Depth-Prior-Guided Detection and Turning-Aware Association for Multi-Object Tracking of Marine Fish in Aquaculture Tanks

Continuous tracking of individual fish provides a basis for movement analysis and visual monitoring in marine aquaculture. Underwater image degradation, similar appearance, and non-rigid motion complicate fish detection and identity preservation. This study develops a tracking framework for golden pompano (Trachinotus ovatus) and large yellow croaker (Larimichthys crocea) recorded in a land-based aquaculture tank. The framework combines a depth-structure-prior-enhanced RT-DETRv4 detector with the Scale-aware and Unscented Tracker (SU-T). A frozen Depth Anything 3 (DA3) model provides supervision based on physics-inspired structural cues derived from relative depth during training, while a lightweight student head predicts the prior without online teacher inference. Fish Turning–Deformation Coupling Consistency (FTDC) uses the empirical relationship between short-term motion-direction change and observable bounding-box variation in trajectory association. On the self-collected dataset, the final detector increased average precision over intersection-over-union thresholds of 0.50–0.95 (AP50:95) from 0.837 to 0.851. With identical detection inputs, SU-T+FTDC achieved 88.3% multiple-object-tracking accuracy, 77.4% identification F1 score, and 62.8% higher-order tracking accuracy; identity switches decreased from 66 to 60 and trajectory fragmentations from 86 to 78 relative to SU-T. These results support improved fish localization and identity continuity, providing a trajectory basis for subsequent movement and behavior analyses in aquaculture.