IEEE Robotics & Automation Magazine - September 2015 - 83

North (m)

North Plot
60
40
20
0

Diver Measured
Diver Estimated
ASV PlaDyPos Estimated
50 100 150 200 250 300 350 400 450 500 550
Time (s)

East (m)

East Plot
20
0
-20

50 100 150 200 250 300 350 400 450 500 550
Time (s)
(a)
Tracking Error

Tracking Error (m)

USBL Measurement
Diver Estimate
Video Validation

2
1.3149
0.9615
0

50 100 150 200 250 300 350 400 450 500 550
Time (s)
(b)

Figure 15. The experimental results obtained from setup 3 with the
human diver: (a) the north and east positions and (b) the tracking
error. A long period of missing acoustic data is visible at around
300 s-the diver position estimator provides estimates based on a
simplified diver model during this period and during other, shorter,
periods when measurements are not available. The overlayed
ground-truth video validation shows that the diver estimator is
providing estimates in line with true diver position. Horizontal
dashed lines denote mean tracking error during the presented
segment of the experiment, while Table 2 contains values for the
complete experiment.

with the result from setup 2, allows us to conclude that the presence of the human diver contributes an additional 0.8 m resulting
in the total error of about 1.8 m. The tracking error calculated
based on the diver position estimates is considerably lower.
Observe that the presence of the human diver has little influence
on the median measured tracking error, as shown in Figure 13.
However, the increase of statistical outliers and the higher spread
are attributed to the diver's presence.
Ground-Truthing the Results
Since during the experiment both the diver and the ROV
were tracking the transect at a depth of about 5 m, it was possible to detect them in the video image. The accuracy of the
visual ground truth is estimated to about !0.5 m based on the
error estimation analysis (due to misalignment of the camera
with the gravity vector) provided before. It should be

mentioned that at larger depths this type of validation would
not be possible due to low visibility.
The tracking error based on video data is overlayed in
Figures 14(b) and 15(b), and the mean tracking error values
are listed in Table 2. These values are very close to the measured and the estimated mean tracking errors showing the
accuracy of the results obtained from measured and estimated positions. It should also be mentioned that the difference
between the setups 2 and 3 mean tracking error from video
data is around 0.6 m, which shows that the influence of the
human uncertainty determined by the acoustic measurements (0.8 m) and the estimated diver positions (0.9 m) are
sufficiently accurate.
The mean error from the diver position estimates is lower
due to inclusion of the diver estimator. However, it should be
mentioned that if this error is too conservative, the diver motion is not estimated properly. By comparing this estimation
with the video validation, we conclude that the diver estimator gives satisfactory results.
Open Data
The experiment described in this article was designed to allow
future replication for comparison with new positioning sensors and methods. All software was implemented within the
ROS (http://www.ros.org) framework which is used by the
worldwide robotics community. During the experiments, all
the relevant data were
logged in an ROS bag forThe diver state estimator
mat. Video validation
footage was time stamped
uses intermittent USBL
and logged into a separate
ROS bag file due to its
measurements, PlaDyPos
size. An a posteriori analysis of the experimental
states, and simplified
data was performed to
identify and extract parts
diver model.
where actual tracking has
taken place. Filtered bag
files were loaded into
MATLAB where the final analysis step was performed. The
data and MATLAB scripts used during analysis are made publicly available at https://bitbucket.org/labust/diver-tracking, together with clear instructions on how to use the data. Making
the data and scripts available in a Git repository makes future
changes and contributions easily trackable.
Conclusions
The benchmark scenario of following a georeferenced transect laid on the seabed allowed us to execute replicable experiments with an ASV for diver tracking. Given that
experiments with human divers introduce a large number of
uncertainties, a structured step-by-step experimental plan
was devised with the intention to identify the influence of
uncertainties introduced by the environment, the unmodeled dynamics, the acoustic sensors, and the human diver.
We have defined a metric in the form of a mean tracking
September 2015

IEEE ROBOTICS & AUTOMATION MAGAZINE

http://www.ros.org https://www.bitbucket.org/labust/diver-tracking

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