physfusion: A Transformer-based Dual-Stream Radar and Vision Fusion Framework for Open Water Surface Object Detection

Detecting water-surface targets for autonomous surface vehicle (ASVs) is challenging due to nonstationary wave clutter, specular reflections, and frequent long-range observations, where small or low-profile objects exhibit weak appearance cues. Although 4-D millimeter-wave radar can complement cameras under degraded illumination, maritime radar point clouds are typically sparse and intermittent, and reflectivity-related attributes often present heavy-tailed variations under scattering and multipath, making it difficult for conventional fusion designs to exploit radar cues effectively and maintain stable predictions across frames. We propose PhysFusion, a physics-informed radar–image detection framework tailored for water-surface perception. Instead of relying on direct fusion of sparse radar returns and dense visual features, PhysFusion improves the multimodal detection from three complementary aspects: radar representation learning under cluttered and heavy-tailed radar observations, query-level radar-guided cross-modal interaction for more effective fusion with multiscale visual features, and lightweight temporal aggregation for stable frame-to-frame prediction. This design enables more reliable radar cue utilization and more consistent multimodal detection in challenging water-surface environments. Experiments on WaterScenes and FLOW demonstrate that PhysFusion achieves 61.5% mAP50:95 and 90.3% mAP50 on WaterScenes (with T = 5 radar history) using 5.6M parameters and 12.5G FLOPs and reaches 94.8% mAP50 and 46.2% mAP50:95 on FLOW under the radar–camera setting. Ablation studies further verify the effectiveness of the proposed design under a consistent evaluation protocol.