Abstract
Monitoring fish growth is critical for intensive aquaculture, yet traditional manual weighing methods are labor-intensive and inefficient. This study proposes FishPose-tracking, an enhanced deep learning model that automates keypoint detection and tracking of fish, improving the efficiency and accuracy of growth monitoring in complex aquaculture environments. Built upon the YOLOv8-Pose model, FishPose-tracking integrates several advanced components: underwater image enhancement, precise keypoint detection, posture estimation, and continuous fish tracking. Image quality is optimized through multi-level bilateral filtering and regional dark channel dehazing, addressing challenges posed by turbid and low-light underwater conditions. To improve feature extraction, FishPose-tracking incorporates the Swin Transformer encoder, which offers robust multi-scale capabilities, and the STIAU module, which enhances temporal consistency across video frames to reduce detection errors from motion blur and occlusion. The model was trained on a diverse dataset of 10,000 images encompassing multiple fish species under varying degrees of occlusion. Experimental evaluations show that FishPose-tracking improves detection and tracking precision, achieving mean Average Precision scores of 99.37% (mAP50) and 87.39% (mAP50:95), representing improvements of 13.74% and 8.86%, respectively, over baseline models. Residual analysis confirmed the high accuracy of the model, with the median keypoint error for the five species being less than or equal to ±2 pixels (IQR ≤ 3 pixels). FishPose-tracking provides a robust and real-time capable monitoring solution adaptable to various aquaculture environments. Future work will explore further enhancements, including integration with multi-source data and large-scale deployments.