Humankind has lived in harmony with the sea for thousands of years, but pop-
ulation growth and consumerism in recent centuries have disrupted this fragile
ecosystem. The exploitation of the ecosystem for food, energy, and transporta-
tion has had a significant impact on biodiversity and the populations of individual
species in recent decades. The first step in species conservation is to monitor their
populations and assess human impact. Video surveillance is a solution for moni-
toring populations that is already well established on land and, to a lesser extent,
underwater. For this purpose, we have designed two types of cameras in collabo-
ration with Anemo Robotics and defined a procedure for installing cameras in the
field, collecting data, training a computer vision model, and analyzing footage to
count the populations of individual species. Battery-powered and wired cameras
enable long-term observations in different areas regardless of the existing infras-
tructure. In the video recordings of the pilot project for the battery-powered
camera in the Danish port of Hunsted, marine biologists identified 10 species
of marine animals, which provided us with an extremely unbalanced dataset on
which we trained two sizes of two models of different architectures: YOLOv8n,
YOLOv8m, RF-DETR Base, and RF-DETR Large. We used general metrics to
evaluate the performance of the models and their computational complexity, and
the confusion matrices showed which species were difficult. We used the mod-
els to count the maximum number of instances of each species in 27 videos and
compared the number with the manually counted number. We also evaluated
generalization by predicting on unseen data from Bilbao, Spain.
|
