IEEE - Aerospace and Electronic Systems - March 2023 - 11

Giusti et al.
Table 5.
Radar Main Parameters Used During the Measurements
B
PRF
Sampling frequency
Sweep duration
Sweep samples
Elevation beamwidth
(3dB)
1 GHz
1000 Hz
2 MHz
8.25e4s
1650
50
Azimuth beamwidth (3dB) 20
MULTISPECTRALDATAANALYSISANDDETECTIONSRESULTS
A semantic segmentation algorithm making use of convolutional
neural network (CNN) has been used to detect targets at
sea. Even though all spectral bands are synchronized, images
need to be aligned. A three step pipeline is implemented: i)
image distortion correction, ii) images warp and alignment
using a homography matrix, iii) combination and
cropping of the aligned to remove nonoverlapping area.
The CNN architecture designed for pixel-level semantic
segmentation is based on an encoder-decoder architecture,
similar to that proposed in [16]. However, instead
of using a VGG16 architecture as a basis for the encoding
scheme, here it is replaced by a multiresolution feature
pyramid network (FPN). The FPN was proposed in
[17] to enhance the performance of multiscale object
detection and showed improved results compared to
other approaches. A pretrained CNN has been retrained
using a portion of the new set of multispectral images
acquired during the measurements. A set of other multispectral
images obtained from the rest of the flights are
used for validation only. Figure 7 shows an example of
aligned and merged images and an example of target
detection and segmentation.
Although the developed algorithm correctly detected a
large part of the debris tied to the vessel, it did not detect
all of the debris, and in the areas where the waste is more
submerged, the system was not capable to detect the presence
of these residues, providing a few pixels that correspond
to false negatives.
Finally, it should also be noted that in certain areas, the
system erroneously predicts the presence of debris (false
positives). It is not a behavior that has been given frequently,
and it is only observed in a very small percentage
ofthe images. This behavior has been observed basically in
some rock, and also in some random area of the ocean,
mainly due to the presence ofreflections from the sun.
The CNN-based detection algorithm exhibits a detection
rate of roughly 80%, which is comparable to the
results presented in [9] and [10].
A dedicated module has been developed to convert the
image coordinates of detections into latitude and longitude
for target localization. The oblique geolocation module is
in charge of estimating the approximate geographical coordinates
of the detections. Figure 8 shows the EO/IR georeferenced
detections close to the true USV GPS location.
The two cameras have been oriented using the same
tilt angle of the radar system, namely, 45. However, this
setup makes the geolocalization process more complicated
Figure 9.
An example of SAR image of the observed area at the sea obtained using the BPA and despeckling. Roughly 16 seconds of radar raw data
have been integrated.
MARCH 2023
IEEE A&E SYSTEMS MAGAZINE
11

IEEE - Aerospace and Electronic Systems - March 2023

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - March 2023