Introduction: Advancements in algorithms and practical applications within aquaculture highlight its potential for the digitalization of the industry. However, accurately detecting and tracking fish in higher -density environments remains a challenge. This study proposes a pose estimation method for monitoring the behaviour of high numbers of fish stocked in RAS tanks.
Materials and Methods: RAS tanks (2m wide and 1m deep) holding ca. 1,000 Atlantic salmon (Salmo Salar ) were continuously recorded using overwater dome cameras for 60 days. To analyse individual fish pose accurately, we employed the state-of-the-art pose model based on Yolo8 architecture , trained to detect fish key points. Using the coordinates of the obtained key points from a one-hour test video, we constructed a vector between the dorsal and snout keypoints for each fish, representing their swimming orientation. Our study investigated fish behaviour based on orientation . We observed rheotaxis (swimming against the current) and cohesive schooling. The cosine similarity of orientation vectors quantified how similarly neighboring fish were oriented. A threshold on the orientation score identified fish deviating from the school. Finally, orientation vectors and keypoints were visualized to analyse behaviour.
Results and Discussion: For the test video the model detected upto 226 out of 1000 fish. While a detection rate of 226 fish (22.6%) might appear low, it is important to consider that this successfully captured most visible fish at the surface, with high pose precision (96.7%). Visualizing different events in the video revealed a strong correlation between the number of detections and orientation scores. This suggests the model has potential to understand fish behavior related to feeding and disturbances.