IEEE Robotics & Automation Magazine - September 2015 - 81

Setup 3 Diver and PlaDyPos [Figure 9(d)]
The final experimental setup, which is in fact the demonstrator
of the final goal of the described robotic system, includes experiments in real conditions with all the abovementioned uncertainties included.
To validate the results, a visual confirmation that gives a
ground truth of the tracking performance is made. A downlooking camera was mounted on the PlaDyPos to validate the
tracking results for setups 2 and 3 (where real agents are being
tracked). Position of the agent is determined within the
image, and its distance to the center of the image is calculated
in pixels. Based on the known size of the ROV and the diver,
the measure in pixels is transferred to meters, giving ground
truth of the tracking performance. Influence of the roll and
pitch of PlaDyPos is compensated using the measurements
from the inertial sensor. An example of the images obtained
from the two setups is shown in Figure 10.
The accuracy of the vision-based ground truth can be
compromised if the camera orientation is not perfectly
aligned with the gravity vector. The upper limit of the error
in the observed position of the agent Dx can be estimated by
using a simplified model Dx = z $ tan a, where z is the
depth of the agent and a is the camera orientation with respect to the gravity vector. If the misalignment is not larger
that a = 5c, the estimate of upper limit of the error is less
than 10% of the depth of the agent.
Experimental Results
A large number of experiments, with previously described experimental setups, were conducted in June 2014 in Split,
Croatia, at the Croatian Navy base. Position plots of all obtained results with a VD, ROV, and a human diver during
transect following in three experimental setups are shown in
Figure 11. Even in these position plots, it can be seen that the
variance of measurements depends on the experimental
setup, from low variance in experiments with the VD to high
variance in experiments with the human diver due to a large
number of sources of uncertainty.
To get a clearer picture of the influence of different sources
of uncertainty on the tracking error, results from each experimental setup are analyzed. Simulation results from setup 0 are
omitted from this article to keep the focus on results obtained

North (m)

80
70
60
0

100

200

300
Time (s)

400

500

600

500

600

500

600

East Time Plot
East (m)

Setup 2 ROV and PlaDyPos [Figure 9(c)]
In this real-environment setup, the human diver is replaced
with an ROV with an acoustic modem pinging the USBL on
the PlaDyPos, and thus introducing the real acoustic channel
uncertainties but eliminating those caused by the diver. This
setup is designed to identify potential deterioration in system
performance due to the acoustic channel characteristics.

North Time Plot

-8
−8.5

VD Measured
VD Estimated
ASV PlaDyPos Estimated
100
200
300
400
Time (s)
(a)
Tracking Error

-9
-9.5

Simulated USBL Measurement
VD Estimate
Tracking Error (m)

with (7). While this experimental setup eliminates acoustic
sensor and diver related uncertainties, it also allows reliable
testing of PlaDyPos behavior under different measurement
update rates and performance evaluation of the diver estimator onboard PlaDyPos in real environmental conditions.

1
0.4852
0

100

200

300
Time (s)
(b)

400

Figure 12. The experimental results obtained from setup 1 with
the VD: (a) the north and east positions and (b) the tracking
error. The largest tracking errors are due to abrupt changes in
the direction of the VD. The diver position is estimated from
the measurements with great precision, and a smooth signal
is ensured for the tracking system at frequency higher than the
measurement availability.

Table 2. Metric for the uncertainties: mean
tracking errors of agents (VD, ROV, human diver).
Mean Tracking Error in (m)
d ( h)

d (hu )

From Video

Setup 1 (h = h virtual diver)

0.4852

0.3906

Setup 2 (h = h ROV)

0.9994

0.4512

0.9169

Setup 3 (h = h diver)

1.7772

1.3510

1.4831

in real environmental conditions. The simulation results can
be found in [10].
Results for Setup 1: VD Tracking
The full experiment with the VD tracking in duration of about
10 min is shown in Figure 12. While the results given in Figure 12(a) indicate that PlaDyPos was following the same path
as the VD, the real VD tracking quality is observed from
September 2015

IEEE ROBOTICS & AUTOMATION MAGAZINE

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2015