A robust maritime target tracking framework for shore-based surveillance via cascaded fusion and occlusion compensation of AIS, visual, and radar data

Abstract

To address multi-modal tracking discontinuities in shore-based maritime surveillance caused by low-frequency AIS updates, visual occlusions, and radar identity voids, this study proposes a cascaded multi-source fusion tracking framework. Unlike conventional parallel or weighted fusion schemes that struggle with heterogeneous data synchronization, the proposed cascaded architecture decouples sensor limitations layer-by-layer. It specifically resolves identity loss during concurrent AIS update gaps and visual failures.

Utilizing YOLOv11 — selected for its superior inference efficiency in maritime edge computing — and DeepSORT as a baseline, two innovations are introduced: 1) A Fusion Prior Occlusion Compensation (FPOC) mechanism that leverages AIS-derived motion priors to maintain visual trajectory continuity. 2) A cascaded feedback loop that re-injects fused ”AIS-Radar” results into the ”AIS-Video” layer to correct identity offsets using radar motion features.

Experimental results on field datasets and the FVessel_v1.0 dataset show that the framework significantly outperforms the DeepSORVF method. Specifically, the proposed model achieves a 3.56% average improvement in MOFA and IDF1 on public benchmarks while maintaining high processing efficiency. This confirms the framework’s effectiveness in enhancing real-time maritime safety management.