End‐to‐End Multi‐Object Tracking With Context‐Guided Inference in Maritime Radar Image Sequences
Zhan Kong, Wei Xiong, Yaqi Cui, Zhenyu XiongABSTRACT
End‐to‐end (E2E) multi‐object tracking (MOT) methods mainly focus on improving model architectures and training strategies, while scene context available in specific applications is often under‐exploited. In maritime radar image sequences, contextual cues such as route priors and temporal relationships between adjacent targets can provide valuable guidance, especially when weak appearance features and target merging cause track fragmentation and identity switches. This letter proposes an E2E tracking framework with context‐guided inference for maritime radar image sequences, which exploits maritime context when available to improve inference‐stage trajectory management through route‐constrained re‐identification, adaptive non‐maximum suppression exemption and context‐aware identity correction. Experiments on real‐world maritime radar data demonstrate that the proposed method achieves 90.49% multi‐object tracking accuracy (MOTA) and reduces identity switches (IDs) by 50% compared with the baseline.