Robust situational awareness for ships and offshore traffic: integrated detection and tracking via buoy-mounted cameras

Effective maritime situational awareness is hindered by environmental clutter and the difficulty of detecting small, distant vessels. This paper proposes Track-YOLO, a unified detection-and-tracking framework. For detection, a Cross-Attention Convolutional Gating (CACG) module suppresses background-induced false positives, while an Omni-Dimensional Dynamic with CSP (ODCSP) network enhances feature extraction for small and fast-moving targets. P-IoU v2 loss further improves localization accuracy by minimizing anchor-target edge distances. For tracking, an Interactive Multiple Model (IMM)-Enhanced Kalman Filter dynamically combines linear and accelerated motion models to handle maneuvering targets and occlusions better. Experiments on the ShipTrack dataset show that Track-YOLO outperforms YOLOv11 and ByteTrack, achieving 9.5% higher mAP@.50:95 and 10.8% higher HOTA. The optimized model contains only 2.45 million parameters and achieves 95.3 FPS on an RTX 4060Ti GPU, demonstrating its suitability for real-time maritime surveillance.