SeaTrack-DeepSORT: A Vision-Based Ship Detection and Multi-Object Tracking Framework for Complex Maritime Environments

With the rapid development of intelligent maritime systems and unmanned surface vehicles (USVs), accurate ship detection and multi-object tracking in complex maritime environments have become essential for autonomous perception and intelligent navigation. However, challenges such as scale variations, background interference, illumination changes, target occlusions, and camera motion significantly degrade tracking performance. To address these issues, this paper proposes SeaTrack-DeepSORT, a maritime detection and tracking framework that connects multiscale detection with confidence-dependent state updates and motion-compensated association. Its SeaDet-YOLO detector incorporates ADown downsampling and an additional P2 detection head into YOLOv11. The DeepSORT-based tracker combines quadratic measurement-noise scaling, ECC-Affine correction of predicted target centers, and two-stage association, using detection confidence to guide both association and the weighting of matched observations. Extensive experiments on multiple maritime datasets show that SeaTrack-DeepSORT improves detection and tracking performance while maintaining stable and continuous trajectories. The resulting framework supports autonomous perception and intelligent monitoring for unmanned surface vehicles.