Abstract: The existing fish trajectory extraction methods fail to balance efficiency and accuracy. This study introduces a fish trajectory extraction approach based on fish landmark recognition and location utilizing the RetinaFace algorithm. The method entails constructing a fish trajectory extraction model through enhanced network structure and loss function for landmark detection, optimizing anchor size design, and encoding and decoding fish landmarks (specifically, the head point and centroid point). Additionally, it involves supplementing landmarks of fish targets with extra labels and generating a fish key point dataset. The findings demonstrate that the proposed research method achieves high accuracy in identifying fish landmarks, with precision evaluation indices including an accuracy rate of 97.12%, a recall rate of 95.72%, and a mean average precision of 96.42%. Moreover, the average relative deviation of the extracted trajectory coordinates is MREx(0.065%,0.092%) and MREy(0.112%,0.011%), aligning closely with the actual swimming trajectory of fish. The recognition rate for landmarks of fish targets reaches 32 frames per second, which meets the real-time extraction requirements for fish trajectory recognition.

Key words: fish, fishway monitoring, detection of fish landmark, fish trajectory extraction, RetinaFace model