Shipborne Electro-Optical Ship Tracking Algorithm Based on YOLOv8s Pruning and Dynamically Weighted ByteTrack

Addressing the issues of model redundancy leading to deployment difficuties and poor tracking continuity in occlusion scenarios in traditional shipborne electro-optical tracking systems, a ship tracking algorithm based on YOLOv8s pruning and improved ByteTrack is proposed in this paper. Firstly, pruning processing is performed on YOLOv8s to simplify redundant structures and reduce the model size. Then, the ByteTrack tracker is improved. A dual-temporal association mechanism combining inter-frame feature persistence and historical trajectory backtracking is introduced, and Kalman filter is integrated to improve the state prediction accuracy in occluded scenarios. Finally, the fusion strategy of temporal features and geometric matching is optimized through weight coefficient adjustment to enhance the stability of target matching. Verified based on the SeaShips dataset covering typical shipborne electro-optical scenarios, the results show that the improved algorithm achieves a mAP50 of 92.9%, MOTA of 82.3%, and an occlusion recovery success rate of 91.2%. Compared with the YOLOv8s+ByteTrack scheme, the effective parameter count (non-zero weights) is reduced by 30%, the total parameter count is reduced by about 30% (from 11.2 M to 7.8 M), the model file size is reduced by 11.8%, and MOTA is increased by 4.2%, which fully meets the core requirements of "real-time monitoring + rapid deployment" for shipborne electro-optical systems.