Knowledge-Assisted Intelligent Maritime Multi-Ship Tracking

Deep learning-based methods for multi-object tracking (MOT) in maritime scenarios often encounter challenges such as high false negative(FN) rates, frequent identity switches, and discontinuous trajectories due to lighting variations and occlusions among ships. These issues arise because traditional tracking methods treat tracking as a subsequent task to detection, resulting in direct failures when detection fails. To overcome these challenges, we propose a novel knowledge-assisted maritime multi-ship intelligent tracking algorithm. By integrating the knowledge that 'objects do not suddenly disappear' into the MOT framework through a target search method, our approach ensures that once an object is tracked, it will not be lost, thus minimizing dependence on the detector, reducing FN rates, and maintaining trajectory continuity. Our experimental results demonstrate that compared to traditional MOT methods that use the same detector, our tracking approach achieves a 20% reduction in FN rates and a 15% increase in Multiple Object Tracking Accuracy (MOTA).