Abstract
<title>Abstract</title> <p>Fixed-camera monitoring of densely stocked fish provides essential trajectory data for intelligent aquaculture; however, frequent crossings, mutual occlusions, and substantial overlaps among visually similar individuals often introduce association ambiguity and compromise identity consistency. Conventional association methods based primarily on motion prediction and intersection over union (IoU) become unreliable when different fish exhibit similar spatial and motion states, whereas deep appearance models may incur considerable computational overhead and exhibit reduced discriminability under degraded imaging conditions. To address these limitations, this study proposes a lightweight appearance-guided auxiliary association module for real-time multi-fish tracking. The module integrates color, gradient, and morphological descriptors and employs extreme gradient boosting (XGBoost) to estimate the identity-matching probability between each candidate detection and existing track. This probability is adaptively incorporated into the association cost matrix of Simple Online and Realtime Tracking (SORT), enabling appearance information to complement spatiotemporal cues when association ambiguity increases. Leave-one-video-out experiments on seven sequences from the Multiple Fish Tracking 2025 (MFT25) dataset demonstrate that the proposed method reduces identity switches from 1,513 to 833 relative to SORT, corresponding to a reduction of 44.94%. Moreover, the average multiple object tracking accuracy increases from 98.52% to 98.82%, while the average identity F1 score improves from 53.75% to 65.49%. Excluding object detection, the tracking speed exceeds 700 frames per second. These results indicate that the proposed module improves identity preservation under frequent occlusion and overlap while maintaining computational efficiency, demonstrating its potential for real-time fish monitoring in intelligent aquaculture.</p>