Abstract
The shift towards Precision Fish Farming (PFF) requires reliable, automated tracking of individual fish in dense, dynamic underwater environments. Many approaches targeting this topic are based on using computer vision methods and AI to detect and track the fish automatically. Most conventional approaches rely on the use of Horizontal Bounding Boxes (HBBs), however this approach has proven inadequate for accurately localising slender fish with variable orientation, particularly in scenes where fish density and occlusion is high. This study presents a new computer stereo vision framework that leverages Oriented Bounding Boxes (OBBs) for fish detection followed by keypoint detection for the localisation of the anatomical features of the detected fish. The method was set up to directly estimate 3D positions of the detected fish and keypoints and track these over time without external calibration targets. When applied to stereo video data collected from two field trials at commercial fish farms, the method estimated fish-camera distances and fish lengths that corresponded with reference data obtained through other sensors and manual sampling. These results imply that this new approach represents a step forward in non-invasive, automated monitoring and analysis of fish behaviour and growth, giving more detailed insights into farmed fish dynamics and contributing to further progress in the realisation of PFF.