Abstract
Aquaculture is of increasing importance as an alternative food supply that mitigates climate change, and an industrial domain with surprising growth as of lately. Unfortunately, due to the notorious logistics of industrial-focused large scale underwater research, and the potential problems that introduces to state-of-the-art methods, only minor focus has been given by the robotics community. This paper aspires to increase the interest in aquaculture robotics, by exposing some major unique challenges this domain poses to relative localization and mapping. Robots are expected to operate in highly dynamic and uncertain environments with no static visual reference, absence of GNSS signals, and interference from a livestock of up to 200,000 units. With this goal in mind, we provide the community a dataset gathered from an industrial-scale fish farm, including monocular cameras, stereo cameras, and acoustics to promote further research. Furthermore, we apply three known state-of-the-art methodologies to demonstrate its value and experiment with different sensing modalities. The first method exploits the net pattern of fish cages to estimate the relative distance via Fourier transformation (FT) on frames collected with a monocular camera, the second method utilizes stereo vision for both acquiring disparity maps and relative pose estimation, and finally acoustic relative localization is shown with a 360° scanning sonar. The contributions are accompanied with relevant analysis and discussion regarding feasibility and potential limitations in aquaculture robotics.