Feature-Enhanced DeepSORT for Shallow-Sea Biological Target Tracking and Water-Intake Invasion Warning

The episodic aggregation of shallow-sea organisms near coastal nuclear power plant water intakes can obstruct filtration facilities and threaten cooling-water circulation. Reliable multi-object tracking is therefore required for continuous monitoring and warning-oriented decision support. However, underwater targets commonly show weak texture, similar appearance, partial occlusion, and current-driven nonlinear motion, which cause trajectory fragmentation and identity switches in conventional trackers. This study proposes a feature-enhanced DeepSORT framework for shallow-sea biological target tracking and intake-invasion warning. An improved YOLOv8 detector is used as the detection front end, while the main methodological contribution is an enhanced tracking module. Efficient Channel Attention, RepVGG, and an enhanced Squeeze-and-Excitation block are incorporated into the Re-ID feature extractor to improve appearance discrimination under turbid and low-texture conditions. An extended Kalman filter is further introduced to improve motion prediction for drifting, turning, and short-term occluded targets. Based on the tracked trajectories, equivalent density, and velocity component toward the intake, an invasion-intensity index and graded warning strategy are established. Experiments on shallow-sea biological video data show that the proposed tracker improves IDF1 from 55.7% to 56.3%, MOTA from 43.6% to 44.7%, and MOTP from 74.3% to 75.1% compared with DeepSORT, while reducing identity switches from 779 to 735. These results indicate that the proposed method can provide more stable trajectory information for early warning of biological blockage risks at coastal nuclear power plant intakes.