IEEE Robotics & Automation Magazine - September 2021 - 74

divers. With these values of a , w , and d , values for H are
computed based on the Jerlov classification.
Image/Video Compression
Underwater robots are employed in practice in a wide range
of applications for a wide range of different tasks, even
within a particular application. Hence, different bandwidth
values or CPU resources
are typically allocated to
each system component
depending on the mission
and the task within the
mission. It is, hence, a common
practice that compression
algorithms are
applied to the images (specifically,
video frames) during
real missions to free
resources for other processes
as well as for data
storage, e.g., for data transmission
in the case of re -
motely operated vehicles.
Often, motion JPEG is
used to optimize for coding
speed and frame-byq=
" 60 20, are used here.
,
Geometry-Contextual Perturbations
In underwater environments, ambient light attenuates exponentially
with depth d and even further with the distance z
between the target object (here, the diver) and the observer
(here, the camera on the AUV). However, the attenuation factor
()
Kd due to the depth can typically be ignored because
the attenuation factor (, )zbm due to distance z and wavelength
m is 2-5 times greater [33]. Note that we can typically
assume an observer-object viewing direction of approximately
i
90=
c (Figure 1).
To obtain z or the depth relative to the image, DispnetC [40]
is used. It is a 100% dense disparity estimator with %4. error in
the KITTI Stereo 2015 benchmark. This accuracy is more than
enough for our purposes, and the method has proven to perform
well in underwater scenarios [41]. Nonetheless, the estimated z
is refined through a bilateral filter to keep the image edge consistency
[Figure 6(g)]. Based on this value, the geometry-contextual
image transformations presented in the following
sections can be applied.
Underwater Haze
A popular haze model used in terrestrial robotics is based on
the following equation:
() () () (( )),
()=
Ix Jxtx Bt x=+ 1-bm
zx() ()
tx e
,
74 * IEEE ROBOTICS & AUTOMATION MAGAZINE * SEPTEMBER 2021
(1)
(2)
where ()
()
Ix is the image received by the camera sensor and
Jx is the original image (scene radiance), which is exponentially
attenuated by the transmission matrix ()
tx at every pixel
x as the range (distance to object) increases and depends on
the wavelength m . B is the ambient light. As mentioned previously,
()
Data were recorded in
the open sea as well as in
indoor and outdoor pools
at three different locations,
namely, in Biograd na Moru
(Croatia), at the Brodarski
Institute in Zagreb (Croatia),
and in Genova (Italy).
frame quality over bitrate. To study the image degradation effects,
compression quality values of
Jx and B ,
,
zx is computed here by DispnetC. It is refined with a
bilateral and a Gaussian filter to avoid discontinuity effects.
But a physically realistic haze model is more complex in the
underwater case [33]. In summary, a different transmission
matrix is needed for each ()
tx e
B ()=
() () () (( )),
J ()=
tx e
-b(, )( )
wd zx
.
Ix Jx tx Bt xJB1=+ -
-bm zx() ()
(3)
(4)
(5)
The underwater haze model from (3) is, hence, combined
with the range map ().zx This allows us to apply systematic
and controlled image degradations to real-world data in the
form of physically realistic underwater haze.
For the values of B and its corresponding (),txB
values as for the underwater alpha blend a 02 . 5=
(see the " Pixel-Based Perturbations " section). For (),txJ
attenuation coefficients ()
the same
are used
bm are chosen to allow for visibility
within a distance of approximately 10 m, which can be
considered to be a reasonable maximum operational distance
in U-HRI. In terms of Jerlov water types, this corresponds
to
b= 05 0150 90
channels, respectively.
Experiments and Results
Data Set and Setup
The presented approach to U-HRI originated within the EU
project CADDY. Major efforts were devoted in this project to the
collection of data, including experiments on the use of underwater
gestures. Data were recorded in the open sea as well as in
indoor and outdoor pools at three different locations, namely, in
Biograd na Moru (Croatia), at the Brodarski Institute in Zagreb
(Croatia), and in Genova (Italy). The data are divided into eight
scenarios representing different diver missions and field experiments.
The scenarios, named Biograd-A, Biograd-B, and Genova-A,
represent trials that were mainly organized for data
collection; they, hence, feature a high number of samples. The
other scenarios, Biograd-C and Brodarski-A-D, cover experimental
or real diver missions. A detailed discussion of the number
of samples and environmental conditions of each scenario is
provided in [39]. For the evaluation of the different ML methods
here, they are trained according to the partition of the data
detailed in Table 2.
Each method described in the " Gesture Detection and Classification "
section has four Model X versions. The partition is
made to gather samples with similar environmental conditions
(e.g., location, light, and so on) and observe how the methods
perform against unseen types of data. Samples from Biograd-C
and Brodarski-B and -D are used only as test sets.
[. ,. ,. ] for the red, green, and blue
IEEE Robotics & Automation Magazine - September 2021

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2021
