Incompleteness-aware radar-vision fusion transformer framework for water-surface object detection
Qingwang Han, Ruidan Su, Jialin Wang, Rui Gao, Li Zhou · Ocean Engineering · 2026
In this article, we address the problem of performance degradation in water-surface object detection under incomplete and unreliable sensory conditions. We propose an incompleteness-aware Real-Time DEtection TRansformer (IA-RTDETR) method, which incorporates radar and vision in a unified transformer-based detection framework to enhance robustness against modality degradation. We first build a radar-to-image pseudo-color mapping module to project sparse radar point clouds into image-aligned representations and resolve cross-modal spatial misalignment. On this basis, an asymmetric dual-stream backbone is adopted to extract modality-dedicated features, where a heavy visual branch captures rich semantic textures and a lightweight radar branch encodes geometric cues for balanced accuracy and efficiency. Further, we design a reliability-guided multi-scale cross-modal fusion (MCF) module with learnable channel gating. Optimized via a tailored loss objective, MCF dynamically assigns higher weights to intact reliable modalities and suppresses degraded visual features, materializing the core incompleteness-aware design of our framework. Experimental results demonstrate that the proposed IA-RTDETR reaches 92.1% in [email protected] and 66.2% in [email protected] - 0.95 , outperforming the upper bound of existing methods by 5.3% and 4.0%.