2011 IEEE International Conference on Robotics and Automation
Shanghai International Conference Center
May 9-13, 2011, Shanghai, China
Water Column Current Profile Aided Localisation combined with
View-based SLAM for Autonomous Underwater Vehicle Navigation
Lashika Medagoda, Stefan B. Williams, Oscar Pizarro , Michael V. Jakuba
Australian Centre for Field Robotics
University of Sydney, Sydney NSW 2006 Australia
{l.medagoda,stefanw,o.pizarro,m.jakuba}@acfr.usyd.edu.au
Abstract— Survey class Autonomous Underwater Vehicles
(AUVs) rely on Doppler Velocity Logs (DVL) for precise
navigation near the seafloor. In cases where the seafloor depth
is greater than the DVL bottom lock range, transiting from
the surface where GPS is available to the seafloor presents a
localisation problem since both GPS and DVL are unavailable
in the mid-water column. This is traditionally addressed by
using acoustic positioning systems, which take extra time to
deploy or require a tracking vessel. Such systems increase the
costs of operating in deep waters and reduce the flexibility of
AUV operations. This paper proposes an alternative approach
to navigation in the mid-water column that exploits the stability
of current profiles of water columns over short periods of
time. Observation of these currents are possible with the ADCP
(Acoustic Doppler Current Profiler) mode of the DVL. Results
with real data from missions with the Sirius AUV show how
the full integration of water column descent with the ADCP,
seafloor view-based SLAM (Simultaneous Localisation And
Mapping), and ascent to the sea surface with ADCP gives results
similar to having continuous bottom lock and shows potential
to act as an alternative to acoustic localisation.
I. I NTRODUCTION
Autonomous Underwater Vehicles (AUVs) have found application in general underwater exploration and monitoring.
This includes high-resolution, georeferenced optical/acoustic
oceanfloor mapping, along with providing measurements of
water column properties including currents, temperature and
salinity. An advantage of AUVs over other methods of ocean
observation is the autonomy and decoupling from the sea
surface that a self-contained vehicle provides.
Georeferencing, or the correct positioning in a global reference frame, is important for AUVs for the purposes of path
planning for mission requirements, registration with independently navigated information, or revisiting a previous mission. For shallow waters, low-frequency (150kHz) Doppler
Velocity Logs (DVL) can be in continuous use for depths
less than 200m. The DVL sensor provides measurements
of the seafloor relative velocity of the AUV. By combining
this information with an appropriate heading reference, the
observations can be placed in the global reference frame
and integrated to facilitate underwater dead reckoning. The
result is accuracies of 22m per hour (2σ) in position error
growth attainable during diving and 8m per hour error growth
(2σ) is possible if coupled with a navigation-grade (>$100K)
Inertial Measurement Unit (IMU) [1]. IMUs provide bodyrelative accelerations and rotation rates to constrain the
978-1-61284-380-3/11/$26.00 ©2011 IEEE
position, velocity and attitude estimates. Additionally, Simultaneous Localisation and Mapping (SLAM) methods
allow further navigation improvements [2] by identifying
previously explored regions.
In cases where the seafloor depth is greater than the DVL
bottom lock range, transiting from the surface, where GPS
is available, to the seafloor presents a localisation problem,
since both GPS and DVL are unavailable in the mid-water
column. Traditional solutions during this descent include
range-limited Long Base Line (LBL) acoustic networks
which require additional ship time to deploy and survey,
along with methods requiring tending, such as ship-based
Ultra Short Base Line (USBL). Acoustic positioning may
also suffer from multipath returns and the sound speed profile
through the water column needs to be accurately known.
These acoustic methods typically give O(10m) accuracy at
1-kilometre ranges [3] [4].
On-board sensors to aid localisation in the mid-water
column, which do not rely on external acoustics such as LBL
or USBL, include IMUs and DVL in water track mode. The
DVL in water track mode provides a measurement of the
velocity of the AUV relative to a user-programmable water
sampling volume that extends away from the instrument. A
navigation-grade IMU coupled with a vehicle model and
DVL water track mode estimates of currents can achieve
120m per hour (2σ) position error growth [5].
For a typical descent rate of 20 metres per minute, a
1 kilometre descent will take 50 minutes. Thus, without
USBL and LBL, relying on a state-of-the-art IMU/Vehicle
model/DVL water track navigation solution will give 100m
(2σ) accuracy. Thus, missions requiring revisit capabilities
in depths beyond a few hundred metres must invest in the
greater effort of external acoustic methods. LBL is the typical
mode of operation for vehicles such as the Autonomous Benthic Explorer (ABE) AUV when searching for hydrothermal
vents [6]. This requirement for using USBL or establishing
LBL infrastructure motivates the need for a higher accuracy
solution using only onboard instruments to provide AUVs
with the potential for true autonomy.
This paper explores an alternative approach to navigation
in the mid-water column that exploits the fact that current
profiles of water columns are stable over short periods of
time (in the minutes scale). With reobservation of these
currents with the ADCP (Acoustic Doppler Current Profiler)
3048
Error
magnitude No sensing
(2 sigma) after GPS
50 m
High end IMU only,
after GPS
A. ADCP sensor operation
High end IMU + Vehicle
model + Water track, after
GPS
40 m
30 m
ADCP aiding + Low
cost IMU after BL
20 m
USBL + GPS (on ship)
10 m
8m
LBL (10 kHz)
0
Depth
900m 1km
100m
Fig. 1. Comparison between different underwater localisation methods for a 1-km descent over 50 minutes where GPS is unavailable.
No sensing at all results in errors due to the unestimated water
currents. An IMU only results in a cubic rate of position error
increase with time. The red line shows the performance of the
proposed algorithm with a low-cost IMU. [7].
mode of the DVL during descent, along with sensor fusion
of other low-cost sensors, position error growth can be
constrained to near the initial velocity uncertainty of the
vehicle at the sea surface during the dive. Following DVL
bottom lock, the entire velocity history is constrained to an
error similar to the DVL velocity uncertainty, and coupled
with a low-cost IMU ($16K), 12m (2σ) per hour position
error growth is possible [8]. This result is compared to other
methods in Figure 1.
The experimental data presented here includes a descent
(with ADCP) following GPS on the sea surface, and incorporates the sea floor portion of the dive with DVL and viewbased SLAM [9]. View-based SLAM utilises the underwater
imagery of the seafloor that is collected to determine whether
the vehicle is revisiting a previously explored region, and
incorporates this information as a loop closure in SLAM.
Additionally, an ascent from the sea floor to the sea surface
(with ADCP) is completed, acquiring GPS positioning once
more, resulting in a fully georeferenced survey mission
without the need for external acoustic localisation.
The remainder of this paper is organised as follows.
Section II introduces the ADCP sensor and its operation,
along with previous related work, and an explanation of the
proposed method and theory. Section III presents results to
validate the proposed method. Section IV concludes with
discussion of future direction and the implications of this
work.
II. ADCP AIDING
An alternative to using the bulk water volume relative
velocity from the DVL water track mode [5] mentioned in
Section I, is to instead use the Acoustic Doppler Current
Profiler (ADCP) mode to provide finer depth resolution
current estimation. This opens the possibility of improved
vehicle motion estimates, given the observation of fine-scale
current structure.
The ADCP operates by sending out an acoustic pulse,
and relying on scatterers, such as plankton, to reflect back
the pulse. Using the Doppler effect, the velocity of the
scatterers relative to the instrument can be determined in the
along-beam direction. Since it is assumed that the scatterers
move with the water currents, the ADCP is measuring the
along-beam velocity of the water column relative to the
sensor. By gating the returned signal with time, currents at
different ranges from the ADCP sensor can be measured,
segmenting the observation into measurement cells. By using 4 differently aligned sensors and assuming horizontally
homogeneous currents, the 3D velocity of the current can
be determined. The fourth sensor provides redundancy in
the estimation of the current profile velocities [10]. Echo
intensity can also be used to check if there are anomalies
corrupting the returns, such as schools of fish [11]. The
result is a current estimate with accuracies according to
specification of O(10mm/s) (2σ) observing 2 m/s currents1.
B. Related work
Previous work regarding the use of ADCP sensors falls
within either the oceanographic community or the underwater localisation field.
The focus within the oceanographic community is on
estimating water current profiles, such as applying least
squares methods to fuse lowered ADCP and DVL bottom
lock information [12]. Accounting for ADCP sensor biases
[10] and sensor uncertainties is not tackled, as the effect
on the overall current profile is minimal, although there are
implications in the velocity estimates of the ADCP sensor
itself during the descent or ascent, which is important for
localisation as the velocity uncertainty relates to the position
error growth rate.
In the underwater localisation field, currents are typically
treated as a single time-varying parameter [13] using observations from the DVL water track mode. This approach does
not take advantage of the finer structure of the currents nor
exploit their stability over time.
In previous work [8], we report a method of incorporating
ADCP measurements to localise within an estimated water
column current profile. It includes a simulation for 1000
seconds of descent, and real data from an ABE AUV mission
searching for hydrothermal vents, but full integration with an
AUV mission is not illustrated due to errors with the data
due to the configuration of the sensor.
Recent developments in vector field SLAM [14] show
parallels with this work, which show that this problem
can be categorised into a wider set of problems which try
to simultaneously estimate the state of a vector field and
localise. An interesting difference between prior work in this
area and the estimation of the water current vector field is
that the vehicle is perturbed by the currents around it and
will tend to drift with the currents unless it works against
1 Sourced from Email from RD Instruments providing standard deviation
of the instrument
3049
them. This has important implications for path planning.
Horizontally looking ADCPs have been employed to map
current fields in [15] for path planning, but have not been
used to aid with localisation. In [16], an autonomous aerial
glider estimates a wind vector field for the purpose of
path planning as it navigates through it, but simultaneous
localisation is not attempted.
C. ADCP estimation and navigation aiding method
The following is an outline of the method and its operation, as originally presented in [8]. We assume that initially,
the AUV has position and velocity estimates in the navigation
frame at the sea surface from GPS. With the ADCP sensor,
body-relative water current bin velocities below the vehicle
are observed with each ADCP measurement cell. These
observations can be used to estimate the full current profiles
in the navigation frame using the estimated vehicle velocity
at the surface. This is illustrated in Figure 2(a).
After another ADCP measurement is made, the vehicle
reobserves the same current bins. Given the estimated water
current velocity of the reobserved current bin and the bodyrelative velocity of these current bins from the ADCP, a filter
can be used to simultaneously update the estimate of the
vehicle velocity and current profile velocities. This relies
on the assumption that the water current velocity in this
bin remains constant, which is realistic over a reobservation
period of minutes. This is shown in Figure 2(b).
New current bins can also now be estimated as the vehicle
changes depth as shown in Figure 2(c). The result is an
estimate of the vehicle motion and the water column current
profile. When the vehicle is within DVL range of the seafloor,
seafloor-relative observations of velocity provided by the
DVL, with accuracy of 6mm/s (2σ) for a 1500 kHz DVL
are also incorporated into the filter. This can serve to update
the estimate of both the vehicle state and the current profile
due to the correlations that exist between these estimates..
1) Information filter with current profiling: Vehicle pose
states (position, velocity and attitude), ADCP bias states and
water current velocity states are stored in a state vector in
the form

 +
x̂p (tk )

(1)
x̂+ (tk ) =  x̂+
bc (tk )
+
x̂vc (tk )
T
+T
x̂+T
is a vector of
where x̂+
p1 (tk ), ..., x̂pnp (tk )
p (tk ) =
past and present pose states where np is the number of
T
+T
+T
(t
)
(t
),
...,
x̂
x̂
(t
)
=
vehicle pose states, x̂+
k
k
k
bc,nb
bc,1
bc
is a vector of past and present ADCP bias states where
nb is the number of ADCP bias states and x̂+
vc (tk ) =
T
+T
+T
x̂vc,1 (tk ), ..., x̂vc,nv (tk )
is a vector of past and present
ADCP water current velocity states where nv is the number
of water current velocity states. The covariance between the
pose states and the water current states are in the form


P̂+
P̂+
P̂+
pvc (tk )
pp (tk )
pbc (tk )


+
+
P̂+ (tk ) =  P̂+T
(2)
pbc (tk ) P̂bc bc (tk ) P̂bc vc (tk ) 
+T
+
(t
)
(t
)
P̂
(t
)
P̂
P̂+T
vc vc k
pvc k
bc bc k
A sparse extended information filter (SEIF) estimates the
states of the vehicle given the various vehicle sensors [17].
The utilisation of the SEIF allows natural integration with
view-based SLAM [9] which occurs on the seafloor, as it
shares the same filter framework. Previous work in vector
field SLAM [18] implements a SEIF to maintain the state
space due to having computational performance benefits to
keep the filter running in constant time, while accomodating
loop closure information in the vector field. This simply
accommodates the water current states and allows relinearisation or smoothing if required. The SEIF maintains the
correlations between states.
In the information form, the filter maintains the matrix Y,
which is the inverse of the covariance matrix
Ŷ+ (tk ) = [P̂+ (tk )]−1
(3)
and the information vector y, which is related to the state
estimate by
ŷ+ (tk ) = Ŷ+ (tk )x̂+ (tk )
(4)
Observations, which include ADCP measurements, are assumed to be made according to
z(tk ) = h[x(tk )] + ν(tk )
(5)
in which z(tk ) is an observation vector, h[x(tk )] is the
sensor model relating states to observations, and ν(tk ) is
a vector of observation errors with covariance R(tk ). New
information from sensor measurements are incorporated into
the information vector and matrix
ŷ+ (tk ) = ŷ− (tk ) + i(tk )
(6)
Ŷ+ (tk ) = Ŷ− (tk ) + I(tk )
(7)
in which
i(tk ) =
−1
∇T
(tk )(z(tk ) . . .
x h(tk )R
−
−h[x̂ (tk )] + ∇x h(tk )x̂− (tk ))
−1
I(tk ) = ∇T
(tk )∇x h(tk )
x h(tk )R
(8)
(9)
where x̂− (tk ) is the a priori state estimate and ∇x h(tk )
is the Jacobian of the observation with respect to the state.
Poses and current states which are no longer observed in the
filter can be marginalised out to reduce the computational
requirements of the filter, although this results in those past
states being lost in the estimation. By maintaining the entire
state history, the filter also can be relinearised to correct for
linearisation errors. Cholesky modifications as outlined in [9]
can be used to keep the filter running in amortised constant
time.
2) ADCP observation: The observation function for each
ADCP measurement is
Wj vnc,j ) + bc,i + νadcp
(10)
zadcp,i = Cbn (−vnv +
where:
th
• zadcp,i is the ADCP-measured current vector in the i
measurement cell
3050
(a)
(b)
(c)
Fig. 2. ADCP aiding method sequence (a) Initial GPS position and velocity are known, and water velocities can be deduced. (b) The
AUV moves, and reobserves the same current bins.(c) The AUV velocity in the world frame can be deduced, along with new current bins
shown in red.
Cbn is the Coordinate transform from navigation/world
frame to ADCP/body frame
n
• vv is the Vehicle velocity in the world/navigation frame
• Wj is the Weighting function for each water current
velocity from current bin j
n
• vc,j is the water current velocity from current bin j. Each
depth cell contains a current velocity state in the X and
Y direction, which represents the average velocity of
the current through that layer
th
• bc,i is the Bias in the i
measurement cell in the body
frame
• νadcp is the Random noise in the ADCP measurement
whose magnitude is given by the sensor manufacturer
Sources of ADCP biases [10] [19] include measurement
cell-dependent biases, due to beam and sensor misalignment,
beam geometry and signal/noise ratio. Sources of ADCP
biases due to changing depth include temperature, pressure,
scatterers and sound speed estimate error.
Further details regarding ADCP sensor modelling can be
found in [8].
Visual Loop Closures: Loop-closure observations are created using a six degree-of freedom stereovision relative
pose estimation algorithm [9] [20]. The SIFT algorithm is
used to extract and associate visual features, and epipolar
geometry is used to reject inconsistent feature observations
within each stereo image pair. Triangulation is performed to
calculate initial estimates of the feature positions relative to
the stereo rig, and a redescending M-estimator, is used to
calculate a relative pose hypothesis that minimises a robustified registration error cost function. Any remaining outliers
with observations inconsistent with the motion hypothesis
are then rejected. Finally, the maximum likelihood relative
vehicle pose estimate and covariance are calculated from the
remaining inlier features.
•
III. R ESULTS
In previous work, simulation results illustrated the performance of the proposed algorithm using sensor fusion of a
low cost IMU and DVL bottom lock at the sea floor [8].
It was shown that navigation performance of 12m per hour
(2 σ) position error growth [8] during descent was possible
following DVL bottom lock, which is equivalent to having
DVL bottom lock the entire ascent and descent. For these
simulations, the ADCP operated at its maximum specified
update rate (5 Hz). This accuracy is competitive with acoustic
localisation methods with USBL and LBL.
Fig. 3.
The Sirius AUV imaging cuttlefish populations.
The following results are obtained through the use of
Sirius [2], the University of Sydney’s Australian Centre for
Field Robotics (ACFR) oceangoing AUV. It is a modified
version of the mid-sized SEABED robotic vehicle from
Woods Hole Oceanographic Institution [21]. This class of
AUV is designed for relatively low-velocity, high-resolution
imaging and is passively stable in roll and pitch. The Sirius
AUV is pictured in Figure 3 performing a mission. It is
equipped with a suite of oceanographic sensors. Navigation
sensors onboard include a 1500 kHz RDI DVL/ADCP,
Tracklink USBL and a Lassen iQ GPS receiver, along with a
stereo imaging platform, which allows for view-based loop
closures. The 1500 kHz RDI DVL/ADCP uses less power
and is more accurate than the 150 kHz version, and can
be used at a lower minimum range, which is important for
localisation during image acquisition which occurs at 2m
3051
Run−time position estimate with partial DVL + ADCP
20
2 σ position estimate uncertainty for
partial DVL+ADCP in−run
18
2 σ position estimate uncertainty for
full DVL+ADCP in−run
16
2 σ position estimate uncertainty for
partial DVL+ADCP for full state history
14
2 σ position estimate uncertainty for
full DVL+ADCP for full state history
Full state history with partial DVL + ADCP
−20
Full state history with full DVL + ADCP
USBL measurements
20
Descent
Ascent
Position uncertainty (m)
Depth (m)
0
40
60
−800
−900
12
10
8
6
GPS
availability
4
GPS
availabilitiy
−1000
Correction when DVL bottom lock is acquired
Local Northing (m)−1100
840
820
800
780
760
Seafloor surveying + view−based SLAM
2
740
Descent
+ DVL
0
0 blackout
Local Easting (m)
500
1000
Time since mission start (s)
Ascent
+ DVL
blackout
1500
(a)
Trajectory of the 45m depth short mission, where DVL
bottom lock is available the entire dive. A simulated partial DVL
blackout for 150 seconds during ascent and descent results in only
ADCP measurements for this time.
Fig. 4.
Difference between in−run filter estimate
and USBL
35
2 σ uncertainty of the difference between
the in−run filter estimate and USBL
30
Difference between full state history filter
estimate and USBL
2 σ uncertainty of the difference between
the full state history filter estimate and USBL
25
Position difference (m)
altitude. This comes at the cost of maximum range for bottom
lock, which is 45m for the 1500 kHz compared to 200m for
the 150 kHz.
Data from two missions recently completed by the vehicle
in Tasmania are used to illustrate the performance of the
proposed ADCP-aided navigation filter. The first is a shorter
mission in relatively shallow water such that DVL lock
was available throughout the dive. The second dive was
completed in deeper water and relies on USBL observations
to validate the positioning accuracy.
20
15
10
5
0
0
500
1000
Time since mission start (s)
1500
A. Shallow water dive
The Sirius vehicle was used in a mission which involved
descending in water which was just within DVL bottom
lock range (about 45m depth for the 1500 kHz DVL),
completing some subsurface manoeuvres with visual loop
closures for SLAM, and then ascending. The total dive time
was approximately 1000 seconds. This mission’s position
estimates are shown in Figure 4. In order to show how the
ADCP-aided method compares with the ground truth from
DVL, the DVL measurements were not fused into the filter
during the descent and ascent phase for 150 seconds each.
This simulates a greater depth where DVL bottom lock would
not be possible, while providing a comparison with ground
truth. Comparisons can illustrate how the ADCP-only case
compares to the DVL while descending and ascending. The
ADCP in this case operated at 1 Hz and the DVL operated
at 1 Hz in an alternating configuration at all times.
As shown in Figure 5(a), during descent for the online runtime filter, the error in position quickly grows as the velocity
uncertainty is quite high given GPS, at about 10cm/s (2σ).
Once DVL bottom lock is acquired, the position uncertainty
during the descent is reduced. The reason for this is that
once DVL bottom lock is acquired after descent, the velocity
estimates of the water currents in the entire water column
(b)
Fig. 5. (a) Shallow water dive position uncertainty estimates for
filters with partial and full DVL, in-run and for the full state history
(b) Differences between filter results and USBL measurements
for the shallow water dive, along with 2 σ uncertainties of this
difference, showing that the filter is consistent
are improved. By maintaining correlations between states
during the descent, the filter propagates the accurate velocity
information attained upon reacquiring bottom lock back
through the entire state history.
During run-time, the position uncertainty of the mission is
12m (2σ) just prior to post-ascent GPS acquisition, and after
is within 6m (2σ) for the full state history. This compares
with the error estimate of at most 5m (2σ) for the full state
history when using DVL the entire time. So even with only 1
Hz ADCP measurements for 150 seconds during the ascent
and descent, the uncertainty associated with the estimates of
the entire mission approaches the full DVL localisation case.
This correction is accurate to almost the DVL velocity
accuracy. It does not have the same accuracy to the equivalent
DVL during this time, because only a finite number of
3052
B. Deeper water dive
5
North current estimate
on descent
10
East current estimate
on descent
15
North current estimate
on ascent
Run−time position estimate with DVL + ADCP
East current estimate
on ascent
Full state history with full DVL + ADCP + USBL
Full state history with DVL + ADCP
Descent
0
25
20
Depth (m)
Depth (m)
20
30
35
Correction when DVL bottom lock is acquired
40
60
Survey grid with loop closures
Ascent
80
100
400
40
500
45
600
−0.05
0
0.05
0.1
0.15
0.2
Current velocity (m/s)
0.25
0.3
0.35
−750
700
−700
−650
800
−600
900
−550
−500
1000
Fig. 6. Current estimates for 45m-depth short mission, with ascent
Local Easting (m)
1100
−450
−400
Local Northing (m)
and descent occuring approximately 1000 seconds apart.
(a)
Run−time trajectory with
DVL + ADCP
1000
Full state history with
DVL + ADCP
950
Full state history with
DVL + ADCP + USBL
900
USBL observations
850
Local Easting (m)
measurements with the noisy ADCP sensor is used to observe
the water column currents. Thus, there is a slight information
loss, which can be reduced by increasing the frequency with
which ADCP measurements are taken, which remains as
future work.
The action of a view-based SLAM loop closure is seen
at about 800 seconds of mission time, evident from the
sudden decrease in uncertainty for the filter while in-run,
as the AUV has detected a revisit to a previous site in
the mission through image feature matching. This acts to
limit the position uncertainty during the seafloor mapping
portion of the dive as the filter runs online, along with
subsequent loop closure observations. Additionally, a greater
improvement in the localisation solution following the postascent GPS acquisition is possible, as there exist stronger
correlations throughout the dive to link the prior-descent GPS
positioning to the post-ascent GPS positioning.
The descent rate was 0.29 m/s, while the ascent rate was
0.15 m/s. As a result, the descent phase of the mission had
less information stored within each current layer, resulting
in a slightly higher error in positioning after the descent,
as opposed to prior to the ascent, after the post-ascent GPS
acquisition. This is evident by the larger position uncertainty
disparity between the partial and full DVL filters for the
full state history for the descent compared to the ascent in
Figure 5(a). Quantifying the trade-off in navigation accuracy
involved with the selection of the descent rate remains as
future work.
Figure 5(b) compares the filter result with the USBL observations available as independant information for comparison.
It can be seen that the ADCP filter is consistent with the
USBL observations, validating the results.
Figure 6 shows the final estimated current for the mission,
illustrating the water profile current structure which the
ADCP-aided method measures in order to navigate against.
800
750
700
650
600
550
500
−900
−800
−700
−600
−500
Local Northing (m)
−400
−300
(b)
Fig. 7. (a) Oblique and (b) Bird’s eye view of the trajectory for
100m depth deeper water mission, where DVL bottom lock is only
available at 45m altitude.
The vehicle also completed a longer dive in 100m of
water in which DVL bottom lock was not available through
the descent and ascent. The entire mission time is over 3
hours. The mission’s position estimates are show in Figure
7. Ground truth in this case is more reliant on the USBL,
which exhibit difficult-to-model errors due to multipath and
the acoustic pulse travelling through an unmodelled water
column.
Results in Figure 8(a) show how the ADCP method, without the USBL, results in georeferencing for the subsequent
seafloor view-based SLAM aided mission. Georeferencing
uncertainty is within 20m (2σ) position accuracy while the
mission is underway on the seafloor, and after post-ascent
GPS acquisition, the accuracy of the seafloor portion is
within 11m (2σ). The ADCP operated at 0.5 Hz and the DVL
operated at 2 Hz in an alternating configuration where one
3053
ADCP measurement is made every five DVL measurements.
Even with such a low rate of ADCP measurements, it is
possible to localise without an external acoustic source such
as USBL, although the localisation uncertainty is much
higher than the equivalent of having DVL bottom lock the
entire time due to information loss.
10
20
30
Depth (m)
40
40
2 σ position estimate uncertainty from DVL+ADCP for full state history
East current estimate
on descent
2 σ position estimate uncertainty from DVL+ADCP in−run
35
2 σ position estimate uncertainty from DVL+ADCP in−run without
loop closure
30
North current estimate
on ascent
60
2 σ position estimate uncertainty from DVL+ADCP for full state history
without loop closure
Position uncertainty (m)
North current estimate
on descent
50
East current estimate
on ascent
70
25
80
20
90
15
100
−0.2
−0.15
−0.1
−0.05
0
Current velocity (m/s)
0.05
0.1
10
5
0
Fig. 9. Current estimates for 100m deep water dive, with ascent
and descent occuring approximately 3 hours apart.
0
2000
4000
6000
8000
Time since mission start (s)
10000
12000
(a)
80
Difference between in−run filter estimate and USBL
70
2 σ uncertainty of the difference between the in−run
filter estimate and USBL
Difference between full state history filter estimate and USBL
2 σ uncertainty of the difference between the full state history
filter estimate and USBL
Position difference (m)
60
50
40
30
20
10
0
0
2000
4000
6000
8000
Time since mission start (s)
10000
12000
(b)
Fig. 8. (a) Deeper water dive position uncertainty estimates for filter
in-run and for the full state history, with and without loop closures
(b) Differences between filter results and USBL measurements
for the deeper water dive, along with 2 σ uncertainties of this
difference, showing that the filter is consistent after DVL bottom
lock
The action of a view-based SLAM loop closure is seen
at about 3500 seconds of mission time in Figure 8(a), and
subsequent loop closures limit the uncertainty in position
for the mission. The advantages of this coupled with the
ADCP-aided descent and ascent are the same as in Section
III-A for the shallow dive mission, even with a 3 hour long
seafloor portion of the dive. A further advantage of viewbased SLAM is a significantly improved localisation for the
seafloor portion of the mission after the post-ascent GPS
acquisition due to the increased correlation of temporally
distant poses, as seen in Figure 8(a).
Figure 8(b) compares the filter result with the USBL
observations available as independant information for comparison. Prior to DVL bottom lock during the descent using
the in-run filter estimate, there is accumulating linearisation
error due to the inaccurate velocity estimates and the nonlinear rotation in the ADCP sensor model, and the filter
becomes inconsistent. Once DVL bottom lock is acquired,
relinearisation can occur with the SEIF and the linearisation
error is reduced. Subsequently, it can be seen that the ADCP
filter is consistent with the USBL observations, validating
the results.
The resultant water column current profile is shown in
Figure 9. It can be seen that a small but noticeable change
has occured over the 3 hours between the ascent and descent,
along with the changing horizontal water currents observed
with depth.
C. Implications of results
The above results show how ADCP-aided navigation during the descent and ascent of a mission, coupled with viewbased SLAM on the seafloor, allows georeferencing even
with infrequent ADCP measurements.
In the case of untended long-term monitoring and exploration AUVs or underwater gliders, tighter constraints on
power consumption are imposed. This requires operating
sensors sparingly. As shown by the results in this paper,
even infrequent ADCP measurements provide information
which permits localisation. This represents a viable solution
for accurate autonomous underwater navigation, which at
the moment does not exist. Additionally, an accurate water
current vector field estimate is output, which is also a useful
a data product.
3054
IV. C ONCLUSION
This paper has demonstrated the ability to achieve constrained error growth in position by incorporating ADCP
measurements into the navigation solution while a vehicle
is transitting between the sea surface and the seafloor. This
makes it appropriate for long-term, accurate navigation of an
AUV which dives and resurfaces, and requires underwater
position accuracy close to the seafloor, without DVL bottom
lock the entire mission. This alleviates the requirement for
a tending vessel or setup of an acoustic network to achieve
precise navigation.
Work is extended from [8] with an analysis of performance
with infrequent ADCP observations in the context of real
missions, and showing how integration with view-based
SLAM can be achieved naturally in this framework.
Accounting for prior information about water currents
will also aid the localisation. Correlating the descent phase
with the ascent phase water currents should be possible by
accounting for tidal changes and other factors using spatiotemporal water current evolution models. Accommodating
horizontally looking ADCPs, or using the slanted beam
characteristics of the downward-looking ADCP, along with
more generalised methods to discretise the state space (into
3D current voxels) will further generalise this method to
other situations beyond vertical dives.
ACKNOWLEDGMENT
This work is supported by the ARC Centre of Excellence
programme, funded by the Australian Research Council
(ARC) and the New South Wales State Government, the Integrated Marine Observing System (IMOS) through the Department of Innovation, Industry, Science and Research (DIISR) National Collaborative Research Infrastructure Scheme
and the Commonwealth Environment Research Facilities
(CERF) biodiversity hub who supported field work to collect
the data used to validate the methods proposed in this work.
The authors would like to thank the Captain and crew of
the RV Challenger. Without their sustained efforts none of
this would have been possible. We also acknowledge the
help of all those who have contributed to the development
and operation of the AUV, including Ian Mahon, Matthew
Johnson-Roberson, the late Duncan Mercer, George Powell,
Donald Dansereau, Lachlan Toohey, Ritesh Lal, Paul Rigby,
Jeremy Randle, Bruce Crundwell and the late Alan Trinder.
[5] O. Hegrenaes and E. Berglund, “Doppler water-track aided inertial
navigation for autonomous underwater vehicle,” in IEEE OCEANS
2009 - EUROPE, May 2009.
[6] M. Jakuba, “Stochastic mapping for chemical plume source localization with application to autonomous hydrothermal vent discovery,”
Ph.D. dissertation, Massachusetts Institute of Technology and Woods
Hole Oceanographic Institution, 2007.
[7] D. Titterton and J. Weston, Strapdown inertial navigation technology.
Peter Peregrinus Ltd, 2004.
[8] L. Medagoda, M. Jakuba, O. Pizarro, and S. Williams, “Water Column Current Profile Aided Localisation for Autonomous Underwater
Vehicles,” in IEEE OCEANS 2010, May 2010.
[9] I. Mahon, S. Williams, O. Pizarro, and M. Johnson-Roberson, “Efficient view-based slam using visual loop closures,” Robotics, IEEE
Transactions on, vol. 24, no. 5, pp. 1002 –1014, oct. 2008.
[10] R. Gordon and R. Instruments, “Principles of Operation a Practical
Primer,” RD Instruments, San Diego, 1996.
[11] P. Plimpton, H. Freitag, and M. McPhaden, “ADCP Velocity Errors
from Pelagic Fish Schooling around Equatorial Moorings*,” Journal of
Atmospheric and Oceanic Technology, vol. 14, no. 5, pp. 1212–1223,
1997.
[12] M. Visbeck, “Deep velocity profiling using lowered acoustic Doppler
current profilers: Bottom track and inverse solutions,” Journal of
Atmospheric and Oceanic Technology, vol. 19, no. 5, pp. 794–807,
2002.
[13] R. McEwen, H. Thomas, D. Weber, and F. Psota, “Performance of an
AUV navigation system at Arctic latitudes,” IEEE Journal of Oceanic
Engineering, vol. 30, no. 2, pp. 443–454, 2005.
[14] J. Gutmann, G. Brisson, E. Eade, P. Fong, and M. Munich, “Vector
field SLAM,” in Robotics and Automation, 2010. ICRA 2010. Proceedings 2010 IEEE International Conference on. IEEE, 2010, pp.
236–242.
[15] B. Garau, A. Alvarez, and G. Oliver, “AUV navigation through
turbulent ocean environments supported by onboard H-ADCP,” in
Robotics and Automation, 2006. ICRA 2006. Proceedings 2006 IEEE
International Conference on, 2006, pp. 3556–3561.
[16] N. R. Lawrance and S. Sukkarieh, “Simultaneous exploration and
exploitation of a wind field for a small gliding uav,” in AIAA Guidance,
Navigation and Control Conference, Toronto, Ontario, Canada, 2010,
aIAA Paper 2010-8032.
[17] S. Thrun, Y. Liu, D. Koller, A. Ng, Z. Ghahramani, and H. DurrantWhyte, “Simultaneous localization and mapping with sparse extended
information filters,” The International Journal of Robotics Research,
vol. 23, no. 7-8, p. 693, 2004.
[18] J.-S. Gutmann, E. Eade, P. Fong, and M. Munich, “A constant-time
algorithm for vector field slam using an exactly sparse extended
information filter,” in Proceedings of Robotics: Science and Systems,
Zaragoza, Spain, June 2010.
[19] C. Atkinson, “Analysis of shipboard ADCP data from RRS Discovery
Cruise D324: RAPID Array Eastern Boundary,” 2008.
[20] S. Williams, O. Pizarro, J. Webster, R. Beaman, I. Mahon, M. JohnsonRoberson, and T. Bridge, “Autonomous underwater vehicle–assisted
surveying of drowned reefs on the shelf edge of the Great Barrier
Reef, Australia,” Journal of Field Robotics, vol. 27, no. 5, pp. 675–
697, 2010.
[21] H. Singh, A. Can, R. Eustice, S. Lerner, N. McPhee, O. Pizarro,
and C. Roman, “Seabed AUV offers new platform for high-resolution
imaging,” EOS, Transactions of the AGU, vol. 85, no. 31, pp. 289–294,
2004.
R EFERENCES
[1] F. Napolitano, A. Chapelon, A. Urgell, and Y. Paturel, “PHINS: The
autonomous navigation solution,” Sea Technology, vol. 45, no. 2, pp.
55–58, 2004.
[2] S. Williams, O. Pizarro, I. Mahon, and M. Johnson-Roberson, “Simultaneous Localisation and Mapping and Dense Stereoscopic Seafloor
Reconstruction Using an AUV,” in Experimental Robotics. Springer,
2009, pp. 407–416.
[3] J. Kinsey, R. Eustice, and L. Whitcomb, “A survey of underwater vehicle navigation: Recent advances and new challenges,” in Proceedings
of the 7th Conference on Maneuvering and Control of Marine Craft
(MCMC2006). IFAC, Lisbon. Citeseer, 2006.
[4] M. Mandt, K. Gade, and B. Jalving, “Integrating DGPS-USBL position
measurements with inertial navigation in the HUGIN 3000 AUV,” in
Proceedings of the 8th Saint Petersburg International Conference on
Integrated Navigation Systems, Saint Petersburg, Russia, 2001.
3055