Multilevel Depth and Image Fusion for Human Activity Detection

Low resolution pedestrian
detection using light robust
features and hierarchical system
Yun-FuLiu
Jing-MingGuo
Che-HaoChang
Pattern Recognition,
Volume 47, Issue 4, April 2014, pp 1616–1625
Reporter: LIM ZI YI
Outline
Introduction
 Problem descriptions
 Proposed probability-based pedestrian
mask pre-filtering
 Orientation/magnitude-based AdaBoost
algorithm
 Experimental results
 Conclusion
 Personal Remark

2
Introduction

How to deal with the prospective
complexities of background and light
conditions is still a key issue in this field.
3
Introduction


Methods with temporal information
consider movement or depth information
have an inherent drawback that when the
camera is moving or pedestrian is not
standing still, the backgroundand foreground
will be misclassified.
common and intuitive features, e.g., color
and brightness, the performance can be
easily affected by the dresses of pedestrians
or environmental lighting conditions.
4
Introduction
This study presents a hybrid pedestrian
detection scheme for significantly
reducing detection time.
 two topics are focused in this work:
(1) The feature descriptors for
pedestrians, and
(2) the resolution of a pedestrian image.

5
Problem descriptions
Features for pedestrian
 Sample resolution

6
Features for pedestrian
Fig. 2. Conceptual diagram of the edgelet feature applied to pedestrians.
7
Sample resolution

Another issue which directly affects
driving safety is the practical detection
distance.
Fig. 4. Relationship between the height of pedestrian and the distance from camera.
8
Sample resolution
Fig. 5.The difference between the same pedestrian captured with
different resolutions.
9
Proposed probability-based
pedestrian mask pre-filtering(PPMPF)
Initial weight mask generation
 Refinement

Fig. 6.Algorithm of the proposed pedestrian detection scheme.
10
Initial weight mask generation

Two main components of the first stage:
◦ probability-based pedestrian mask
◦ the corresponding probability table
(histogram table)

This study uses a convolutional pedestrian
mask detection strategy
11
Initial weight mask generation
12
Initial weight mask generation
(4)
(5)
13
Initial weight mask generation
Fig. 7. Pedestrian pre-filtering by orientation mask.
14
Refinement
15
Refinement
Fig. 8. Pedestrian pre-filtering by magnitude mask.
16
Orientation/magnitude-based
AdaBoost algorithm
Features
 Training procedure
 Detection process

Fig. 9. Flowchart of the proposed training process.
17
Features
(10)
(11)
18
Training procedure

The AdaBoost can be divided into two
parts:
◦ Training
◦ voting.

the process trains the sorted datasets to
obtain the classifiers, and uses the trained
classifiers to construct a strong classifier
to predict the samples of interest.
19
Training procedure
In this training procedure, the iteration
updates the weights to upgrade the effect
of the weak classifier to a strong classifier.
 As the training process is completed, a
voting mechanism is applied to form a
strong classifier which can yield a precise
detection capability.

20
Training procedure
Fig. 10. Orientation/magnitude-based features obtained from the average
value of the blocks which are not easily affected during walking.
21
Detection process
Fig. 11. Flowchart of the orientation/magnitude-based detection process.
22
Experimental results

Datasets:
◦ INRIA
◦ Daimler Chrysler Pedestrian dataset
(15)
(16)
23
Experimental results
Fig. 12. Precision-recall plots with different features and the corresponding
derived ways (1-D and the Sobel edge detector are considered).
24
Experimental results
Fig. 13. Precision-recall performances of the HOG, Haar-like, edgelet , and
the proposed feature obtained with
(a) the INRIA dataset and
(b) the Daimler Chrysler Pedestrian dataset.
25
Experimental results
Fig. 14. Practical detection results with various features under diverse scenarios.
26
Experimental results
Table1 Detection speed comparison.
Detect
windows (s)
HOG
Haar-like
features
Edgelet
features
Proposed
feature
Proposed entire
system (with PPMPF)
62.35
957.78
1364.07
1339.17
3773.58
27
Conclusion
lighting change and low- resolution
scenario, are considered simultaneously
for pedestrian detection application.
 the proposed scheme can offer an
accuracy similar to that of the HOG,
while the proposed scheme is much faster
 the processing efficiency is similar to that
of the edgelet, while the propose scheme
is of higher accuracy.

28
Personal Remark
On Table 1, Author didn’t explain what the
unit is.
 We can use other features like depth
information to increase accuracy(but time
of computing will become slower).
 If need to apply such system on driving
safety system, we can use laser radar to
replace this type of method.

29
Thank you
for your listening !
30
Q&A
31