A practical adverse-weather maritime vessel detection–tracking framework via an enhanced YOLOv11 detector and BoxMOT association

Maritime visual surveillance is increasingly used as a complementary sensing modality to radar/AIS in coastal VTS and onboard situational awareness, yet its reliability degrades under sea clutter and adverse weather. This paper presents a practical tracking-by-detection framework that couples an enhanced YOLOv11 detector with lightweight data association. The detector adopts a GELAN backbone and a weighted bi-directional feature pyramid network (BiFPN) to strengthen multi-scale representation, and replaces the regression objective with Unified IoU (UIoU) to improve box tightness for IoU-based association. To stabilize detections across domains and mitigate fine-grained label jitter, we construct a cross-domain single-class training set (Ship-1C) by fusing SeaShips images with frames sampled from SMD. For adverse-weather evaluation, we build SeaShips-Complex via physically motivated fog/rain/snow synthesis and further conduct external validation on real rainy and snowy WaterScenes images. On two representative maritime video sequences, including a self-collected low-visibility hazy video and a public waterway/port video, a maritime-tuned ByteTrack configuration (high detection threshold + low IoU matching threshold) attains 89.74% MOTA with low identity switches. The proposed detector runs at 344.83 FPS at the detector level on an RTX 4060 GPU. The results support the feasibility of deploying the proposed framework for shore-based monitoring and engineering-grade maritime perception.