
-53.6
-60.0
-80.0
-60.4
-63.6
-53.9
-68.6
-80.7 -77.4
-76.2
-100.0
User Absence
User Presence
Figure 5.8: Average RSS Measured from access point B3, B4, A5 and A6 at location
L2 with user absence in the first two hours and presence for the following two hours
Measurements in this section suggest that user proximity will bring
measurable changes in RSS value at a given location. Therefore, it is essential to
collect RSS location fingerprint based on the application scenario of the indoor LES.
If the system will be carried and used by human user, it is essential to collect RSS
location fingerprint with human user presence in order to take into consideration the
human body effect.
5.7
Effects of User’s Orientation
Human body consists of 70% of water and resonance frequency of water is
2.4 GHz [6]. Thus the WLAN signal will be absorbed when the human user’s body
obstructs the signal path and causing extra attenuation leading to lower RSS value.
Therefore, RSS value measured at one location can be varied when user facing
different orientations such as North, South, East and West.
80
The effect of user’s orientation on the RSS will be investigated here.
Measurements are performed at Location L3 and L4 as shown in Figure 4.5. RSS is
measured in “forward” and “backward” orientations. “Forward” is the orientation
when user walk in forward direction along a given path. “Backward” is the opposite
orientation of “forward”. At location L3, the forward orientation is the direction
facing AP A1. At location L4, the forward orientation is the direction facing AP A11.
RSS from APs are recorded for 5 minutes with sampling interval of 1 second at each
orientation. The experiment is conducted at night where people movement is
minimal. Results are shown in Table 5.5.
Table 5.5: Average RSS at location L3 and L4 with forward and backward
orientations
Location
L3
L4
Access Point
Average RSS (dBm)
Difference (dB)
Forward
Backward
|Forward - Backward|
A1
-56.0
-69.7
13.7
A3
-48.5
-42.7
5.8
A12
-55.8
-44.8
11.0
B1
-69.3
-70.8
1.5
B3
-82.5
-82.1
0.4
A4
-67.0
-73.1
6.1
A8
-74.1
-65.3
8.8
A11
-53.4
-63.8
10.4
B1
-73.7
-73.4
0.3
B3
-73.4
-76.7
3.3
81
At location L3, the WLAN client device’s antenna had a LOS to AP A1 and
A3 when user is facing forward orientation. When user is facing backward
orientation, the WLAN client device’s antenna only had LOS to AP A12. Referring
to Table 5.5, the average RSS measured from AP A1 at location L3 is -56.0 dBm
when user is facing forward orientation and decreased to -69.7 dBm when user is
facing backward orientation. This is mainly because in forward orientation, the
WLAN client device’s antenna has a clear LOS to AP A1, while in the backward
direction the LOS to AP A1 is blocked by user’s body. The average RSS is
attenuated by 13.7 dB due to the obstruction from the user’s body. The user’s body
also attenuated the average RSS from AP A3 and A12 by 5.8 dB and 11.0 dB
respectively when user’s orientation is changed.
At location L4, the WLAN client device’s antenna has a LOS to AP A11
when user is facing forward direction. When user is facing backward orientation, no
AP is in the LOS of the WLAN client device’s antenna. The average RSS from AP
A4, A8 and A11 is attenuated by 6.1 dB, 8.8 dB and 10.4 dB respectively due to the
obstruction from the user’s body.
In both location L3 and L4, variation in user’s orientation doesn’t affect
significantly the average RSS of APs installed in floor 2, AP B1 and B3. The
attenuation of average RSS for AP B1 and B2 is less than 4 dB for both location L3
and L4 when user’s orientation is changed. This is mainly because the APs are
located vertically on top of the user. User’s body doesn’t obstruct significantly their
signal paths when moved horizontally in both forward and backward orientations.
Measurements in this section suggest that user’s orientation affects the RSS at
a given location. Thus, RSSs measured from different orientations at a given location
can be treated as distinct location fingerprints.
82
5.8
Effects of Different Environments
In indoor environment, RF signal propagation is affected by various
environmental factors such as the presence and movement of people, opening or
closing of doors and furniture relocation.
The effects of environmental changes on the RSS will be investigated here.
RSS measurements are performed in two different environments: busy office and
empty office. The main differences between these two environments are the presence
and movement of people and opening and closing of doors.
Measurements are taken at location L1 as shown in Figure 4.5. Location L1 is
situated closed to the main entrance of the test bed. Movement of people and open
and closing of doors always happened here during busy office environment and are
minimal in the empty office environment. For each environment, RSS measurement
is taken for a period of 1 hour with sampling interval of 2 seconds. The measurement
for busy office environment is taken around lunch hour, 1:30 p.m. to 2:30 p.m.,
where there are around ten people moving around the test bed. Measurement for
empty office environment is taken after office hour, 8:00 p.m. to 9:00 p.m., where
there are no people in the test bed.
The RSS distributions measured from AP A7, A11 and B3 is presented in the
following section. At location L1, the WLAN client device’s antenna has a clear
LOS to AP A7 and NLOS to AP A11 and B3. Results of measurements are shown in
Figure 5.9, Figure 5.10 and Figure 5.11. The averages and standard deviations of the
RSS distributions for AP A7, A11 and B3 in both environments are summarized in
Table 5.6.
83
0
1200
2400
3600
RSS (dBm)
-35
-40
-45
(a)
-50
-55
-60
-65
RSS (dBm)
0
1200
2400
3600
-35
-40
-45
-50
-55
-60
-65
(b)
Time of Measurement (s)
Figure 5.9: RSS measured from access point A7 at location L1 during (a) Busy
office environment (b) Empty office environment
84
0
1200
2400
3600
RSS (dBm)
-70
-75
(a)
-80
-85
-90
0
1200
2400
3600
RSS (dBm)
-70
-75
(b)
-80
-85
-90
Time of Measurement (s)
Figure 5.10: RSS measured from access point A11 at location L1 during (a) Busy
office environment and (b) Empty office environment
85
0
1200
2400
3600
RSS (dBm)
-65
-70
(a)
-75
-80
-85
0
1200
2400
3600
RSS (dBm)
-65
-70
(b)
-75
-80
-85
Time of Measurement (s)
Figure 5.11: RSS measured from access point B3 at location L1 during (a) Busy
office environment and (b) Empty office environment
Table 5.6: Average and standard deviation of the RSS distributions for access point
A7, A11 and B3 at location L1 during busy office and empty office environments
Access Point
A7
A11
B3
Environment
Statistic
Average (dBm)
Standard Deviation (dB)
Average (dBm)
Standard Deviation (dB)
Average (dBm)
Standard Deviation (dB)
Busy Office
Empty Office
-47.8
-42.0
3.5
3.1
-80.6
-82.8
2.5
0.6
-73.9
-76.4
2.5
1.7
86
Figure 5.9, Figure 5.10 and Figure 5.11 show the comparisons between RSS
distributions during busy office and empty office environments for AP A7, A11 and
B3 respectively. From these figures, RSS distributions for busy office and empty
office environments are different. RSS distributions during empty office are more
stable compared to busy office. In addition, Table 5.6 shows that the standard
deviations for RSS distributions during busy office environment are higher compared
to empty office environment. This is mainly due to the changes in the environment.
Changes in environment condition such as presence and movement of people and
opening and closing of doors are likely to cause disturbances and higher variation or
fluctuation in RSS during busy office environment.
Besides fluctuation in RSS distribution, changes in environment condition
also affect the strength of signal. From study in [42], RSS is stronger during empty
office environment compared to busy environment. However, in this experiment, the
results obtained show that changes in environment may increase or decrease the
strength of the signal. Referring to Table 5.6, average RSS for AP 7 during empty
office is higher compared to busy office environment. In contrast, averages of RSS
for AP A11 and B3 are lower during empty office compared to busy office
environment.
Measurements in this section show that distribution of RSS is affected by
environmental conditions such as movement of people and opening and closing of
doors. This suggests that different radio maps should be collected based on
application scenario. If the indoor LES will be used during busy office environment,
then off-line phase should be conducted during busy office environment.
87
5.9
Effects of Types of WLAN Client Device
WLAN client devices are available in different interfaces such as Personal
Computer Memory Card International Association (PCMCIA), Universal Serial Bus
(USB) and MiniPCI card. Due to the different physical architecture, operating
environment and size, the receiver design and antenna characteristics for different
types of WLAN client device may vary. Characteristics such as type of antenna,
receiver sensitivity and transmit power will affect the value of RSS measured.
IEEE 802.11 standard defines a mechanism by which RF energy is to be
measured by the circuitry on a WLAN client device. It is measured as Received
Signal Strength Indicator (RSSI) which is an integer ranged from 0 to 255. It is used
internally by WLAN client device to report signal quality [43]. RSSI is intended to
be used in a relative manner and the absolute accuracy of RSSI reading is no
specified in IEEE 802.11 standards. Therefore, different vendors measure RSSI of
WLAN client device in a different ways. Instead of measuring 256 signal levels from
0 to 255, actual implementation of RSSI measurement for each vendor is limited to
value between 0 and a specific maximum RSSI value called RSSI_Max. For
instance, value of RSSI_Max for Cisco’s WLAN client device is 100, while Atheros
chipset uses RSSI_Max value of 60 and Symbol uses RSSI_Max value of 31.
The RSS value in dBm reported by most WLAN client devices is a
conversion of the RSSI value. The RSSI value can be mapped to an equivalent value
in dBm using a table or certain calculation. This mapping between RSSI and actual
RF energy also vary among vendors [44]. Due to the differences in measuring RSSI
and conversion of RSSI to RSS value in dBm among vendors, the RSS reported by
different WLAN client devices are different.
The effects of different types of WLAN client device on RSS value measured
at a given location will be investigated here. Five different IEEE 802.11b/g WLAN
client devices are used to measure RSS at location L5 as shown in Figure 4.5. Table
5.7 lists the WLAN client devices used in this measurement.
88
Table 5.7: List of IEEE 802.11b/g WLAN client device
Name
Vendor
Netgear_A1 Netgear
Netgear_A2 Netgear
Netgear_B
Netgear
D-Link_A
D-Link
Intel_A
Intel
Model
WG511
WG511
WG511T
DWL-G650+
Intel
PRO/Wireless
2200BG
Standard
Interface
IEEE 802.11b/g
32-bit Cardbus
IEEE 802.11b/g
32-bit Cardbus
IEEE 802.11b/g
32-bit Cardbus
IEEE 802.11b/g
32-bit Cardbus
IEEE 802.11b/g
Embedded
MiniPCI Type
3B
From Table 5.7, Netgear_A1, Netgear_A2, Netgear_B and D-Link_A are
WLAN client device with 32-bit Cardbus interface or better known as PC card.
These four devices are equipped with integrated internal diversity antenna. The types
and characteristics of the antenna are not clearly mentioned in the product data sheet.
Netgear_A1 and Netgear_A2 came from same vendor and model. Intel_A is the
MiniPCI Type 3B WLAN client device embedded in the HP Compaq tc1100 Tablet
PC used for the measurement in this study. The antenna of Intel_A is mounted in the
tablet computer’s display closure.
Measurements are taken at night to minimize people movement interference.
Each WLAN client device collected RSS from AP for a duration of 20 minutes with
sampling interval of 1 second. Ideally, the RSS measurements from all the WLAN
client devices should be taken simultaneously for more accurate results. But due to
limited number of MT, the RSS measurements for different WLAN client devices are
taken sequentially in this experiment. Although the results obtained may vary with
the results taken simultaneously due to the possible changes in measurement
environment, but it will be sufficient to prove that different types of WLAN client
device will report different RSS value at a given location. Results are shown in
Figure 5.12.
89
A3
Access Point
A8
B3
Average RSS (dBm)
0.00
-20.00
-40.00
-60.00
-80.00
-100.00
Netgear_A1
Netgear_A2
Netgear_B
D-Link_A
Intel_A
Figure 5.12: Average RSS measured from access point A3, A8 and B3 at location
L5 with five different WLAN client devices
Figure 5.12 depicts the average RSS measured from AP A3, A8 and B3 at
location L5 with five different WLAN client devices. At location L5, the antenna of
the WLAN client devices has a LOS to AP A3 and NLOS to AP A8 and B3. The
average RSSs for both Netgear_A1 and Netgear_A2 are nearly same for all the AP
shown in Figure 5.12. The differences in average RSS value for AP A3, A8 and B3
between Netgear_A1 and Netgear_A2 are 0.44 dB, 0.68 dB and 1.95 dB
respectively. The differences are small because both devices belong to the same
model and vendor. In contrast, the maximum difference in average RSS measured
from different devices for AP A3, A8 and B3 are 14.25 dB, 14.02 dB, 13.61 dB
respectively. These results suggest that WLAN client devices from same vendor and
model will give similar RSS reading, while WLAN client devices from different
vendors and models will give different RSS reading.
Measurements in this section indicate that different types of WLAN client
device will give different RSS measurement at a given location. This suggests that
same WLAN client device should be used in collecting location fingerprints and also
real-time location estimation in order to minimize location estimation error.
90
5.10
Effects of Multi-Floor Environment
Concrete floors in an indoor building attenuate RSS in a measurable way.
According to finding in [38], a single concrete floor can reduce RSS approximately
from 15 dB to 35 dB.
The effects of floor on RSS will be investigated here. In order to measure the
floor attenuation, two APs are placed in vertically symmetric locations on different
floor in the test-bed. The first set of measurement is performed at location L6 and L7,
while second set of measurement is conducted at location L8 and L9 as shown in
Figure 4.5 and Figure 4.6. Both locations in these two measurements are vertically
symmetric locations on different floors, where location L6 and L8 are on first floor
and L7 and L9 are on second floor. At both locations L6 and L7, a total of 500 RSS
samples are recorded from AP A7 and B4 respectively, with sampling interval of 1
second. On the other hand, a total of 500 RSS samples are recorded from AP A4 and
B3 at both location L8 and L9. Measurements are conducted at night where people
movement is minimal. Results are shown in Figure 5.13 and Figure 5.14.
Figure 5.13 and Figure 5.14 illustrate the RSS distributions on two different
floors for AP A7, B4, A4 and B3 respectively. For each distribution, there are two
horizontal regions that are separated from each other with at least 10 dB differences.
RSS values measured at the floor where the AP is installed are higher than the values
measured at the vertically symmetric location on the next floor. This difference is
caused by the floor attenuation.
91
1
101
201
301
401
-30
RSS (dBm)
-40
-50
(a)
-60
-70
-80
-90
1
101
201
301
401
-30
RSS (dBm)
-40
-50
(b)
-60
-70
-80
-90
Time of Measurement (s)
X
Floor 1 (L6)
Floor 2 (L7)
Figure 5.13: RSS measured at different floors from access point (a) A7 and (b) B4 at
location L6 and L7 respectively
92
1
101
201
301
401
-30
RSS (dBm)
-40
-50
(a)
-60
-70
-80
-90
1
101
201
301
401
-30
RSS (dBm)
-40
-50
-60
(b)
-70
-80
-90
Time of Measurement (s)
X
Floor 1 (L8)
Floor 2 (L9)
Figure 5.14: RSS measured at different floors from access point (a) A4 and (b) B3 at
location L8 and L9 respectively
Average and difference of the RSS distributions in Figure 5.13 and Figure
5.14 are summarized in Figure 5.15. From Figure 5.15, the minimum average floor
attenuation is 13.9 dB and the maximum is 24.9 dB. AP A7 and B4 are placed
approximately in vertically symmetric locations on floor 1 and floor 2 respectively.
Although signals from both AP experienced the attenuation from the same concrete
floor, but the amount of average floor attenuation experienced by AP A7 and B4 are
different as shown in Figure 5.15. The same situation happened to AP A4 and B3.
93
This suggests that the value of floor attenuation is not fixed but may varied according
other factors such as different propagation effects at different floors.
In addition, from Figure 5.15, the average floor attenuation values for AP A4
and B3 are less than values for AP A7 and B4. This is mainly due to AP A4 and B3
are installed near the staircase area which is an open space between floor 1 and floor
2. It is possible that the RSS values measured at this area are similar.
A7
Access Point
B4
A4
Average RSS (dBm)
40
24.9dB
17.3 dB
20
B3
20.2dB
13.9 dB
0
-20
-40
-60
-80
-53.1
-70.4
-51.7
-76.6
-61.4
-75.2
-52.4
-72.6
-100
Floor 1
Floor 2
Difference of Average RSS
Figure 5.15: Average and difference of RSS measured from access point A7, B4, A4
and B3 at different floors
Measurements in this section indicate that floor attenuates WLAN RSS in a
measurable way. Two locations vertically symmetry on two consecutive floors will
have different RSS value detected from a given AP. This suggests that WLAN RSS
can be used as location fingerprint to differentiate floor in a multi-floor indoor
environment.
94
5.11
Summary
The premise of WLAN RSS location fingerprinting is that RSS information
provides a means of inferring MT’s location. Therefore, before development of the
proposed MILES, a series of measurements are conducted to study the characteristics
of WLAN RSS in indoor environment. The results of observations and their effects
on design and deployment of the proposed system are summarized in Table 5.8.
Study in this chapter indicates that it is feasible to use WLAN RSS as
location fingerprint to estimate indoor location. Various characteristics of WLAN
RSS for indoor location estimation application have been identified. Some of the
environmental factors which will affect the characteristics of WLAN RSS in
indoor environment are investigated. Results obtained from this study are used in
design, development and deployment of the proposed system in the next chapter.
Table 5.8: Summary on WLAN RSS characteristics
Type of Study
Location dependency of
WLAN RSS
WLAN RSS Distribution
at a Stationary Indoor
Location
Characteristics
Conclusion
ƒ RSS changes with
ƒ RSS can be used as
distance traveled.
location fingerprint to
ƒ RSS tuple collected
estimate MT’s
from multiple APs are
location.
unique at distinct
ƒ Distance between two
physical location.
consecutive RLs where
ƒ RSSs measured from a
RSS location
given AP approximately
fingerprint will be
same over short
collected should not be
distance (< 1 meter).
too close to avoid
similarity in location
fingerprints.
ƒ RSS distribution at a
ƒ More than one RSS
stationary indoor
from a given AP should
location is not stable but
be collected to represent
fluctuates around a
the AP’s RSS
mean value.
distribution at that
ƒ RSS from AP with LOS
given location.
path fluctuate more
compared to AP with
NLOS path
95
Table 5.8: Summary on WLAN RSS characteristics (Continued)
Type of Study
Statistical
Representation of
WLAN RSS Distribution
at a Stationary Indoor
Location
Variation of Numbers of
APs Detected at a
Stationary Indoor
Location
Characteristics
ƒ RSS distribution at a
ƒ RSS histogram is used
stationary indoor
as location fingerprint
location is asymmetric
instead of mean RSS
and has multiple modes.
value.
ƒ RSS histograms created ƒ Small number of RSS
with little and many
samples is sufficient to
RSS samples are almost
create the RSS
similar.
histogram.
ƒ RSS from some APs
collected may be
unreliable due to too
little samples.
ƒ Presence of human
influenced the RSS in
different ways.
Effects of user proximity
Effects of user’s
orientation
ƒ RSS is affected by
user’s orientation.
Effects of types of
WLAN client device
ƒ RSS distribution is
affected by
environmental factors
such as people
movement and objects.
ƒ RSS measured by
different WLAN client
device may vary.
Effects of Multi-Floor
Environment
ƒ RSS collected at two
consecutive floors
varied at least 10dB.
Effects of different
environments
Conclusion
ƒ RSS histogram
collected from
unreliable APs should
be omitted from
location fingerprint.
ƒ RSS should be collected
based on types of
application. If the
system will be carried
by human user, then
off-line phase should be
conducted with human
proximity.
ƒ RSS collected at a given
location with different
orientations can be
regarded as different
location fingerprint.
ƒ RSS collection should
be collected based on
operating environment.
ƒ Same wireless client
device should be used
during off-line and online phases.
ƒ RSS can be used as
location fingerprint to
differentiate between
floors.
CHAPTER 6
SYSTEM DESIGN AND DEVELOPMENT
6.1
Proposed Multi-Floor Indoor Location Estimation Framework
Functional blocks of the proposed MILES are illustrated in Figure 6.1. In the
proposed system, multi-floor indoor environment is represented as a topological map.
Topological map represents the environment using a set of nodes and connection
edges. The nodes define the possible MT locations in the indoor environment and the
edges define the possible transition between those nodes.
During the off-line phase, a series of RSS calibration measurements will be
taken at every topological node. These calibration measurements will be stored in
radio map as the environmental model.
During on-line phase, the MT’s will collect the current RSS observations at
the unknown indoor location. These current RSS observations will be fed to a
probabilistic Bayesian filtering algorithm to estimate the MT’s current location.
Besides that, the topological map and motion constraints will be incorporated into the
algorithm to provide more accurate location estimation. The estimated MT’s location
is the topological node which has the highest probability distribution.
Since the topological map is a discrete abstraction of the indoor environment,
the location estimated is discrete space location. The accuracy of the system can be
further enhanced if the location can be estimated in continuous space. Therefore,
97
centre of mass and time averaging algorithms are introduced into the system to
provide continuous space location estimation.
Besides physical location, the proposed system is capable to provide
symbolic location estimation. Instead of location estimation in form of 3D
coordinates (x, y, floor), location is reported in form of descriptive logical area such
as “Reception area” and “Room 4A”.
The details on the design for each functional blocks of the proposed system
will be explained in the following sections.
RSS Calibration
Measurements
Current RSS
Observations
Location Estimation Module
Environment
Model
Bayesian
Filtering
Discrete Location
Estimation
Discrete Space
Location
Topological Map
Representation
Continuous Space
Location
Centre of Mass
Logical Area
Time Averaging
Figure 6.1: Proposed multi-floor indoor location estimation framework
98
6.1.1
Topological Map Representation
The proposed system represents the indoor environment as a discrete
topological map. In topological map representation, the topology of the real
environment is modeled using a set of graphs. A graph G={V, E} comprises of b
number of topological nodes V={v0, v1, …, vb-1} and c number of edges E={e0, e1, …,
ec-1}. Each edge eij={vi, vj} describes a possible path or transition from topological
node vi to topological node vj. Thus all the permissible routes in the indoor
environment are described by various combinations of edges of the graph G.
Figure 6.2 illustrates an example of topological map representation of the
indoor environment of a two-floor building. In Figure 6.2, permissible MT
movement paths are defined along the corridor, movement between rooms and
corridors and vertical movement between floors.
There are many advantages of using topological map representation. The
topological nodes can be used as the calibration nodes during off-line phase. It
provides guidance to the user about the exact location to collect the location
fingerprint. The topological nodes are also interpreted as the possible MT’s location
candidates during the on-line phase. Instead of using all locations in the continuous
space, a finite set of discrete locations within the indoor environment is used as the
possible MT’s location. This greatly reduces the computational complexity and time.
In addition, the topological map modeled the permissible movement of MT’s in the
indoor environment. Thus, motion constraint can be introduced to increase location
estimation accuracy. With motion constraint, MT’s reported locations will move
naturally, where the MT will not move through walls or other obstacles.
Besides that, the structure of the topological map serves as a good basis for
establishing the probabilistic location estimation model. The topological nodes are
used as the states of the probabilistic model and the edges between topological nodes
are used to govern the state transition probability.
99
Entrance
Floor 1
Floor 2
Figure 6.2: Topological map representation of a two-floor indoor environment
100
6.1.2
Environmental Model
The environmental model is the RSS location fingerprints database or radio
map constructed during off-line phase. The structure of the proposed environmental
model will be discussed here.
Measurements from section 5.4 suggest that RSS histogram can be used as
location fingerprint. Therefore, in the proposed system, normalized histogram is used
to describe the probability distribution of RSS at every calibration nodes.
Besides that, measurements in section 5.7 also discovered that RSS values are
affected by measurement direction. Therefore, in the proposed system, the RSS
location fingerprints are collected from two opposite directions at each calibration
node. Directions of collecting RSS location fingerprints for the topological map in
Figure 6.2 are depicted in Figure 6.3. In Figure 6.3, the two arrows beside each
topological node represent the two directions of RSS measurement. Therefore, a
single topological node will have two calibration nodes
101
Entrance
16
17
18
20
19
23
21
15
0
22
14
1
7
5
6
8
4
10
9
3
11
2
12
43
32
26
25
27 31
30
28
24
42 Staircase
29
33
Floor 1
36
37
38
35
39
40
34
41
Staircase
Floor 2
Figure 6.3: Calibration nodes
13
102
Figure 6.4 shows an example of normalized RSS histogram collected from
one WLAN AP at one calibration node. The histogram describes the RSS probability
distribution of that particular AP at the given location and direction. For instance,
from Figure 6.4, the probability of observed RSS equals to -57 dBm is about 0.20
and the probability of observed RSS equals to -55 dBm is 0.1.
0.25
Probability
0.2
0.15
0.1
0.05
0
-65 -64 -63 -62 -61 -60 -59 -58 -57 -56 -55 -54 -53 -52 -51 -50
RSS (dBm)
Figure 6.4: An example of normalized RSS histogram collected from one access
point at one calibration node
During the off-line phase, the RSS histograms will be collected for every
calibration node. The structure of the RSS histogram is shown in Figure 6.5. At every
calibration node, a total number of Nb APs are detected. Thus, Nb RSS histograms
are created for that particular calibration node. Each RSS histogram is constructed
from a total number of Nm RSS measurements. The RSS value, RSSj ranges from 0
dBm to -255 dBm. Histogram for a given AP may have Np different RSS values. The
normalized probability for each RSS value in the histogram, p(RSSj) is calculated
according to equation (6.1):
103
p(RSS j ) =
Number of RSS measurement with value of RSS j
Total number of RSS measurements, N m
(6.1)
where p(RSSj) is the normalized probability for RSS measurement with value of RSSj.
Nm is the total number of RSS measurements in the given histogram.
Topological node ID
Calibration node ID
Coordinate of calibration node ~ (x, y, floor)
Direction ~ (forward or backward)
MAC address of AP0
RSS Histogram of AP0
MAC address of AP1
RSS Histogram of AP1
……
MAC address of AP(Nb-1)
RSS Histogram of AP(Nb-1)
RSS Histogram of APi
RSS0
Probability ~ P(RSS0)
RSS1
Probability ~ P(RSS1)
……
RSS(Np-1)
Probability ~ P(RSS(Np-1))
Figure 6.5: Structure of the proposed environmental model
104
6.1.3
Bayesian Filtering
Given the topological map and environmental model, the task of the
probabilistic Bayesian filter is to estimate the MT’s location within the topological
map based on currently observed RSS values. The Bayesian Filtering framework
applied in this study is proposed by research in [7]. In their work, the framework
only tested for single-floor location estimation only.
Let the set A = {a0 , a1 ,..., aNa −1} denotes the N a number of AP installed in the
indoor environment and the set L = {l0 , l1 ,..., lNc −1} denotes the N c number of
calibration nodes defined in the topological map. Each calibration node li consists of
it’s coordinates on the floor map ( xi , yi ) , a floor number fi , an orientation θi of the
MT and a collection of RSS histograms from N b detected APs at that node, hik ,
k = 0,1,..., N b − 1 . The orientation θi = 1 when the MT is facing “forward” direction
and θi = -1 when facing “backward” direction. Then set L can be summarized as
below:
L = { l0 = ( x0 , y0 , f 0 , θ 0 , h0 k ), ... ,
lNc −1 = ( xNc −1 , y Nc −1 , f Nc −1 , θ Nc −1 , h( Nc −1)k ) }
(6.2)
During off-line phase, the MT collects RSS measurements at all calibration
nodes L as the environmental model. A single RSS measurement received
from N b APs is represented with M = {m0 , m1 ,..., mNb −1} . The number of APs detected
at each calibration node is always less than or equal to the total number of AP
installed in the indoor environment, N b ≤ N a . A series of N m RSS
measurements M z , z = 0,1,..., N m − 1 are taken at each calibration node li . From these
RSS measurements, the histograms hik are computed for each AP.
Once the off-line phase is completed, the proposed system is ready for
location estimation process. The MT will measure the current RSS observation
105
O = {o0 , o1 ,..., oNb −1} at the unknown location. The probabilistic algorithm will use the
information from the current observation, topological map and environmental model
to estimate the MT’s current location.
Bayesian filter is proposed to probabilistically estimate the MT’s location.
Referring to section 3.5.2, the Bayesian filter only provides a general framework for
recursive state estimation. In actual implementation, the representation of belief
Bel ( st ) , perceptual model p (ot | st ) and system dynamics p( st | st −1 ) need to be
determined. The specifications of the Bayesian filter applied in the proposed system
are listed below:
(i)
Representation of the belief ( Bel ( st ) = p ( st | o0 , o1 ,..., ot ) )
The proposed system represents the belief of the Bayesian filter in discrete space
using topological map. Possible MT’s location is represented with a set of discrete
topological nodes as explained in section 6.1.1. During on-line phase, Bayesian filter
is used to update the belief over these topological nodes. The topological node with
the highest probability or belief is the estimated current MT’s location.
Perceptual model ( p(ot | st ) )
(ii)
Perceptual model describes the likelihood of observing RSS value ot at location or
state st . This conditional probability can be obtained from the RSS histograms in the
environmental model constructed during the off-line phase.
(iii)
Motion model ( p( st | st −1 ) )
Motion model describes how the MT’s state or location changes over time. This
conditional probability predicts the state of MT at time t, st given that the MT was
previously at state st −1 . Motion model of the proposed system is obtained from the
structure of the topological map. Topological map describes the adjacency of the
locations and permissible movement of MT.
Details of the proposed Bayesian filter are discussed in the following sections.
106
6.1.3.1 State Space Representation
Let the state space Si = {S0 , S1 ,..., S N s −1} denotes the N s finite number of states
in the Bayesian filter. Each state S j represents a sequence of n ordered calibration
nodes li , S j = {l 0j , l1j ,..., l nj −1} , j = 0,1,..., N s − 1 . Instead of representing each state as a
single calibration node, state S j represents the last n calibration nodes traversed by the
MT, if it is currently at calibration nodes l nj −1 . By incorporating more than one
calibration node, the Bayesian filter is capable to track the dynamic of MT’s motion.
In order to represent each state with a sequence of ordered calibration nodes,
knowledge about the environment and permissible movement of MT in indoor
environment is needed. This information is obtained from the topological map.
Referring to Figure 6.3, the topological map represents movement of MT in the twofloor indoor environment with six different paths. Calibrations with two opposite
directions at each topological node split a single path into two motion tracks. The
summary of the paths and tracks from Figure 6.3 is given in Table 6.1. With this
defined tracks, motion constraints can be applied to construct the state space.
Motion constraints restricts the movement of MT in such a way that the MT
must traverse a sequence of neighbors in order to move from one location to the
other. Each calibration node in the topological map has a set of neighbors. The task
of motion constraints is to determine which neighbors are allowed to travel to from
current calibration node. For most calibration nodes, three allowed neighbors are
available:
(i)
The calibration node itself. This allows MT’s stationary motion in time.
(ii)
The next calibration node on the same track. This allows MT’s motion in the
same direction.
(iii)
The calibration node at the same location but on the opposite track. This
allows MT’s to change motion’s direction.
107
Table 6.1: Paths and tracks represented by the topological map in Figure 6.3
Path
1
2
3
4
5
6
Location
Floor 1
Floor 1
Floor 1
Floor 1
Floor 1
Connecting
floor 1 and
floor 2
Track
Calibration Nodes
1
0–1–2–3–4–5–6–7
2
8 – 9 – 10 – 11 – 12 – 13 – 14 - 15
3
16 – 17
4
18 – 19
5
20 – 21
6
22 – 23
7
24 – 25
8
26 – 27
9
28 – 29
10
30 – 31
11
32 – 33 – 34 – 35 – 36 – 37
12
38 – 39 – 40 – 41 – 42 – 43
For example in Figure 6.3, calibration node 1 has three allowed neighbors which are
calibration node 1, 2 and 14. Note that although calibration node 0, 15 and 13 are
neighbors of calibration node 1, but direct transitions from calibration node 1 to these
nodes are not allowed. These neighbors are restricted because MT cannot reach them
directly without traversing through other calibration nodes.
Some special calibration nodes may have less or more than three allowed
neighbors, for instance calibration node at the end of track, calibration node at Tjunction and calibration node at the cross-junction. These special cases can be
observed in Figure 6.3, where:
(i)
Calibration node 15 is the ending point of a track. It only has two allowed
neighbors which are calibration node 15 and 0. It doesn’t have next neighbor
on the same track.
108
(ii)
Calibration node 3 is located at a T-junction. It has four allowed neighbors
which are calibration node 3, 4, 12 and 32. Calibration node 32 enables the
MT to change path.
(iii)
Calibration node 6 is located at a cross-junction. It has five allowed neighbors
which are calibration node 6, 7, 9, 16 and 24. Calibration node 16 and 24
enable the MT to change paths.
Motion constraints are encoded into the state representation where each
consecutive element in state S j must be an allowed neighbor. Besides that, in order to
enable smooth motion along tracks, more than one track transition between
consecutive elements in the state sequence is restricted. This is based on assumption
that user carrying a MT doesn’t move fast and doesn’t change tracks more than once
in a short period of time. Based on these motion constraints, the structure of each
state S j can be illustrated as in Figure 6.6.
Calibration node l 0j
Calibration node l
Allowed neighbor & less than one track transition
1
j
Calibration node l 2j
Allowed neighbor & less than one track transition
Calibration node l nj − 2
Calibration node l
n −1
j
Allowed neighbor & less than one track transition
Figure 6.6: Structure of the proposed state representation
A few examples of allowed and restricted state representation for topological
map in Figure 6.3 are listed in Table 6.2. In Table 6.2, every state S j is represented
by n = 4 ordered calibration nodes.
109
Table 6.2: Allowed and restricted state representations
Sequence of
calibration nodes ( n = 4 )
Status
0–0–1–2
Allowed
2 – 3 – 32 – 33
State
0 – 1 – 14 – 15
0–0–0–0
Reason
Allowed neighbors and no track
transition
Allowed neighbors and one path
transition
Allowed neighbors and one track
transition
Allowed neighbors and static
motion
0–2–3–4
2 is not neighbor of 0
0–1–2–1
1 is not allowed neighbor of 2
Restricted
0 – 1 – 13 – 14
13 is not allowed neighbor of 1
State
2 – 3 – 12 – 43
43 is not allowed neighbor of 12
0 – 1 – 14 – 1
Allowed neighbors but two track
transitions in sequence
Allowed neighbors but two track
transitions in sequence
1 – 14 – 1 – 14
6.1.3.2 Topological Markov Localization
The Bayesian filtering technique implemented in the proposed system is
known as topological Markov localization. The name comes from the topological
map which represents the environment and the Markov assumptions applied to the
environment. The environment is assumed to fulfill first-order Markov assumption,
where current MT’s location at time t only depends on the previous location at time
t-1. This assumption is reasonable in the context of MT’s location estimation, since
the locations prior to t-1 will not provide additional information about the current
MT’s location [45].
110
Before the details of the algorithm are presented, the basic concept of
topological Markov localization is explained with a simple example. Consider the
environment depicted in Figure 6.7. Here the MT is assumed to move in onedimensional space. The MT only moves horizontally without turning along a straight
path. The path is represented with a topological map consisted of nodes and edges.
Each topological node has a RSS fingerprint, Oi collected from one AP during offline phase.
The MT’s initial location is unknown. Markov localization represents the
current uncertainty in the belief with a uniform probability distribution over all states,
Bel − ( S t =0 ) , as shown in Figure 6.7(a). The belief probability distribution is discrete
since the path is represented with discrete topological mapping.
In order to estimate its current location, the MT queries its WLAN client
device and detected that current RSS observation is OB. Markov localization updates
and corrects the belief by raising the probability for nodes with RSS fingerprint OB
and lowering it elsewhere, as shown in Figure 6.7(b). The resulting belief is multimodal, reflecting the fact that RSS fingerprinting alone is not sufficient for global
localization. The probabilities for other nodes with RSS fingerprint other than OB are
non-zero. This is to account for the uncertainty in RSS observations.
Now, the MT moves forward. Markov localization predicts and shifts the
belief distribution accordingly, as shown in Figure 6.7(c). Notice that the distribution
becomes smoother and less certain due to the noise in MT’s motion.
Finally, the MT queries its WLAN client device again and detected that
current RSS observation is Oc. Markov localization updates and corrects the belief
by multiplying the predicted belief with current observation distribution, which leads
to the belief shown in Figure 6.7(d). Now, the belief is concentrated at one node and
the location of the MT is more certain.
111
RSS
Fingerprint
Topological
Map
OA
OB
n1
OC
n2
OB
OD
n3
n4
OE
n5
n6
(a)
Bel − ( S t =0 )
OA
OB
n1
OC
n2
OB
n3
OD
n4
OE
n5
n6
(b)
p(O | S j )
Bel ( S t =0 )
OA
OB
n1
OC
n2
OB
n3
OD
n4
OE
n5
n6
(c)
Bel − ( S t =1 )
OA
n1
OB
OC
n2
OB
n3
OD
n4
OE
n5
p(O | S j )
Bel ( S t =1 )
Figure 6.7: Basic operation of topological Markov localization
n6
(d)
112
The key idea of topological Markov localization is to compute the a posteriori
probability for all states at time t, given all RSS observations available since
initialization, p ( S tj | O 0 , O1 ,..., O t ) . At each time step t, the a posteriori probabilities
are given by the following algorithm:
N s −1
p( S tj | O 0 , O1 ,..., O t ) = α t ⋅ p (O t | S tj ) ⋅ ∑ ⎡ p ( S tj | Stj−1 ) ⋅ p ( Stj−1 | O 0 , O1 ,..., O t −1 ) ⎤
⎣
⎦
j = 0
(6.3)
where:
αt
=
1
N s −1
∑ p( S
j =0
t
j
| O 0 , O1 ,..., O t )
= normalizing term to enforce the law of total probability
p (O t | S tj )
= the perceptual model which describes the likelihood of
making RSS observation O at location S j at time step t.
p( S tj | Stj−1 )
= the motion model which describes the system dynamics. It
describes the probability of MT currently at state S j given
that it was previously at state Sj .
At each time step t, the estimated MT’s current state Sc is given by the a posteriori
probability which has the maximum value.
Sc = argmax p( S tj | O 0 , O1 ,..., O t )
j
(6.4)
113
Finally, the estimated MT’s current location Lc is given by the ( n - 1 )th element in
current estimated state Sc . The estimated MT’s current location Lc is one of the
calibration nodes in the topological map.
Sc = {lc0 , lc1 ,..., lcn −1}
(6.5)
Lc = lcn −1
(6.6)
6.1.3.3 Perceptual Model
The perceptual model, p (O t | S tj ) is the likelihood of observing RSS
observation O at state S j during time step t. This a priori probability can be obtained
from the environmental model created during the off-line phase. From the
independence of percepts assumption, the RSS observation, O at time t only depends
on the ( n - 1 )th element of the state S j at time t. Therefore, the a priori probability
can be simplified as
p (O t | S tj ) = p (O t | li = l nj −1 )
(6.7)
P (O t | li ) denotes the likelihood of observing RSS observation O at calibration node
li . This is given by the RSS histograms collected at calibration node li during the offline phase.
Nb −1
P (O t | li ) = ∏ hik (Okt )
k =0
where hik (Okt ) denotes the RSS histogram collected at calibration node li .
(6.8)
114
6.1.3.4 Motion Model
The motion model p( S tj | Stj−1 ) is the state transition probability which
describes how the state of the system changes over time. Topological map
representation provides a convenient way to constraint the transition between states
since the connectivity between states is given by the edges between nodes. This
connectivity is time-invariant once the topological map is drawn. Therefore, the
motion model is represented with time-invariant state transition probabilities
p( S j | Sj ) to promote fluent motion with constant velocity or direction, yet
providing abrupt changes [7]. The notation t is dropped to indicate the time-invariant
characteristics.
In order to control the continuity between locations, each state S j maintains a
set of states which it can transition in the next step. Given S j = {l 0j , l1j ,..., l nj −1} and
Sj = {l0j , l1j ,..., lnj −1} , then the state transition probability p ( S j | Sj ) is non-zero only
for those candidate states S j which obeyed the following rule:
l qj −1 = lqj ;
q = 1, 2,..., n − 1
(6.9)
For example, let’s assume each state contains an ordered sequence of four calibration
nodes, n = 4 , the condition for non-zero state transition probability p ( S j | Sj ) is
illustrated in Figure 6.8. In brief, p ( S j | Sj ) is non-zero if state S j is a left-shift of Sj
with replacement of l nj −1 with another element in L [7].
115
Sj
l0j
l1j
l2j
l3j
1
j
2
j
3
j
Left shift
Sj
l
0
j
l
l
l
Replace with
another element
from L
Figure 6.8: Motion model
After generating all the possible time-invariant state transitions, the next step
is to calculate the values for the non-zero state transition probabilities p ( S j | Sj ) . The
equation to calculate the state transition probabilities p( S j | Sj ) is given by [7]:
1
− 2 ⎡l nj−1 − l ′⎤
1
⎣
⎦
p ( S j | Sj ) =
e 2σ
σ 2π
2
(6.10)
where:
σ
= controls the degree of falloff
l′
= predicted next location
The value of l ′ is predicted by averaging the first n − 1 elements of state S j using
the following equation:
n−2
l′ =
∑l
r =0
r
j
n −1
(6.11)
Equation (6.10) assumes Gaussian noise for the state transition probabilities
p( S j | Sj ) , where the mean of the distribution is the predicted location l ′ . The state
S j which has the value of l nj −1 closer to the predicted next location l ′ will have higher
state transition probabilities. This is illustrated in Figure 6.9.
116
From Figure 6.9, location lan −1 is closer to the predicted location l ′ compared
to lbn −1 . Thus, the state transition probability for state Sa , with probability value
p( S a | Sj ) is higher than state Sb , with probability value of p ( Sb | Sj ) . In other
words, state Sa has higher probability to be the next state of state Sj compared to
state Sa .
p ( S j | Sj )
p( S a | Sj )
p ( Sb | Sj )
l nj −1
lbn −1
l ′ lan−1
Figure 6.9: State transitional probability
In order to promote fluent motion along tracks in indoor environment and
reliable floor estimation, state transition between state S j and state Sj , p( Sa | Sj ) is
multiply by a transition penalty factor, ptrans ; 0 < ptrans < 1 . The state transition
probabilities with transition penalty, p ( S a | Sj ) is given by the following equation:
117
p ( S a | Sj ) = ptrans ⋅ p ( S j | Sj ) ;
0 < ptrans < 1
(6.12)
where:
ptrans
⎧ ptrans1 ;
⎪p
;
⎪
= ⎨ trans 2
⎪ ptrans 3 ;
⎪⎩ ptrans 4 ;
stationary ⎫
step forward ⎪⎪
⎬
change direction ⎪
change path ⎪⎭
(6.13)
Values of ptrans are set based on the usual behavior of the MT. Lower value of
ptrans indicating lower occurrence of corresponding state transition. For example, if
the MT is usually moving, then lower value of ptrans should be set for stationary. The
strategies for setting the value of ptrans will not be investigated in this study.
6.1.3.5 Proposed Location Estimation Algorithm
The previous sections have presented the underlying principles of the
proposed topological Markov localization model. The estimated location of the MT
is updated whenever a new RSS observation is made. The complete recursive
algorithm for updating the a posteriori probability distribution, p ( S tj | O 0 , O1 ,..., O t )
of the MT’s estimated location is given by equation (6.3).
The procedure to recursively update equation (6.3) is listed in Figure 6.10.
118
for each state S j do
// initialize the belief
Bel ( S t =0 = S j ) ←⎯
⎯ p( S t =0 = S j )
end for
forever do
if new RSS observation O t is received do
α t ←⎯
⎯0
// reset
for each state S j do
// apply the perceptual model
Bel ( S t = S j ) ←⎯
⎯ p (O t | S j ) ⋅ Bel ( S t −1 = S j )
α t ←⎯
⎯ α t + Bel ( S t = S j )
end for
for each state S j do
Bel ( S t = S j ) ←⎯
⎯
// normalize the belief
Bel ( S t = S j )
αt
end for
end if
for each state S j do
// apply the motion model
Bel ( S t = S j ) ←⎯
⎯ ∑ ⎡ p( S tj | Stj−1 ) ⋅ Bel ( S t −1 = Stj−1 ) ⎤
⎣
⎦
j
end for
end forever
Figure xx: Procedure for topological Markov localization algorithm
Figure 6.10: Procedure to recursively update the proposed topological Markov
localization algorithm
119
6.1.4
Continuous Space Location Estimation
The discrete Markov localization algorithm proposed in section 6.13
estimates the MT’s location in a discrete space domain. The estimated MT’s location
returned by the system is one of the topological nodes defined in the topological map.
The accuracy of the system will be low if the density of topological nodes in the
indoor environment is low. In order to increase the accuracy, two continuous space
location estimation algorithms proposed by [46] are implemented in the system.
6.1.4.1 Centre of Mass
The first technique applied to estimate the MT’s location in continuous space
is centre of mass algorithm. Before explaining the implementation of this technique,
the general framework of centre of mass algorithm is presented.
In physics, centre of mass is defined as the weighted average of locations. In
a discrete domain, the masses are interpreted as the weights. For a discrete
distribution with N cm number of masses with weights W = {w0 , w1 ,..., wNcm −1}
located at locations X = {x0 , x1 ,..., xNcm −1} ,the location of centre of mass R is given by
N cm −1
R=
∑ wx
i =0
N cm −1
i i
∑w
i =0
(6.14)
i
The location of centre of mass R can be interpreted as the point where the
weights of the masses is equally balanced on a lever as illustrated in Figure 6.11. In
Figure 6.11, the lever system consists of N cm = 2 discrete masses, each with weight
of w1 and w2 respectively ( w1 > w2 ). Both of theses masses are located at a distance
of x1 and x2 from a reference point. The location of centre of mass R is then given
by:
120
1
R =
∑w x
i =0
1
i i
∑w
i =0
=
w1 x1 + w2 x2
w1 + w2
(6.15)
i
The location of centre of mass R is closer to the mass with higher weight.
x2
R
x1
w1
w2
Reference
Location
Centre of Mass’s
Location
Figure 6.11: Concept of centre of mass
Centre of mass concept can be applied in the proposed system to provide
continuous space location determination. The calibration nodes are interpreted as
masses in the discrete physical space whose weights W are equal to the normalized
probabilities assigned by the discrete space location determination process. Centre of
mass’s location R is then calculated from N cm numbers of calibration nodes which
have the highest normalized probability.
Let L = {l 0 ,l1 ,...,l Nc −1} denotes the set of calibration nodes sequence arranged
in descending order according to the normalized probability. The normalized
li ) . The first N cm number of
probability of calibration nodes l i is represented by p (
L are used to calculate the location of centre of mass R .
calibration nodes from 121
Location of centre of mass R is the continuous space location lcm is given by
N cm −1
∑ [ p(l ) ⋅l ]
i
lcm =
i =0
N cm −1
∑
p (l i )
i
; ( N cm ≤ N c )
(6.16)
i =0
The continuous space location lcm is not necessarily an element in the set of discrete
calibration nodes L .
The concept of continuous space location estimation using centre of mass
technique is shown in Figure 6.12. In Figure 6.12, the system consists of two discrete
locations, A and B. Current MT’s true location is at the middle of these two locations.
Discrete space location estimation process calculated that the normalized probability
of MT currently located at location A and location B are 0.6 and 0.4 respectively.
With discrete space location estimation, location A will be estimated as the current
MT’s location since it has the highest probability. By using centre of mass technique,
the estimated MT’s location will be at the middle of location A and B, slightly biased
toward location A. Thus, the continuous space location estimation is capable to
provide higher location estimation accuracy compared to the discrete space approach.
Estimated MT’s
location using discrete space
location estimation
Estimated MT’s
ocation using centre of
mass technique
p(B) = 0.4
Location B
Actual MT’s
location
p(A) = 0.6
Location A
Centre of mass’s
Location
Figure 6.12: Continuous space location estimation using centre of mass technique
122
6.1.4.2 Time Averaging
The second technique applied to estimate the MT’s location in continuous
space is time averaging technique. In this approach, the continuous space location is
determined by averaging a time series of discrete space locations estimated by the
discrete space location estimation process.
Let Tu denotes the location update time during on-line phase of the proposed
system. Tw denotes the length of the time-averaging window to smooth the resulting
discrete space location estimates. The number of discrete location estimates to be
average N ta is given by:
N ta =
Tw
;
Tu
( Tu ≤ Tw )
(6.17)
The stream of discrete location estimates to be average is denoted
as Lta = {l 0 , l1 ,..., l Nta −1} . The continuous space location determination using time
averaging technique lta is then given by the following equation:
∑ li
Nta −1
lta =
i =0
N ta
(6.18)
123
6.1.5
Logical Area Estimation
Besides physical location estimation in terms of 3D coordinates (x, y, floor),
the proposed system is capable to provide symbolic location estimation. Location
space in indoor environment is defined in terms of descriptive names called logical
area such as “Reception”, “Meeting Room” and “Pantry”. Unlike the physical
location, logical area is more human-friendly and suitable for LBS which deliver
services based on indoor areas.
Figure 6.13 illustrates a simple example of representation of indoor location
space with logical areas. In Figure 6.13, a total of five logical areas are defined for
the two-floor indoor environment. The logical area is named based on the function of
the area. During the on-line phase, let’s say the current estimated MT’s physical
location is (x = 40, y = 50, floor = 2), then the estimated logical area reported to the
user will be “Reception”. This logical area location information is more humanfriendly compared to the physical location (x, y, floor).
“Meeting Room”
“Reception”
(x, y, floor) = (40, 50, 2)
“Floor 1’s Hallway”
Floor 1
“Floor 2’s Hallway”
Floor 2
Figure 6.13: Representation of indoor area using logical area
124
6.2
System Development
6.2.1
Software Architecture
Software architecture of the proposed system is shown in Figure 6.14. The
proposed system is divided into four main components, which are the System
Interface, the Data Collection, the Database and the Location Estimation.
The System Interface component is the GUI which assists the user to
configure the proposed system, to construct the radio map during off-line phase and
to show the estimated MT’s location information during on-line phase.
The Data Collection component is used to collect RSS information during the
off-line and on-line phases. It consists of two main modules which are the RSS
Reader and the Access Point Filter. The RSS Reader module is used to collect RSS
samples from the APs. For every detected AP, the information collected is its SSID,
MAC address and RSS value. The Access Point Filter module is used to filter out
RSS samples from unwanted APs.
The Database component stores the RSS information collected during the
off-line phase. It consists of three main modules which are the Topological Map
Builder, the Histogram Builder and the Histogram Reliability Checker. The
Topological Map Builder module is used to construct the topological map. The
Histogram Builder module is used to construct histograms for RSS samples collected
during off-line phase. The histograms constructed will be stored in the Radio Map.
The Histogram Reliability Checker module is used to discard unreliable RSS
histograms.
125
Location Based Services
Estimated Location
System Interface
Database
Location Estimation
Topological Map
Builder
Logical Area
Estimator
Radio Map
Continuous-Space
Location Estimator
Discrete-Space
Location Estimator
Histogram
Reliability Checker
Histogram Builder
Data Collection
Access Point Filter
RSS Reader
IEEE 802.11g WLAN
Infrastructure
Figure 6.14: Software architecture of the proposed system
126
The Location Estimation component is used to estimate the MT’s location
during on-line phase. It consists of three main modules which are the Discrete-Space
Location Estimator, the Continuous-Space Location Estimator and the Logical Area
Estimator. The Discrete-Space Location Estimator module is used to estimate
current MT’s location through Markov localization algorithm. Given current RSS
observation, this module returns the topological node in the Radio Map that has the
maximum probability as current estimated location. The Continuous-Space Location
Estimator module receives the discrete locations estimated by Discrete-Space
Location Estimator module and returns a more accurate location estimation in
continuous-space. The Logical Area Estimator module receives the discrete or
continuous estimated MT’s location and returns the logical area where the MT’s is
currently located.
Details of each individual component and the modules involved will be discussed in
the following sections.
6.2.2
System Interface Component
6.2.2.1 Main GUI
Main GUI for the proposed system is shown in Figure 6.15. This is the main
interface between the user and the proposed system during configuration, off-line
phase and on-line phase.
127
Figure 6.15: The Main GUI of the proposed system
6.2.2.2 System Configuration
Before off-line and on-line phases, user needs to configure the proposed
system. The user has to select the types of location estimation preferred. Through the
Location Options submenu, as shown in Figure 6.16(a), the user can select either
discrete or continuous space location estimation. If continuous space location
estimation is preferred, then the user has to decide on either centre of mass technique
or time averaging technique. Time averaging technique will require the user to select
the time averaging interval, as shown in Figure 6.16(b). The timing averaging
interval is in unit of millisecond.
128
(a)
(b)
Figure 6.16: The Location Options submenu. (a) Discrete or continuous space
location estimation. (b) The time averaging interval for time averaging technique
6.2.3
Data Collection Component
The two main functions of Data Collection component are to collect RSS
measurements during off-line phase and to collect RSS observations during on-line
phase. It consists of two main modules which are the RSS Reader module and
Access Point Filter module. Flow chart for Data Collection component is illustrated
in Figure 6.17.
129
Start
Call RSS Reader module
to read RSS tuple
<SSID, Access Point MAC, RSS>
Call
Access Point Filter module
to discard unwanted access point
Valid RSS
tuple?
No
Yes
Write RSS tuple to text file
< Access Point MAC, RSS>
Yes
Scan again?
No
End
Figure 6.17: Flow chart of Data Collection Component
130
6.2.3.1 RSS Reader
Function of RSS Reader is to capture real-time WLAN beacons over the air
and then extract the AP information from the beacons. This module is used during
off-line phase and on-line phase.
In order to capture the WLAN beacons, the WiFiSpotter API from Place Lab
project [47] is used in this module. This library is used because it is an open source
Java library and it is capable to retrieve information from AP in passive mode.
Besides that, it is functioning well with most of the off-the-shelf WLAN client
device. The WiFiSpotter API is used to obtain three parameters from the WLAN
beacons which are SSID, AP’s MAC address and RSS value. This information is
called RSS tuple, <SSID, Access Point MAC, RSS>.
The information reported by the WiFiSpotter Reader is stored in a text file for
further processing by other module. The flow chart and java source code for
retrieving information from WiFiSpotter API is shown in Figure 6.18 and Figure
6.19 respectively.
131
Start
Create a new WiFiSpotter to
access the WLAN client device
Open the WiFiSpotter
Scan for available WLAN
beacons
Beacons
available?
No
Yes
Read detected beacons
Extract information from beacons
<SSID, Access Point MAC, RSS>
Close the WiFiSpotter
End
Figure 6.18: Flow chart of RSS reader module
132
Spotter s = new WiFiSpotter();
// Create a new WiFiSpotter
// Access the WLAN interface
try {
s.open();
// Collect detected beacons into vector
BeaconMeasurement m = (BeaconMeasurement) s.getMeasurement();
if (m.numberOfReadings() > 0) {
// If beacons available
// Iterate through the beacons vector and extract information
for (int i = 0; i < m.numberOfReadings(); i++) {
WiFiReading r = (WiFiReading) m.getReading(i);
String SSID = r.getSsid();
// WLAN SSID
String MAC = r.getTd(); // Access Point MAC address
int RSS = r.getRssi();
// Access Point RSS
//write to text file
……
……
}
}
} catch (SpotterException ex) {
// No WLAN interface detected or access failure
ex.printStackTrace();
}
// Close the access to WLAN interface
try {
s.close();
} catch (SpotterException ex) {
// Closing WLAN Interface failure
ex.printStackTrace();
}
Figure 6.19: Source code for retrieving RSS tuple using WiFiSpotter API
133
6.2.3.2 Access Point Filter
During off-line phase, the RSS Reader module may detect WLAN beacons
from other WLANs in and around the indoor environment. Some of these WLANs
especially those from other buildings are not reliable to be used as RSS location
fingerprint because the AP’s location, hardware and software settings, operation time
and other parameters may be changed without prior notice. These changes will
change the value of RSS at a given calibration node. If the RSS readings from these
types of WLAN are used as RSS fingerprints, it may affect the performance of
location estimation during on-line phase.
Therefore, it is preferable to collect RSS reading only from trusted WLANs
that will operate during off-line and on-line phases. The function of this module is to
make sure only RSS readings from trusted WLANs are used as the RSS fingerprint.
The SSID of trusted WLANs are defined by user before off-line phase and
stored in a text file named goodSSIDFile.txt. An example of the mentioned text file is
shown in Figure 6.20. In this example, only RSS from APs in network with SSID
WCCWireless will be processed as RSS location fingerprint. The flow chart of
Access Point Filter module is shown in Figure 6.21.
Figure 6.20: Example of text file which defines the SSID of the trusted WLANs.
134
Start
Read trusted SSID from user
defined text file
Read RSS information from
RSS Reader module
<SSID, Access Point MAC, RSS>
Compare SSID with
the trusted SSID
Trusted?
No
Yes
Can be used as RSS fingerprint
Discard current RSS information
End
Figure 6.21: Flow chart of access point filter module
6.2.4 Database Component
The main function of Database component is to get the RSS tuples return by
the Data Collection component and transform them into RSS location fingerprints.
The flow chart for this process is shown in Figure 6.22. These RSS location
fingerprints will then be stored in radio map together with the topological map
information. This component consists of the Histogram Builder module, Histogram
Reliability Checker module and Topological Map Builder module.
135
Start
Create empty
RSS histogram table
Open RSS tuples text file created
during data collection
Read all RSS tuples into array
Close RSS tuples text file
Call
Histogram Builder module
to construct RSS histograms
Call
Histogram Reliability Checker
module to insert reliable
RSS histograms into table
End
Figure 6.22: Flow chart of Database component
6.2.4.1 Histogram Builder
The function of this module is to construct RSS histograms based on RSS
tuples collected by Data Collection module. Flow chart in Figure 6.23(a) and Figure
6.23(b) show the steps for constructing histograms for each calibration node.
136
Start
Read RSS tuples array
Get detected access points array
Read a RSS tuple
< Access Point MACi, RSSi>
Find histogram for this access
point, Access Point MACi
Histogram for this
access point exists?
No
Create new histogram for
this access point
No
Insert the new RSS value
to the histogram, RSSij
Yes
Find RSS value in histogram for
this access point, RSSij
RSS value exists?
Yes
Count (RSSij) = Count (RSSij) + 1
Total (RSSi) = Total (RSSi) + 1
No
Last RSS tuple?
Yes
A
Figure 6.23(a): Flow chart of Histogram Builder Module
137
A
Convert RSS histogram
count to probability
Get histogram for access point,
Access Point MACi
Get total RSS samples in
histogram, Total(RSSi)
Get RSS value in histogram, RSSij
Get count of RSS value
in histogram, Count(RSSij)
Probability of RSSij,
Probabilityij =
No
Count(RSSij)
Total(RSSi)
Last RSS value?
Yes
Last access point?
No
Yes
Store RSS Histograms for all access points
< Access Point MACi, RSSij, Probabilityij >
End
Figure 6.23(b): Flow chart of Histogram Builder Module (Continued)
138
6.2.4.2 Histogram Reliability Checker
Histogram Builder module creates the probability RSS histograms for all
trusted AP detected at a given calibration node. Some of these histograms may not be
reliable to be used as RSS fingerprints due to too few RSS samples compared to
average RSS samples per histogram or the RSS values are too low.
Histogram Reliability Checker module is used to discard these unreliable
RSS histograms. The number of samples in a given histogram must exceed certain
reliability level in order to be accepted as reliable RSS histogram. The steps involved
in this module are illustrated in Figure 6.24.
139
Start
Get all RSS histograms for
calibration node
Calculate average RSS samples per histogram
Average =
Total RSS samples in all histograms
Number of histograms
Calculate reliability threshold
Reliability = Average * Reliability Percentage
Threshold
Get histogram for access point,
Access Point MACi
Get total RSS samples in
histogram, Total(RSSi)
Total(RSSi) >=
Reliability Threshold?
No
Discard histogram
Yes
Store histogram as
RSS fingerprint
No
Last access point?
Yes
End
Figure 6.24: Flow chart of Histogram Reliability Checker module
140
6.2.4.3 Topological Map Builder
The function of Topological Map Builder is to construct the topological map
based on user input. A simple graph is used to represent the uncalibrated topological
map, while a directed graph is used to represent the calibrated topological map.
Simple graph and directed graph are illustrated in Figure 6.25. The differences
between uncalibrated map and calibrated map are the node representation.
Uncalibrated node or referred as topological node in the proposed system contains
the coordinate only, (x, y, floor) while calibrated node or referred as calibration node
contains the coordinate and also direction of measurement, (x, y, floor, direction).
Since RSS measurement taken at two directions at each topological node n1, each
topological node will have two corresponding calibration nodes c1 and c2.
A simple graph is an unweighted, undirected graph containing no loops or
multiple edges as shown in Figure 6.25(a). On the other hand, directed graph is graph
with directed edges and contains loops, as shown in Figure 6.25(b). The edges are bidirected to represents bidirectional movement between nodes and loops represent
stationary motion at a given node.
c1 , c2
n1
(a)
(b)
Figure 6.25: (a) Simple graph (b) Directed graph
In order to implement these graphs in the proposed system, a Java graph library
called JGraphT is used [48]. JGraphT library provides an easy implementation of
simple and directed graphs in Java programming. The SimpleGraph( ) and
141
DefaultDirectedGraph( ) classes in JGraphT library are used to construct the simple
graph and directed graph, respectively. The source code of these two
implementations is listed in Figure 6.26.
// Create a new simple graph for uncalibrated topological map
SimpleGraph
topological_map = new SimpleGraph();
// Create a new directed graph for calibrated topological map
DefaultDirectedGraph calibration_map = new DefaultDirectedGraph();
5.2.5
… … Location Estimation Component
……
/***** Uncalibrated topological map***********/
// Create new topological nodes n1 and n2
TopologicalNode
n1 = new TopologicalNode(x1, y1);
TopologicalNode
n2 = new TopologicalNode(x2, y2);
// Add topological nodes n1 and n2 to uncalibrated topological map
topological_map.addVertex(n1);
topological_map.addVertex(n2);
// Add undirected edge between topological nodes n1 and n2
topological_map.addEdge(n1,n2);
/***** Calibrated topological map***********/
// Create new calibration nodes c1 and c2 topological node n1
CalibrationNode
c1 = new CalibrationNode(n1, direction1);
CalibrationNode
c2 = new CalibrationNode(n1, direction2);
// Add calibration nodes c1 and c2 to calibrated topological map
calibration_map.addVertex(c1);
calibration_map.addVertex(c2);
// Add directed edge from calibration node c1 to c2
calibration_map.addEdge(c1,c2);
// Add directed edge from calibration node c2 to c1
calibration_map.addEdge(c2,c1);
Figure 6.26: Source Code for constructing topological map with simple graph and
directed graph using JGraphT API
142
6.2.5
Location Estimation Component
The function of this component is to estimate MT’s location based on current
RSS observation. It provides both physical and symbolic location reports. This
component consists of three main modules which are Discrete-Space Location
Estimator module, Continuous-Space Location Estimator module and Logical Area
Estimator module.
6.2.5.1 Discrete-Space Location Estimator
The Discrete-Space Location Estimator estimates the MT’s location in
discrete space based on the probabilistic algorithm presented in section 6.1.3. The
basic steps involved in this module are:
(i)
Step 1: Initialization of state probabilities
(ii)
Step 2: Get current RSS observation
(iii) Step 3: Calculate RSS observation probabilities
(iv) Step 4: Predict state transition probabilities (Prediction of current state)
(v)
Step 5: Update state probabilities (correction of state predicted using current
RSS observation)
(vi) Step 6: Normalization
(vii) Step 7: Get estimated location
The details of these steps are illustrated in Figure 6.27(a) to Figure 6.27(i)
143
Start
Generate all allowed states
Initialize location estimation process
Si = {S0 , S1 ,..., S N s −1}
Normalize the estimated state probabilities
p ( S tj | O 0 , O1 ,..., O t ) = α t ⋅ p ( S tj | O 0 , O1 ,..., O t )′
Get current RSS observation
Generate all possible
state transitions, Sj → S j
n −1
j
};
n −1
j
};
S j = {l , l ,..., l
0
j
1
j
Sj = {l , l ,..., l
0
j
l qj −1 = lqj ;
1
j
Calculate RSS observation probabilities
Get the estimated MT’s state, Sc
p (O t | S tj )
Sc = argmax p ( S tj | O 0 , O1 ,..., O t )
j
Predict current state
q = 1, 2,..., n − 1
p( S tj )′ =
Calculate time-invariant
state transition probabilities
p( S j | Sj )
Initialize all state
probabilities, p ( S j )
with global probability
N s −1
⎡ p( S tj | St −1 ) ⋅ p( St −1 | O 0 , O1 ,..., O t −1 ) ⎤
∑
j
j
⎣
⎦
j = 0
Get the estimated MT’s location, Lc
(Discrete calibration node)
Sc = {lc0 , lc1 ,..., lcn −1}
Lc = lcn −1
Correct the state prediction with
RSS observation probabilities
p ( S tj | O 0 , O1 ,..., O t )′ = p (O t | S tj ) ⋅ p ( S tj )′
End
Figure 6.27(a): Flow chart of the Discrete-Space
143
Location Estimator Module
144
Start
Define state length, n
Get all calibration nodes
L = {l0 , l1 ,..., lNc −1}
Create empty array to hold
ordered sequence of calibration
nodes with size n
sequnce[ ]
Get the first neighbor and
insert as the next element
in sequnce[counter ]
sequnce[counter ] = neighbors[0]
counter = n − 1 ?
No
B
Yes
C
Initialize sequence array
sequnce[ ] = null
Initialize counter
counter = 0
Get calibration node li and
insert as first element in
sequence array
Store sequnce[ ] as
possible state S j
Last calibration
node, lNc −1 ?
No
Yes
End
sequnce[counter ] = li
B
Get all neighbors of
sequnce[counter ] by traversing
through the topological map
neighbors[ ]
counter = counter + 1
Figure 6.27(b): Sub-Function: Generate all Possible States (Continued)
C
145
Start
Get state length, n
Get possible state generated, S j
Get the sequence of ordered
calibration nodes from S j
sequnce[ ]
Initialize sequence counter
counter = 1
Initialize change counter
change = 0
Get direction of calibration
node sequnce[counter ] and
sequnce[counter − 1]
Yes
Same direction?
counter = counter + 1
No
Get track ID of calibration
node sequnce[counter ] and
sequnce[counter − 1]
counter = n ?
No
Yes
change ≥ 2 ?
No
No
Yes
Store state S j as
State S j is
allowed state
not allowed
Same track?
Yes
change = change + 1
End
Figure 6.27(c): Sub-Function: Generate all Allowed States (Continued)
146
Start
Get state length, n
counter = counter + 1
Get all allowed states
counter = n ?
Si = {S0 , S1 ,..., S N s −1}
No
D
Yes
Store as valid state transition
from state Sj to state S j
F
Initialize counter
counter = 0
Reset counter = 0
Get one allowed state and set
as current state Sj
Last candidate?
Get the sequence of ordered
calibration nodes from Sj
No
E
Yes
sequnce[ ]
Last state?
E
No
F
Yes
Get candidate state S j where
state Sj will possibly transit to
G
Get the sequence of ordered
calibration nodes from S j
candidate[ ]
D
Compare calibration nodes
sequnce[counter + 1] and
candidate[counter ]
Yes
Same node?
No
Invalid state transition
G
End
Figure 6.27(d): Sub-Function: Generate all Possible State Transitions (Continued)
147
Start
Get all allowed states
Si = {S0 , S1 ,..., S N s −1}
Get number of allowed state, N s
Get one allowed state S j
Set the initial state probability
for state S j
p( S j ) =
1
Ns
Last state?
No
Yes
End
Figure 6.27(e): Sub-Function: Initialize all state probabilities with global probability
(Continued)
148
Start
Get state length, n
Get all allowed states
Si = {S0 , S1 ,..., S N s −1}
Calculate time invariant state
transition probability
1
p ( S j | Sj ) =
− 2 ⎡l nj−1 − l ′⎤
1
⎣
⎦
e 2σ
σ 2π
2
I
Get one allowed state and set
as current state Sj
Get the sequence of ordered
calibration nodes from Sj
n −1
j
Sj = {l , l ,..., l
0
j
1
j
}
Store time invariant state
transition probability
Last following
state?
No
H
Yes
H
Get one following state S j where
Last state?
state Sj are allowed to transit to
Yes
No
I
End
Get the sequence of ordered
calibration nodes from S j
S j = {l 0j , l1j ,..., l nj −1}
Calculate predicted next location
n−2
l′ =
∑l
r =0
r
j
n −1
Determine penalty value, Ptrans
for transition from state
Sj to state S j
Figure 6.27(f): Sub-Function: Calculate time-invariant state transition probabilities
(Continued)
149
Start
Get current
RSS observation, O t
Calculate average RSS for
every access point collected in
current RSS observation, O t
< APk , AvgRSS k >
Get all calibration nodes
L = {l0 , l1 ,..., lNc −1}
Get calibration node li
Get RSS histogram, hik of access
point APk at calibration node li
Get probability of RSS value
AvgRSSk in histogram hik
hik ( AvgRSSk )
Calculate observation probability
at calibration node li
p(O t | li ) = p(O t | li ) ⋅ hik ( AvgRSS k )
No
Last
access point?
Yes
Store current observation
probability for calibration node li
p(O t | li )
Last calibration
node?
Yes
End
Figure 6.27(g): Sub-Function: Calculate RSS observation Probabilities (Continued)
150
Start
Get all allowed states
Si = {S0 , S1 ,..., S N s −1}
Store the predicted state
probability for current state S j
p( S j )
J
Get one allowed state and set
as current state S j
Get all possible preceding states
of current state S j
Last state?
Yes
End
Get one preceding state Sj
Get state probability of state Sj ,
p( Sj )
Get the state transition
probability, p ( S j | Sj )
for transition from
state Sj to state S j
Calculate partial state probability
p( S j )′ = p ( S j | Sj ) ⋅ p(Sj )
Calculate total state probability
p( S j ) = p( S j ) + p( S j )′
No
Last preceding
state?
Yes
Figure 6.27(h): Sub-Function: Predict Current State (Continued)
No
J
151
Start
Get all allowed states
Si = {S0 , S1 ,..., S N s −1}
Get one allowed state and set as
current state S j
Get observation probability for state S j
p (O t | S j )
Get state probability of state S j , p( S j )
Correct the state probability
p ( S j )′′ = p (O t | S tj ) ⋅ p ( S j )
Store the corrected state probability for
current state S j , p( S j )′′
Calculate total state probability for all
state
total = total + p( S j )′′
No
Last state?
Yes
Normalize every corrected
state probability
p( S j )′′
p( S j ) =
total
Store the normalized state probability for
current state S j , p( S j )
End
Figure 6.27(i): Sub-Function: Correct and Normalize the Predicted State
Probabilities (Continued)
152
6.3
System Implementation
6.3.1
Off-Line Phase
After the proposed system is initialized and configured, the following steps are
taken to create the radio map:
(i)
Step 1: Create a new model.
(ii)
Step 2: Load floor maps for multi-floor environment.
(Figure 6.28, Figure 6.29)
(iii)
Step 3: Define topological map. (Figure 6.30)
(iv)
Step 4: Define logical area.(Figure 6.31)
(v)
Step 5: Start calibration process. (Figure 6.32)
(vi)
Step 6: Go to calibration node and face the direction of arrow shown.
(vii)
Step 7: Press start collect RSS button and wait until progress bar completed.
(viii) Step 8: Repeat Step 6 and Step 7 for all calibration nodes.
(ix)
Step 9: Save model. This is the radio map created during off-line phase.
(Figure 6.33)
(x)
Step 10: The histograms in radio map can be observed through analyzer.
(Figure 6.34)
6.3.2
On-Line Phase
Once the off-line phase is completed, the MT’s location can be estimated in
on-line phase according to the following steps:
(i)
Step 1: Load radio map model created in off-line phase. (Figure 6.35)
(ii)
Step 2: Start location estimation process. (Figure 6.36)
(iii)
Step 3: Textual and graphical location report - <x, y, floor, logical area>
(Figure 6.36)
153
Figure 6.28: Choose and insert floor map
Figure 6.29: Floor maps for multi-floor environment loaded
154
Figure 6.30: Define topological map
Figure 6.31: Define logical area
155
Figure 6.32: Calibration Process. Go to calibration node and face the direction of
arrow shown. Press start collect RSS button and wait until progress bar completed.
Figure 6.33: Save the model as radio map
156
Figure 6.34: The histograms in radio map can be observed through analyzer. This
example shows a part of the histograms collected at calibration node at location
(420.0, 476.0).
Figure 6.35: Load radio map model created in off-line phase
157
Estimated
Current Location
Figure 6.36: Textual and graphical location report, <x, y, floor, logical area>
6.4
Summary
This chapter describes the design, development and implementation processes
of the proposed system. Firstly, the working algorithms behind the proposed system
framework are explained. Next, the software architecture, modules and flow charts of
the proposed system are presented. Finally, the implementation of the proposed
system is shown. At the end of these processes, a WLAN-based MILES using RSS
location fingerprinting technique is successfully designed and developed.
The performance of the developed prototype is evaluated in real-time
experiments in the test-bed. Besides that, various factors affecting the performance of
the proposed system are investigated. Results from the experiments and
investigations are presented in the following chapter.
CHAPTER 7
RESULTS AND DISCUSSION
7.1
Introduction
Prototype of the proposed MILES is constructed and evaluated in real-time
environment. A series of experiments are conducted to evaluate the performance of
the proposed system in terms of stationary and mobile multi-floor indoor location
estimation. Besides that, experiments are conducted to investigate various factors
which affect the performance of the proposed system.
Results of the experiments conducted are presented and discussed in this
chapter. Firstly, the performance metrics used for evaluation is explained. Then, the
types of experiment conducted are introduced. Next, the results obtained from the
experiments are presented. This is followed by analyses and discussions on the
results.
7.2
Performance Metrics
The performance of the proposed MILES is evaluated in terms of location
estimation accuracy, precision, percentage of correct logical area estimation and
percentage of correct floor estimation.
159
7.2.1
Accuracy and Precision
Accuracy is expressed using the average Euclidean distance error between the
estimated location ( xi , yi ) and the actual location ( xii , i
yi ).
Euclidean Distance Error =
Accuracy =
i = 1, 2,3,..., n
(7.1)
Total Euclidean Distance Error
Total Number of Location Estimate
n
∑
=
( xi − xii ) 2 + ( yi − i
yi ) 2 ;
i =1
( xi − xii ) 2 + ( yi − i
yi ) 2
n
(7.2)
Precision is obtained from the cumulative distribution function (cdf) of the Euclidean
distance error.
7.2.2
Percentage of Correct Logical Area Estimation
Percentage of correct logical area estimation is used to evaluate the logical
location estimation capability of the proposed system. Higher percentage of correct
logical area estimation indicates the higher capability of the proposed system in
estimating the actual logical area.
Logical Area Estimation Error =
Number of Correct Logical Area Estimation
× 100%
Total Number of Logical Area Estimation
(7.3)
160
7.2.3 Percentage of Correct Floor Estimation
Percentage of correct floor estimation is used to evaluate the floor estimation
capability of the proposed system. Higher percentage of correct floor estimation
indicates the higher capability of the proposed system in estimating the actual floor.
Floor Estimation Error =
Number of Correct Floor Estimation
× 100%
Total Number of Floor Estimation
(7.4)
7.3
Types of Experiment
The objectives of the experiments conducted on the proposed system are:
(i)
To study the performance of the proposed system in terms of stationary
and mobile multi-floor indoor location estimation.
(ii)
To compare the performance of the three different location modes which
are discrete mode, centre of mass mode and time averaging mode.
(iii) To investigate various factors which affect the performance of the
proposed system.
In order to achieve these objectives, experiments listed in Table 7.1 are carried out.
Details and results obtained from these experiments are presented in the following
sections.
161
Table 7.1: Objectives and types of experiment carried out to study the performance
of the proposed system and various factors affecting the performance
Objective (i)
To study the performance of the proposed system in terms of stationary and mobile
multi-floor indoor location estimation.
Objective (ii)
To compare the performance of the three different location modes which are discrete
mode, centre of mass mode and time averaging mode.
Experiment 1a
Stationary Multi-floor Indoor Location Estimation
Experiment 1b
Mobile Multi-floor Indoor Location Estimation
Objective (iii)
To investigate various factors which affect the performance of the proposed system.
Experiment 2a
Experiment 2b
Effects of number of access points
Effects of number of topological nodes
Experiment 2c
Effects of number of RSS samples per fingerprint
Experiment 2d
Effects of off-line phase sampling interval
7.4
Stationary and Mobile Multi-Floor Indoor Location Estimation
In Experiment 1a, the MT was stationary during real-time multi-floor indoor
location estimation. In contrast, the MT was moving at walking speed during
Experiment 1b in the same test-bed. The test bed setups and parameter settings for
both of these experiments are presented in the following sections.
162
7.4.1
Experimental Setup
7.4.1.1 Test-bed
Experiment 1a and Experiment 1b are carried out in the two-floor building of
WCC, UTM (Figure 4.5 and Figure 4.6). The test bed are covered with sixteen IEEE
802.11g DLink DWL-2000AP+ WLAN APs. Details on the test-bed, AP’s
specification and installation locations are available in section 4.4, section 4.5 and
section 4.6. HP Compaq tc1100 Tablet PC is used to collect RSS fingerprint during
off-line phase and acted as the MT to be tracked during the on-line phase. The tablet
PC’s build-in Intel PRO/Wireless 2200BG MiniPCI WLAN is used as the WLAN
client device in these two phases.
The test-bed is consisted of hallways, open area and a small pantry. It was
divided into nine logical areas as illustrated in Figure 7.1 and Figure 7.2.
163
Legend
Topological Node
Hallway_3
Hallway_4
Hallway_5
Reception
Waiting_Area
Pantry
Figure 7.1: Distribution of topological nodes and logical areas on floor 1 of WCC,
UTM
164
Legend
Topological Node
Floor2_Entrance
Hallway_1
Hallway_2
Figure 7.2: Distribution of topological nodes and logical areas on floor 2 of WCC,
UTM
165
7.4.1.2 Off-line Phase
In off-line phase, RSS fingerprints are collected at 29 distinct topological
nodes. Floor 1 contains 21 topological nodes, while floor 2 contains 8 topological
nodes. The distribution of topological nodes is illustrated in Figure 7.1 and Figure
7.2. The physical distance between topological nodes is averagely 3.5 meters. The
maximum and minimum distances are approximately 5 meters and 2 meters
respectively. The topological nodes are logically link to together with 5 paths. Floor
1 contains 3 paths along the hallway. Floor 2 contains 1 path along the hallway.
Floor 1 and floor 2 is connected by 1 path via the staircase. At each topological node,
50 RSS measurements are recorded for each direction (forward and backward) with
sampling interval of 0.1 second.
7.4.1.3 On-line Phase for Stationary Multi-floor Indoor Location Estimation
In Experiment 1a, the MT is stationary at a given testing location during realtime location estimation. A total of 15 topological nodes are selected as the test
locations. Floor 1 contains 10 testing locations and another 5 are located at floor 2.
The physical location of the test locations are illustrated in Figure 7.3 and 7.4. At
each testing location, the MT stood and faced either forward or backward direction
randomly. The MT initialized the proposed MILES application and started the
location estimation.
At each testing location, a total of 50 location estimations are carried out
using 3 location estimation modes separately. Discrete location and centre of mass
modes estimated the locations every 0.5 second. Centre of mass mode used the first 5
highest probability calibration nodes to calculate the centre of mass, N cm = 5. Time
averaging used location estimation time of 0.5 second, Tu = 0.5 s and time averaging
window of 2 seconds, Tw = 2 s. The estimated locations ( xi , yi ), floors and logical
areas are recorded for performance analysis.
166
10
9
5
8
7
6
4
2
3
1
Legend
Testing Path 1
Testing Path 2
Testing Path 3
Testing Path 4
n
Direction of Path
Continue on Floor 2
Testing Location n
Figure 7.3: Testing locations and testing paths used for performance evaluation on
floor 1 of WCC, UTM
167
11
15
12
13
14
Legend
n
Testing Path 4 (Continued)
Testing Path 5
Testing Location n
Direction of Path
Continue on Floor 1
Figure 7.4: Testing locations and testing paths used for performance evaluation on
floor 2 of WCC, UTM
168
7.4.1.4 On-line Phase for Mobile Multi-floor Indoor Location Estimation
In Experiment 1b, the MT is moving with walking speed along a pre-defined
testing path. A total of 5 testing paths are defined at the test bed. Floor 1 contains 3
testing paths. Floor 2 contains 1 testing path. Floor 1 and floor 2 are connected with
1 testing path via the staircase. The test paths are illustrated in Figure 7.3 and 7.4. At
the beginning of each testing path, the MT initialized the proposed MILES
application and started the location estimation. Then the MT walked along the testing
path with average speed of 0.5 meter per second. The experiment is repeated twice
for each testing path.
Location estimations are carried out using three location estimation modes
separately. Discrete location and centre of mass modes estimated the locations every
0.5 second. Centre of mass mode used the first 5 highest probability calibration
nodes to calculate the centre of mass, N cm = 5. Time averaging used location
estimation time of 0.5 second, Tu = 0.5 s and time averaging window of 2 second, Tw
= 2 s. The estimated locations ( xi , yi ), floors and logical areas are recorded for
performance analysis.
169
7.4.2
Experiment 1a: Stationary Multi-floor Indoor Location Estimation
7.4.2.1 Accuracy and Precision
Figure 7.5 shows the cdf of location error using different location modes. In
this experiment, the discrete mode achieved an accuracy of 2.85 meters 75% of the
time and 4.86 meters 90% of the time. Centre of mass mode has a lower accuracy of
4.01 meters 75% of the time and 4.98 meters 90% of the time. Time averaging mode
has an accuracy of 2.85 meters 75% of the time and 4.56 meters 90% of the time.
Cumulative Probablity
1
0.9
0.8
0.7
0.6
Discrete
Center of Mass
Time Averaging
0.5
0
1
2
3
4
5
6
7
8
9
10
11
Location Error (m)
Figure 7.5: CDF of location error for stationary multi-floor location estimation
The average location error using different location modes is shown in Figure
7.6. From Figure 7.6, the discrete, centre of mass and time averaging modes achieved
an average accuracy of 1.13 meters, 1.67 meters, 1.21 meters respectively.
170
2.0
1600
1.8
1400
1200
1.4
1.2
1000
1.0
800
0.8
600
0.6
Time (s)
Location Error (m)
1.6
400
0.4
200
0.2
0.0
0
Discrete
Centre of Mass
Time Averaging
Location Mode
Average Location Error
Duration for 750 location estimates
Figure 7.6: Comparison of average location error and time duration for 750 location
estimates in stationary multi-floor location estimation
7.4.2.2 Percentage of Correct Logical Area Estimation and Percentage of
Correct Floor Estimation
Figure 7.7 shows the percentage of correct logical area estimation and
percentage of correct floor estimation using different location modes. From Figure
7.7, discrete, centre of mass and time averaging modes achieved 93.73%, 88.27%
and 96.80% of correct logical area estimation respectively during the stationary
experiment.
For floor estimation performance, all three location modes achieved 100% of
correct floor estimation during the stationary experiment, as depicted in Figure 7.7.
171
100
Percentage (%)
80
60
40
20
0
Discrete
Centre of Mass
Time Averaging
Location Mode
Correct Logical Area Estimation
Correct Floor Estimation
Figure 7.7: Percentage of correct logical area estimation and percentage of correct
floor estimation for stationary multi-floor location estimation
7.4.2.3 Discussion
The performance of the proposed system during stationary multi-floor
location estimation is summarized in Table 7.2. All three different modes achieved
an average accuracy lower than 2 meters during the stationary location estimation. In
term of location estimation error, performances of discrete and time averaging modes
are better than centre of mass mode. For precision of 75%, discrete and time
averaging mode are 28.93% more accurate than centre of mass mode.
From Table 7.2, discrete and time averaging modes have similar performance.
However, discrete mode is better than time averaging mode in term of location
estimation time required. For a given number of location estimates, discrete mode
only requires 25% of the total time duration required by time averaging mode. This is
an advantage for applications where short location estimation time is crucial such as
location based emergency services.
172
The percentage of correct logical area estimation is more than 80% for all
location modes. Most of the logical area estimation errors happened at junction of
hallways such as test location 5 in Figure 7.3 and test location 13 in Figure 7.4.
These areas are the border between different logical areas. Since the location
fingerprints at these areas are usually similar, the proposed system will mistakenly
estimate the MT to be in the nearby but wrong logical area.
In stationary multi-floor location estimation, the proposed system achieved
100% of correct floor estimation. When the MT is stationary, the RSS fingerprints
observed by the MT are stable. This helps the proposed system to estimate the floor
more effectively.
Table 7.2: Performance of three location modes during stationary multi-floor
location estimation
Statistics of Location Error (m)
Location
Mode
Time
Percentage Percentage
Duration
of Correct of Correct
Average
75th
90th
for 750
Logical
Floor
Percentile Percentile
location
Area
Estimation
estimates Estimation
(s)
(%)
(%)
Discrete
1.13
2.85
4.86
375
93.73
100.00
Center of
Mass
Time
Averaging
1.67
4.01
4.98
375
88.27
100.00
1.21
2.85
4.56
1500
96.80
100.00
173
7.4.3
Experiment 1b: Mobile Multi-floor Indoor Location Estimation
7.4.3.1 Accuracy and Precision
Figure 7.8 shows the cdf of location error using different location modes. In
this experiment, the discrete mode achieved an accuracy of 4.20 meters 75% of the
time and 5.58 meters 90% of the time. Centre of mass mode has a lower accuracy of
3.48 meters 75% of the time and 4.92 meters 90% of the time. Time averaging mode
has an accuracy of 3.12 meters 75% of the time and 4.54 meters 90% of the time.
1
0.9
Cumulative Probability
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
Discrete
Centre of Mass
Time Averaging
0
0
1
2
3
4
5
6
7
8
9
10
Location Error (m)
Figure 7.8: CDF of location error for mobile multi-floor location estimation
The average location error using different location modes is shown in Figure 7.9.
From Figure 7.9, the discrete, centre of mass and time averaging modes achieved an
average accuracy of 2.96 meters, 2.51 meters, 2.23 meters respectively.
3.5
0.8
3.0
0.7
0.6
2.5
0.5
2.0
0.4
1.5
0.3
1.0
0.2
0.5
0.1
0.0
Average Walking Speed (m/s)
Loction Error (m)
174
0
Discrete
Centre of Mass
Time Averaging
Location Mode
Average Location Error
Average Walking Speed
Figure 7.9: Comparison of average location error and average MT’s walking speed
during on-line phase for mobile multi-floor location estimation
7.4.3.2 Percentage of Correct Logical Area Estimation and Percentage of
Correct Floor Estimation
Figure 7.10 shows the percentage of correct logical area estimation and
percentage of correct floor estimation using different location modes. From Figure
7.10, discrete, centre of mass and time averaging modes achieved 72.34%, 77.80%
and 75.90% of correct logical area estimation respectively during the mobile
experiment.
For floor estimation performance, discrete, centre of mass and time averaging
modes achieved 99.43%, 99.63% and 100.00% of correct floor estimation
respectively during the mobile experiment, as depicted in Figure 7.10.
175
100
Percentage (%)
80
60
40
20
0
Discrete
Centre of Mass
Time Averaging
Location Mode
Correct Logical Area Estimation
Correct Floor Estimation
Figure 7.10: Percentage of correct logical area estimation and percentage of correct
floor estimation for mobile multi-floor location estimation
7.4.3.3 Discussion
The performance of the proposed system during mobile multi-floor location
estimation is summarized in Table 7.3. All three different modes achieved an average
accuracy lower than 3 meters during the mobile location estimation. In term of
location estimation error, performances of centre of mass and time averaging modes
are better than discrete mode. For precision of 75%, centre of mass mode and time
averaging modes are 11.83% and 18.64% more accurate than discrete mode
respectively.
From Table 7.3, centre of mass and time averaging modes have similar
performance. However, it must be noticed that the average MT’s walking speed
during centre of mass mode is 54.35% faster than time averaging mode. This
suggests that for a given accuracy level, centre of mass mode is suitable for
application where MT moves faster and time averaging mode is suitable for
application where MT moves slower.
176
The percentage of correct logical area estimation is more than 70% for all
location modes. Since the MT is moving from a given logical area to another logical
area, logical area estimation error usually happened at the border between logical
areas and also when the MT moved between different floors.
In mobile multi-floor location estimation, time averaging mode has 100% of
correct floor estimations. This is mainly due to the slower movement in time
averaging mode where the proposed system has longer time to estimate the floor.
Centre of mass and discrete mode have approximately 100% of correct floor
estimations. In these modes, the MT moved faster and floor estimation error usually
happened at area near the staircase. Since this area is an open area connecting the two
floors, the RSS fingerprints at this area may be similar and this may caused errors in
floor estimations.
Table 7.3: Performance of three location modes during mobile multi-floor location
estimation
Statistics of Location Error (m)
Location
Mode
Discrete
Center of
Mass
Time
Averaging
Average Percentage Percentage
MT’s
of Correct of Correct
Average
75th
90th
Walking
Logical
Floor
Percentile Percentile
Speed
Area
Estimation
during Estimation
On-line
(%)
(%)
Phase
(m/s)
2.96
4.20
5.58
0.73
72.34
99.43
2.51
3.48
4.92
0.71
77.80
99.63
2.23
3.12
4.54
0.46
75.90
100.00
177
7.5
Factors Affecting the Performance of the Proposed System
The effects of number of APs, number of calibration nodes, number of RSS
samples per fingerprint and off-line phase sampling interval on the performance of
the proposed system will be investigate in the following section.
7.5.1
Experimental Setup
Experiments are conducted at two areas within the test-bed used in previous
experiments (Figure 7.1 and Figure 7.2). The first test-bed is the rectangular hallway
in floor 1 as shown in Figure 7.11. The second test-bed is the area around the
staircase as shown in Figure 7.12. The topological nodes and testing locations for
each experiment are shown in Figure 7.11 and Figure 7.12.
The rectangular hallway test-bed in Figure 7.11 is used to study the
performance in terms of average accuracy and percentage of correct logical area
estimation. The staircase area test-bed in Figure 7.12 is used to study the
performance in terms of percentage of correct floor estimation. The off-line and online system parameters settings are similar to section 7.4.1.
178
Floor 1
Floor 1
(b)
(a)
Legend
Floor 1
Topological Node
Testing Location
Hallway_1
Hallway_2
Hallway_3
Hallway_4
(c)
Figure 7.11: Rectangular hallway test-bed with (a) 4 (b) 10 (c) 22 topological nodes
179
(a)
Floor 2
Floor 1
(b)
Floor 2
Floor 1
(c)
Floor 2
Legend
Floor 1
Topological Node
Testing Location
Figure 7.12: Staircase area test-bed with (a) 4 (b) 7 (c) 9 topological nodes
180
7.5.2
Experiment 2a: Effects of number of access points
The test-beds used for Experiment 2a are shown in Figure 7.11(c) and Figure
7.12(c). The location estimation performance is investigated with 1, 2, 3, 4 and 5 APs
respectively. The APs used are listed in Table 7.4 and their location is available in
Figure 4.5 and Figure 4.6. The results of experiment 2a will be presented and
discussed in the following sections.
Table 7.4: Access points used to investigate the effects of number of access points
on location estimation performance
Test-bed
Number of Access Points
Rectangular hallway
Staircase area
1
A3
A3
2
A3, A4
A3, B3
3
A3, A4, A8
A3, B3, A6
4
A3, A4, A8, A1
A3, B3, A6, B2
5
A3, A4, A8, A1, A12
A3, B3, A6, B2, B4
7.5.2.1 Accuracy
Figure 7.13 and Figure 7.14 show the average location error under different
number of APs during stationary and mobile location estimation respectively. From
these figures, the average location error is inversely proportional to the number of
APs installed for both stationary and mobile location estimation. Generally, for all
three location modes, increment in number of APs used increased the location
estimation accuracy. For example using discrete mode in Figure 7.13, the average
location error using one AP is 5.71 meters and average location error using five APs
is 0.28 meters, a 95.10% improvement in accuracy.
181
Average Location Error (m)
7
6
5
4
3
2
1
0
1
2
3
4
5
Number of Access Point
Discrete
Centre of Mass
Time Averaging
Figure 7.13: Effects of number of access points on average location error during
stationary location estimation
Average Location Error (m)
8
7
6
5
4
3
2
1
0
1
2
3
4
5
Number of Access Point
Discrete
Centre of Mass
Time Averaging
Figure 7.14: Effects of number of access points on average location error during
mobile location estimation
182
7.5.2.2 Percentage of Correct Logical Area Estimation
Figure 7.15 and Figure 7.16 show the percentage of correct logical area
estimation under different number of APs during stationary and mobile location
estimation respectively. From these figures, the percentage of correct logical area is
proportional to the number of APs installed for both stationary and mobile location
estimation. Generally, for all three location modes, increment in number of APs used
increased the percentage of correct logical area estimation. For example using
discrete mode in Figure 7.15, the percentage of correct logical area estimation is
increased by 36.80% when the number of APs changed from one to five.
100
Percentage (%)
80
60
40
20
0
1
2
3
4
5
Number of Access Point
Discrete
Centre of Mass
Time Averaging
Figure 7.15: Effects of number of access points on percentage of correct logical area
estimation during stationary location estimation
183
100
Percentage (%)
80
60
40
20
0
1
2
3
4
5
Number of Access Point
Discrete
Centre of Mass
Time Averaging
Figure 7.16: Effects of number of access points on percentage of correct logical area
estimation during mobile location estimation
7.5.2.3 Percentage of Correct Floor Estimation
Figure 7.17 shows the percentage of correct floor estimation under different
number of APs during stationary location estimation. From Figure 7.17, the
percentage of correct floor estimation is proportional to the number of APs installed.
Increment in number of APs used increases the percentage of correct floor estimation.
For example, the percentage of correct logical area estimation is increased by
90.00% when the number of APs changed from one to five.
184
100
Percentage (%)
80
60
40
20
0
1
2
3
4
5
Number of Access Point
Correct Floor Estimation
Figure 7.17: Effects of number of access points on percentage of correct floor
estimation during stationary location estimation
7.5.2.4 Discussion
Results from Experiment 2a show that more APs detected at a given location
will bring higher location estimation accuracy with better logical area and floor
estimations. Higher number of detected APs will increase the dimension of the
location fingerprint. This will directly increase the uniqueness of the fingerprint.
Improvement in accuracy is tremendous when number of APs increased from
one to two and from two to three. But with more than three APs, it only improved the
accuracy slightly. This suggests that there is a limit of improvement which can be
achieved by increasing the number of APs.
Therefore, for MILES based on WLAN RSS location fingerprinting, the
number of APs to be installed in the indoor environment is crucial in order to achieve
a targeted accuracy. Generally, higher number of APs detected at a given location
will bring higher accuracy. However, this will incur higher hardware and
maintenance costs.
185
7.5.3
Experiment 2b: Effects of number of topological nodes
In rectangular hallway test-bed, the effects of using 4, 10 and 22 topological
nodes are investigated as shown in Figure 7.11. This test-bed used five APs. In
staircase area test-bed, the effects of 4, 7 and 9 topological nodes are investigated as
shown in Figure 7.12. This test-bed used four APs. The results of experiment 2b will
be presented and discussed in the following sections.
7.5.3.1 Accuracy
Figure 7.18 and Figure 7.19 show the average location error under different
number of topological nodes during stationary and mobile location estimation
respectively. From these figures, the average location error is inversely proportional
to the number of topological nodes for both stationary and mobile location estimation.
For example using discrete mode in Figure 7.18, the average location error using 4
topological nodes is 2.76 meters and average location error using 22 topological
nodes is 0.28 meter, an 89.86% of improvement in accuracy.
Average Location Error (m)
3
2.5
2
1.5
1
0.5
0
4
10
22
Number
NumberofofCalibration
TopologicalNode
Node
Discrete
Centre of Mass
Time Averaging
Figure 7.18: Effects of number of topological nodes on average location error during
stationary location estimation
186
Average Location Error (m)
3.5
3
2.5
2
1.5
1
0.5
0
4
10
22
NumberofofCalibration
TopologicalNode
Node
Number
Discrete
Centre of Mass
Time Averaging
Figure 7.19: Effects of number of topological nodes on average location error during
mobile location estimation
7.5.3.2 Percentage of Correct Logical Area Estimation
Figure 7.20 and Figure 7.21 show the percentage of correct logical area
estimation under different number of topological nodes during stationary and mobile
location estimation respectively. From these figures, the percentage of correct logical
area is generally proportional to the number of topological nodes for both stationary
and mobile location estimation. Increment in number of topological nodes increases
the percentage of correct logical area estimation. For example using discrete mode in
Figure 7.20, the percentage of correct logical area estimation is increased by 40.00%
when the number of topological nodes changed from 4 to 22.
187
100
Percentage (%)
80
60
40
20
0
4
10
22
Number
Numberof
ofCalibration
TopologicalNode
Node
Discrete
Centre of Mass
Time Averaging
Figure 7.20: Effects of number of topological nodes on percentage of correct logical
area estimation during stationary location estimation
100
Percentage (%)
80
60
40
20
0
4
10
22
Number of
of Topological
Calibration Node
Number
Node
Discrete
Centre of Mass
Time Averaging
Figure 7.21: Effects of number of topological nodes on percentage of correct logical
area estimation during mobile location estimation
188
7.5.3.3 Percentage of Correct Floor Estimation
Figure 7.22 shows the percentage of correct floor estimation under different
number of topological nodes during stationary location estimation. From Figure 7.22,
the percentage of correct floor estimation is proportional to the number of
topological nodes. Increment in number of topological nodes increases the
percentage of correct floor estimation. For example, the percentage of correct floor
estimation is increased by 47.00% when the number of topological nodes changed
from 4 to 9.
100
Percentage (%)
80
60
40
20
0
4
7
9
Number
Number of
of Calibration
TopologicalNode
Node
Correct Floor Estimation
Figure 7.22: Effects of number topological nodes on percentage of correct floor
estimation during stationary location estimation
7.5.3.4 Discussion
Results from Experiment 2b show that, more topological nodes will bring
higher location estimation accuracy with better logical area and floor estimations.
Higher number of topological nodes will increase the number of location fingerprints
189
collected during off-line phase. Therefore, the radio map created will represent the
indoor environment more accurately.
However, referring to Figure 7.21, results for mobile location estimation
using centre of mass mode show that accuracy obtained from 10 topological nodes
are better than 22 topological nodes. Possible reasons for this result will not be
investigated in here. This may suggest that there is a limit of improvement which can
be achieved by increasing the density of the topological nodes.
Therefore, for MILES based on WLAN RSS location fingerprinting, the
number of topological nodes tabulated around the indoor environment is crucial in
order to achieve a targeted accuracy. Generally, higher number of topological nodes
will bring higher accuracy. However, this will need more off-line location
fingerprinting time and efforts. Besides that, higher number of topological nodes will
increase the size of the radio map which will consume more data storage spaces. This
will be a constraint for MT with limited data storage spaces such as PDA.
7.5.4
Experiment 2c: Effects of number of RSS samples per fingerprint
The test-bed used for Experiment 2c is shown in Figure 7.11(c). The location
estimation performance is investigated with 20, 50 and 100 RSS samples per
fingerprint respectively. The results of experiment 2c will be presented and discussed
in the following sections.
7.5.4.1 Accuracy
Figure 7.23 shows the average location error under different number of RSS
samples per fingerprint during stationary location estimation. From Figure 7.23, the
average location error is inversely proportional to the number of RSS samples per
fingerprint. Increment in number of RSS samples per fingerprint decreases the
190
averaged location error. In Figure 7.23, the average location error using 20 RSS
samples per fingerprint is 2.83 meters and average location error using 100 RSS
3
12000
2.5
10000
2
8000
1.5
6000
1
4000
0.5
2000
0
Time (ms)
Average Location Error (m)
samples per fingerprint is 1.92 meters, a 32.16% of improvement in accuracy.
0
20
50
100
Number of Samples Per Fingerprint
Average
Mesurement Duration at Each Calibration Node
Figure 7.23: Effects of number of RSS samples per fingerprint on average location
error during stationary location estimation
7.5.4.2 Percentage of Correct Logical Area Estimation
Figure 7.24 shows the percentage of correct logical area estimation under
different number of RSS samples per fingerprint during stationary location
estimation. From Figure 7.24, the percentage of correct logical area is proportional to
the number of RSS samples per fingerprint. Increment in number of RSS samples per
fingerprint increases the percentage of correct logical area estimation. In Figure 7.24,
the percentage of correct logical area estimation is increased by 19.20% when the
number of RSS samples per fingerprint changed from 20 to 100.
191
100
Percentage (%)
80
60
40
20
0
20
50
100
Number of Samples Per Fingerprint
Correct Logical Area Estimation
Figure 7.24: Effects of number of RSS samples per fingerprint on percentage of
correct logical area estimation during stationary location estimation.
7.5.4.3 Discussion
Results from Experiment 2c show that, more RSS samples per location
fingerprint will bring higher location estimation accuracy with better logical area
estimations. Higher number of RSS samples per location fingerprint will help to
construct more accurate RSS histograms. RSS histograms with more RSS samples
will capture the RSS distribution at a given location more accurately.
Therefore, for MILES based on WLAN RSS location fingerprinting, the
number of RSS samples collected per location fingerprint at a given calibration node
is crucial in order to achieve a targeted accuracy. Generally, higher number of RSS
samples per location fingerprint will bring higher accuracy. However, this will need
more off-line location fingerprinting time and efforts. For instance, referring to
Figure 7.23, time duration needed to collect 20 RSS samples per location fingerprint
is only 20% of the time required by 100 RSS samples per location fingerprint.
Besides that, higher number of RSS samples per location fingerprint will increase the
size of the radio map which will consume more data storage spaces. This will be a
constraint for MT with limited data storage spaces such as PDA.
192
7.5.5
Experiment 2d: Effects of off-line phase sampling interval
The test-bed used for Experiment 2d is shown in Figure 7.11(c). The location
estimation performance is investigated with 10 milliseconds, 50 milliseconds and
100 milliseconds of off-line phase sampling interval respectively. The results of
experiment 2d will be presented and discussed in the following sections.
7.5.5.1 Accuracy
Figure 7.25 shows the average location error under different off-line phase
sampling intervals during stationary location estimation. From Figure 7.25, the
average location error is inversely proportional to the off-line phase sampling
interval. Increment in off-line phase sampling interval decreases the averaged
location error. In Figure 7.25, the average location error using 10 milliseconds of offline phase sampling interval is 3.44 meters and average location error using 100
milliseconds of off-line phase sampling interval is 2.08 meters, a 39.53% of
improvement in accuracy.
6000
3.5
5000
3
4000
2.5
2
3000
1.5
Time (ms)
Average Location Error (m)
4
2000
1
1000
0.5
0
0
10
50
100
Sampling Interval (ms)
Average
Measurement Duration at Each Calibration Node
Figure 7.25: Effects of off-line phase sampling interval on average location error
during stationary location estimation
193
7.5.5.2 Percentage of Correct Logical Area Estimation
Figure 7.26 shows the percentage of correct logical area estimation under
different off-line phase sampling intervals during stationary location estimation.
From Figure 7.26, the percentage of correct logical area is proportional to the off-line
phase sampling interval. Increment in off-line phase sampling interval increases the
percentage of correct logical area estimation. In Figure 7.26, the percentage of
correct logical area estimation is increased by 22.00% when the off-line phase
sampling interval changed from 10 milliseconds to 100 milliseconds.
100
Percentage (%)
80
60
40
20
0
10
50
100
Sampling Interval (ms)
Correct Logical Area Estimation
Figure 7.26: Effects of off-line phase sampling interval on percentage of correct
logical area estimation during stationary location estimation
7.5.5.3 Discussion
Results from Experiment 2d show that, performance of higher off-line
sampling interval will bring higher location estimation accuracy with better logical
area estimations. The possible reason for this improvement is that, longer sampling
interval between consecutive RSS samples helps to capture more accurately the RSS
characteristics and distribution at a given location.
194
Therefore, for MILES based on WLAN RSS fingerprinting, the off-line
sampling interval is crucial in order to achieve a targeted accuracy. Longer off-line
sampling interval will bring higher accuracy. However, this will require more offline location fingerprinting time. For instance, referring to Figure 7.25, in order to
collect 50 RSS samples per location fingerprint, system with sampling interval of 10
milliseconds only needs 10% of the time required by system with sampling interval
of 100 milliseconds.
7.6
Summary
The proposed system is evaluated in real-time in the test-bed. Firstly,
experiments are conducted to evaluate the performance of the proposed system in
terms of stationary and mobile multi-floor indoor location estimation. Besides that,
performances of the three location modes proposed are validated. Finally, four
factors which affect the performance of the proposed system are investigated.
In stationary experiments, the best accuracy achieved by the proposed system
is 2.85 meters with 75% precision and 4.56 meters with 90% precision. On the other
hand, in mobile experiments, the best accuracy achieved is 3.12 meters with 75%
precision and 4.54 meters with 90% precision. Both experiments achieved a 100% of
correct floor estimation. The proposed system performed better in stationary situation.
The proposed system managed to achieve targeted accuracy of less than 5 meters
with 90% precision for both stationary and mobile multi-floor indoor location
estimation.
Performance of discrete, centre of mass and time averaging modes are
compared. The comparison discovered that every mode perform best in different
environments. For instance, discrete mode has high performance during stationary
experiment and centre of mass mode excelled in mobile experiment. This suggests
that location modes should be chosen according to application environment for
optimum performance.
195
Investigations also discovered that number of WLAN APs, number of
topological nodes, number of RSS samples collected per location fingerprint and offline phase sampling interval affected the performance of the proposed system in a
measurable way. Settings of these factors should be considered based on the targeted
accuracy and precision in the initial stage of system deployment.
The proposed system is successfully evaluated. The next chapter will
conclude this study and provide future improvements for the proposed system.
CHAPTER 8
CONCLUSION
8.1
Conclusion
The main goal of this study is to design and develop an indoor LES for multifloor environment using off-the-shelf IEEE 802.11g WLAN infrastructures. Aligned
with the design and development processes, four objectives are identified for this
study.
At the initial stage of the study, characteristics of WLAN RSS in indoor
environment for WLAN-based indoor location estimation application are studied
through a series of measurements. From these measurements, the feasibility of
WLAN RSS locating fingerprinting technique is proven. In addition, findings
obtained from these measurements are applied in the design, development and
deployment of the proposed system.
Based on the literature review and understanding of the WLAN RSS
characteristics, the software-based proposed system is design and developed. The
technical specifications, test-bed, hardware and software components are identified
and setup. RSS histograms collected in the off-line phase are adopted as the location
fingerprint. Probabilistic Bayesian filtering framework is then applied to estimate the
MT’s location in the on-line phase. Topological Markov localization algorithm is
selected to implement the Bayesian filter. Three location modes: discrete, centre of
mass and time averaging are introduced to cater different requirements in different
197
environments. The estimated physical and symbolic location information are
reported to the end-user in textual and also graphical illustration.
Next, prototype of the proposed system is validated through real-time
experiments. The proposed system is required to estimate the location of stationary
and mobile user in a two-floor building. The proposed system is targeted to provide
accuracy less than 5 meters with 90% of precision and 90% of correct floor
estimation. Results from the experiments verified that the proposed system has
achieved the targeted performance. In stationary experiments, the best accuracy
achieved by the proposed system is 2.85 meters with 75% precision and 4.56 meters
with 90% precision. On the other hand, in mobile experiments, the best accuracy
achieved is 3.12 meters with 75% precision and 4.54 meters with 90% precision.
Both experiments achieved a 100% of correct floor estimation. Since the proposed
system is only evaluated in WCC, UTM, the results obtained here are only applicable
to the given two-floor test-bed and the settings used during off-line and on-line
phases.
Finally, various factors which will affect the performance of the proposed
system are studied. Four factors investigated are number of WLAN APs, number of
topological nodes, number of RSS samples collected per location fingerprint and offline phase sampling interval. Results from experiments indicate that these factors
will affect the performance of the proposed system in a measurable way. Therefore,
settings for these factors should be considered before the deployment of the proposed
system. Guidelines on setting these factors are briefly discussed.
All objectives of this study are successfully accomplished. A WLAN-based
indoor LES for multi-floor environment using RSS location fingerprinting technique
is successfully developed. The proposed system can be implemented in any multifloor buildings which are equipped with WLAN. However, the performance of the
proposed system may varied due to different building structures, building materials,
environments, off-line and on-line phases settings and others.
198
8.2 Future Works
Works in this study provides a platform for the future researches on efficient
indoor LESs. The prototype of the proposed system still has limitations which can be
further improved in future. Possible future development directions of this study are
identified and listed as below:
(i)
Time dependency of radio map
In this study, the proposed system uses a single radio map for different environments
such as office hour and off-office hour. The environment is assumed unchanged
during the experiments. However in reality, a MILES should be able to achieved
same location estimation performance in different environments. Thus, future
research should address this issue. One of the possible solutions is automatic
environmental profiling where different radio maps will be selected automatically
based on observations of the current environment situation.
(ii)
Sharing of radio map
The proposed system requires both off-line and on-line phases to be conducted with
same WLAN client devices. This is mainly because RSS values measured by
different WLAN client devices may vary. Thus, radio map created with a given
WLAN client device is suitable to be used by itself only. If other WLAN client
devices use this radio map, the location estimation performance may not be
optimum. Therefore, the solution of sharing a single radio map for different MTs
should be incorporated in the proposed system in future. If the sharing of radio map
is possible, then the off-line phase only needed to be conducted by one MT and the
radio map created can be shared by different WLAN client devices.
(iii) Reducing off-line calibration efforts
Location fingerprinting provides better performance compared to other techniques.
However, the main disadvantage of location fingerprinting technique is the time and
energy consuming off-line phase. Efforts required for off-line phase will grow
proportionally with the size of indoor area to be covered by the proposed system. In
addition, changes in the environment or movement of the AP will require re-
199
calibration in order to obtain a optimum radio map. Therefore, solutions on how to
reduce and simplify this off-line phase process are an interesting enhancement to the
proposed system.
(iv) Further testing and modification
In this study, the proposed system only evaluated in a two-floor building. For future
development, more experiments in different test-bed with more floors should be
conducted to study the limitations and weaknesses of the proposed system. In
addition, besides the factors studied here, there are many other factors which will
affect the performance of WLAN RSS location fingerprinting system. Factors such
as location of APs, design of topological map structure, design of logical area and
others should be investigated. After further investigations on factors affecting the
performance of WLAN RSS location fingerprinting system, a user guideline can be
listed in order to assist the user in planning, implementing and improving WLAN
RSS location fingerprinting system.
(v)
Integration with other wireless technologies
The proposed system is implemented solely on top of IEEE 802.11g WLAN
infrastructures. In order to improve the proposed system performance and scalability,
it can be integrated with other wireless technologies such as Bluetooth, RFID and
ZigBee. By integrating two or more wireless-based indoor LESs into a single
platform, the location estimation accuracy can be improved. Besides that, the
proposed system currently only covers indoor area. In future, it can be incorporated
with outdoor LES to become a universal LES. Outdoor system such as GPS will
cover outdoor area and WLAN-based system will operate whenever GPS signal is
not available.
REFERENCES
1. Djuknic, G.M. and Richton, R.E. Geolocation and Assisted GPS. IEEE Computer
Magazine. 2001. 34(2): 123-125.
2. Guolin Sun, Jie Chen, Wei Guo and Liu, K. J. R. Signal Processing Techniques
in Network-Aided Positioning: A Survey of State-of-the-Art Positioning Designs.
IEEE Signal Processing Magazine. 2005. 22(4): 12-23.
3. Kolodziej, K. W. and Hjelm, J. Local Positioning System. Florida: CRC Press.
2006.
4. Bahl, P. and Padmanabhan, V. N. RADAR: An In-Building RF-Based User
Location and Tracking System. Proceedings of the 19th Annual Joint Conference
of the IEEE Computer and Communications Societies. March 26-30, 2000. TelAviv, Israel: IEEE. 2000. 775-784.
5. Bahl, P., Padmanabhan, V. N. and Balachandran, A. Enhancements to the
RADAR User Location and Tracking System. Microsoft Research Technical
Report: MSR-TR-00-12. 2000.
6. Ladd, A. M., Bekris, K. E., Rudys, A., Kavraki, L. E., Wallach, D. S. and
Marceau, G. Robotics-based Location Sensing Using Wireless Ethernet.
Proceedings of the 8th Annual International Conference on Mobile Computing
and Networking. September 23-28, 2002. Atlanta, Georgia: ACM Press. 227-238.
7. Gentile, C. and Klein-Berndt, L. Robust Location Using System Dynamics and
Motion Constraints. The 2004 IEEE International Conference on
Communications. June 20-24, 2004. Paris: IEEE. 2004. 1360-1364. \
8. Pahlavan, K., Xinrong Li and Makela, J. P. Indoor Geolocation Science and
Technology. IEEE Communications Magazine. 2002. 40(2): 112-118.
9. Siegwart, R. and Nourbakhsh, I. R. Introduction to Autonomous Mobile Robots.
London: The MIT Press. 2004.
10. Hightower, J. and Borriello, G. Location Systems for Ubiquitous Computing.
IEEE Computer Magazine. 2001. 34(8): 57-66.
201
11. Kupper, A. Location-Based Services: Fundamentals and Operation. West
Sussex: John Wiley & Sons Ltd. 2005.
12. Want, R., Hopper, A., Falcao, V. and Gibbons, J. The Active Badge Location
System. ACM Transactions on Information Systems. 1992. 10(1): 91-102.
13. Harter, A., Hopper, A., Steggles, P., Ward, A. and Webster, P. The Anatomy of
A Context-Aware Application. Proceedings of the 5th Annual ACM/IEEE
International Conference on Mobile Computing and Networking. August 15-19,
1999. Seattle, Washington: ACM Press. 1999. 59-68.
14. Addlesee, M., Curwen, R., Hodges, S., Newman, J., Steggles, P., Ward, A. and
Hopper, A. Implementing a Sentient Computing System. IEEE Computer
Magazine. 2001. 34(8): 50-56.
15. Priyantha, N. B., Chakraborty, A. and Balakrishnan, H. The Cricket LocationSupport System. Proceedings of the 6th Annual International Conference on
Mobile Computing and Networking. August 6-11, 2000. Boston, Massachusetts:
ACM Press. 2000. 32-43.
16. Fontana, R. J. and Gunderson, S. J. Ultra-Wideband Precision Asset Location
System. 2002 IEEE Conference on Ultra Wideband Systems and Technologies.
May 21-23, 2002. Baltimore, Maryland: IEEE. 2002. 147-150.
17. Fontana, R. J., Richley, E. and Barney, J. Commercialization of an Ultra
Wideband Precision Asset Location System. 2003 IEEE Conference on Ultra
Wideband Systems and Technologies. Nov 16-19, 2003. Reston, Virginia: IEEE.
2003. 369-373.
18. Ni, L. M., Yunhao Liu, Yiu Cho Lau and Patil, A. P. LANDMARC: Indoor
Location Sensing Using Active RFID. Proceedings of the 1st IEEE International
Conference on Pervasive Computing and Communications. March 23-26, 2003.
Texas: IEEE. 2003. 407-415.
19. Gast, M. S. 802.11 Wireless Networks: The Definitive Guide. 2nd. ed.
Sebastopol, California: O’Reilly Media. 2005.
20. Prasithsangaree, P., Krishnamurthy, P. and Chrysanthis, P. On Indoor Position
Location with Wireless LANs. The 13th IEEE International Symposium on
Personal, Indoor and Mobile Radio Communications 2002. September 15-18,
2002. Lisboa, Portugal: IEEE. 2002. 720-724.
202
21. Kanaan, M. and Pahlavan, K. A Comparison of Wireless Geolocation Algorithms
in the Indoor Environment. IEEE Wireless Communications and Networking
Conference 2004. March 21-25, 2004. Atlanta, Georgia: IEEE, 2004. 177-182.
22. Saha, S., Chaudhuri, K., Sanghi, D. and Bhagwat, P. Location Determination of
A Mobile Device Using IEEE 802.11b Access Point Signals. IEEE Wireless
Communications and Networking 2003. March 16-20, 2003. New Orleans,
Louisiana: IEEE. 2003. 1987-1992.
23. Battiti, R. and Brunato, M. Statistical Learning Theory for Location
Fingerprinting in Wireless LANs. Computer Networks. 2005. 47(6): 825-845.
24. Roos, T., Myllymaki, P., Tirri, H., Misikangas, P. and Sievanen, J. A
Probabilistic Approach to WLAN User Location Estimation. International
Journal of Wireless Information Networks. 2002. 9(3): 155-164.
25. Fox, D., Hightower, J., Lin Liao, Schulz, D. and Borriello, G. Bayesian Filtering
For Location Estimation. IEEE Pervasive Computing. 2003. 2(3): 24-33.
26. Krumm, J. Probabilistic Inferencing for Location. Proceedings of the 2003
Workshop on Location-Aware Computing. October 12, 2003. Seattle,
Washington: Microsoft Research. 2003. 25-27.
27. Fox, D., Hightower, J., Kautz, H., Lin Liao and Patterson, D. Bayesian
Techniques for Location Estimation. Proceedings of the 2003 Workshop on
Location-Aware Computing. October 12, 2003. Seattle, Washington: Microsoft
Research. 2003. 16-18.
28. Jan, R. H. and Lee, Y. R. An Indoor Geolocation System for Wireless LANs.
Proceeding of the 32th International Conference on Parallel Processing
Workshops 2003. October 6-9, 2003. Kaohsiung, Taiwan: IEEE. 2003. 29-34.
29. Wong, W. H., Ng, J. K. and Yeung, W. M. Wireless LAN Positioning with
Mobile Devices in a Library Environment. 25th IEEE International Conference
on Distributed Computing Systems Workshops 2005. June 6-10, 2005. Columbus,
Ohio: IEEE. 633-636.
30. Harder, A., Song, L. and Wang, Y. Towards an Indoor Location System Using
RF Signal Strength in IEEE 802.11 Networks. International Conference on
Information Technology: Coding and Computing 2005. April 4-6, 2005. Las
Vegas, Nevada: IEEE. 228-233.
203
31. Castro, P., Chiu, P., Kremenek, T. and Muntz, R. A Probabilistic Room Location
Service for Wireless Networked Environments. Proceedings of the 3th
International Conference on Ubiquitous Computing. September 30 - October 2,
2001. Atlanta, Georgia: Springer Berlin. 2001. 18-34.
32. Youssef, M. and Agrawala, A. The Horus WLAN Location Determination
System. Proceedings of the 3rd International Conference on Mobile Systems,
Applications, and Services. June 6-8, 2005. Seattle, Washington: ACM Press.
205-218.
33. Youssef, M. and Agrawala, A. Handling Samples Correlation in the Horus
System. The 23th Annual Joint Conference of the IEEE Computer and
Communications Society. March 7-11, 2004. Hong Kong: IEEE. 2004. 10231031.
34. Youssef, M. and Agrawala, A. Small-scale Compensation for WLAN Location
Determination Systems. IEEE Wireless Communications and Networking
Conference 2003. March 16-20, 2003. New Orleans, Louisiana: IEEE. 2003.
1974-1978.
35. Youssef, M. A., Agrawala, A. and Shankar, A. U. WLAN Location
Determination via Clustering and Probability Distributions. Proceedings of the
1st IEEE International Conference on Pervasive Computing and
Communications 2003. March 23-26, 2003. Texas: IEEE. 2003. 143-150.
36. Xiang, Z., Song, S., Chen, J., Wang, H., Huang, J. and Gao, X. A Wireless LANbased Indoor Positioning Technology. IBM Journal of Research and
Development. 2004. 48(5/6). 617-626.
37. Krishnan, P., Krishnakumar, A. S., Ju, W. H., Mallows, C. and Gamt, S. N. A
System for LEASE: Location Estimation Assisted by Stationary Emitters for
Indoor RF Wireless Networks. The 23rd Annual Joint Conference of the IEEE
Computer and Communications Society. March 7-11, 2004. Hong Kong: IEEE.
2004. 1001-1011.
38. Komar, C. and Ersoy, C. Location Tracking and Location Based Service Using
IEEE 802.11 WLAN Infrastructure. European Wireless 2004. February 24-27,
2004. Barcelona, Spain: 2004. 1-7.
39. Hashemi, H. The Indoor Radio Propagation Channel. Proceedings of the IEEE.
1993. 81(7): 943-968.
204
40. Kaemarungsi, K. and Krishnamurthy, P. Properties of indoor received signal
strength for WLAN location fingerprinting. The 1st Annual International
Conference on Mobile and Ubiquitous Systems: Networking and Services 2004.
August 22-26, 2004. Boston, Massachusetts: IEEE. 2004. 14-23.
41. Ghaddar, M., Talbi, L. and Denidni, T.A. Human Body Modeling for Prediction
of Effect of People on Indoor Propagation Channel. IEE Electronics Letters.
2004. 40(25): 1592-1594.
42. Patil, A. P., Kim, D. J., and Ni, L. M. A Study of Frequency Interference and
Indoor Location Sensing with 802.11b and Bluetooth Technologies. Wireless
Telecommunications Symposium 2005. April 28-30, 2005. Pomona, California:
IEEE. 2005. 174-183.
43. Institute of Electrical and Electronics Engineers. IEEE Std 802.11-1997
Information Technology- telecommunications And Information exchange
Between Systems-Local And Metropolitan Area Networks-specific Requirementspart 11: Wireless Lan Medium Access Control (MAC) And Physical Layer (PHY)
Specifications. New York. IEEE Std 802.11-1997. 1997.
44. Bardwell, J. Converting Signal Strength Percentage to dBm values. White Paper.
November 2002.
45. Fox, D. Markov Localization: A Probabilistic Framework for Mobile Robot
Localization and Navigation. Ph. D. University of Bonn, Germany: 1998.
46. Youssef, M. and Agrawala, A. Continuous Space Estimation for WLAN
Location Determination Systems. Proceedings of the 13th International
Conference on Computer Communications and Networks 2004. October 11-13,
2004. Chicago, Illinois: IEEE. 2004. 161-166.
47. LaMarca, A., Chawathe, Y., Consolvo, S., Hightower, J., Smith, T., Scott, J.,
Sohn, T., Howard, J., Hughes, J., Potter, F., Tabert, J., Powledge, P., Borriello, G.
and Schilit, B. Place Lab: Device Positioning Using Radio Beacons in the Wild.
Proceedings of the 3rd International Conference on Pervasive Computing 2005.
May 8-13, 2005. Munich, Germany: Springer. 2005. 116-133.
48. JGraphT Development Team. JGraphT version 0.5.3.
http://jgrapht.sourceforge.net/. January 3, 2005.
APPENDIX A
GRAPHICAL USER INTERFACE
OF THE PROPOSED SYSTEM
Main GUI
206
Menu: File
Menu: File
Submenu
Icon
Function
New Model
Create new environmental
model.
Open Model
Open existing
environmental model.
Save Model
Save current
environmental model.
New Floor Map
Connect Floor
Add new floor map to
current environmental
model.
Create floor connection in
topological map
207
Menu: Edit
Menu: Edit
Submenu
Icon
Function
Create topological map.
Define Topological Map
Create logical area.
Define Logical Area
Topological Map
Information
Show topological map
information.
Calibrate All Nodes
Calibrate all topological
nodes.
Go to Next Node
Go to next calibration
node for calibration.
(Re)Calibrate One Node
(Re)Calibrate one selected
topological nodes.
Untrain All Nodes
Start Recording Signals
Discard all location
fingerprints in existing
environmental model.
Start recoding WLAN
RSS.
Start indoor location
estimation.
Location Mode
Stop indoor location
estimation.
208
Menu: Options
Menu: Options
Submenu
Icon
Function
Location Options
Select type of location
modes.
Enable Analyzer
Enable and display
histogram analyzer.
Disable Analyzer
Disable histogram
analyzer.
Show Tutorial Messages
Show help messages.
Hide Tutorial Messages
Hide help messages.
APPENDIX B
AWARD
Event
: British Invention Show 2005 (BIS 2005)
Date
: 20 – 23 October 2005
Venue
: Alexandra Palace, London.
Organizer
: British Inventors Society, London.
Title
: WiFiGeoLoc : A Multi-Floor IEEE 802.11 b/g Wireless Local
Area Network Based Indoor Geo-Location System
Award
: Silver medal, International Invention of the Year Awards.
APPENDIX C
BAYES’ THEOREM
The mathematical foundations of Bayes’ theorem are presented here. The
Bayes’ theorem is derived from the definition of the relationship between the
conditional and joint probability of random variables A and B.
Let B be an event such that probability of B, P(B) > 0. Then for any event A, the
conditional probability of A given B, P ( A | B ) can be defined as:
P( A | B) =
P( A ∩ B)
P( B)
(C.1)
where the joint probability of A and B, P ( A ∩ B) is given by:
P ( A ∩ B ) = P ( A | B ) P ( B ) = P ( B | A) P ( A)
(C.2)
Then Bayes’s theorem can be sated as:
P( A | B) =
P( B | A) P( A)
P( B)
(C.3)
Bayes’ theorem is capable to infer the conditional probability P ( A | B ) from the
conditional probability P ( B | A) .
Equation (C.3) can be expressed in discrete form as:
P (ai | b j ) =
P (b j | ai ) P (ai )
P (b j )
;
i = 0,1, 2,..., ( M − 1) ;
where M represents the total number of element in set A.
j = 0,1, 2,...
(C.4)