A Real-time Tracking Algorithm for Underwater Fish Based on Track by Detection Framework

Fish tracking under natural waters has a wide range of application values for fish behavior analysis, species observation and water quality condition monitoring. Aiming at the low efficiency and poor real-time performance of current underwater fish tracking algorithms, A real-time marine fish tracking algorithm based on the Track by Detection framework is proposed in this paper. The algorithm improves the Yolov5 detector and Sort tracker, effectively enhancing detection and tracking accuracy while considering both accuracy and speed. Firstly, the CBAM attention mechanism is introduced into the Yolo5 model to better capture key information in the image and improve task accuracy and robustness. In this paper, an optimised version of the SORT algorithm is presented for the real-time tracking of underwater fish, where the Kalman filter has been removed to improve the tracking accuracy and speed. AP and FPS are chosen as the detection evaluation indexes, and MOT challenge standard is used as the tracking evaluation indexes in the Experiments. The experimental results demonstrate that the improved algorithm has increased the tracking accuracy index MOTA by 0.8% compared to the previous version. Additionally, the tracking precision index MOTP has improved by 2.1% compared to the pre-improvement version. The tracking speed has also increased to 75.6 FPS, meeting the requirements for real-time tracking.