FHA-MTT: A fish health assessment method based on multi-target tracking

• Proposed SPSE-YOLO-pose, a novel detector that fuses SPD-CBS downsampling with SE channel attention to improve small-target and keypoint localization. • Combined SPSE-YOLO-pose with BotSORT to overcome identity switches and trajectory fragmentation in dense/occluded aquaculture scenes. • Extracted and validated four quantitative behavioral health indicators — TSD, MSA, MTA and MC — for individual fish health assessment. • Achieved high detection and tracking performance (mAP@50(Box) 99.31%, mAP@50:95(Pose) 98.28%; BotSort+SPSE HOTA 92.77%, MOTA 99.15%, IDF1 91.2%, IDSW = 12) with behavioral measurement errors controlled within −7.5% to +5%. • Provided a machine vision–based framework (FHA-MTT) for automated, non-invasive assessment of fish health through multi-target tracking. In aquaculture, the health status of crucian carp directly impacts production efficiency, and swimming behavior is an important indicator reflecting its physiological state. Existing vision-based methods for analyzing fish behavior lack a systematic correlation with health status. Furthermore, under complex scenarios such as target occlusion and posture changes, issues like detection drift and trajectory fragmentation exist, making it difficult to achieve stable and continuous monitoring of individual behavior. To address this, this paper proposes a fish health analysis method based on multi-target tracking (FHA-MTT) to achieve precise capture of crucian carp swimming behavior and quantitative assessment of health status. This paper constructs the SPSE-YOLO-pose detection model based on YOLOv8n-pose, introducing the SPD-CBS module to enhance spatial detail retention capability and embedding the Squeeze and Excitation attention mechanism to improve keypoint localization accuracy. For the tracking stage, we employ the BotSORT algorithm, which combines motion prediction with re-identification feature matching to effectively suppress identity switches. Finally, based on keypoint trajectories, behavioral metrics such as total swimming distance, mean acceleration, mean turn angle, and mean curvature are extracted to construct a health status assessment system. Experimental results show that SPSE-YOLO-pose improves the mAP@50:95 for object detection and keypoint localization tasks to 66.40% and 98.28%, respectively; the BotSORT tracker performs excellently on multiple metrics, achieving a higher order tracking accuracy of 92.77%, a multi-object tracking accuracy of 99.15%, with the number of identity switches being only 12; the extraction error for swimming behavior parameters is controlled between -7.5% and +5%. This study verifies the feasibility of the FHA-MTT method for automated and refined assessment of crucian carp health status in complex aquaculture environments, providing a reliable technical pathway for early disease warning in fish.