R3-BEV: Routing, Rasterization, and Reliability-Aware Fusion for Radar-Camera 3-D Object Detection
Jinhao Li, Xueying Bai, Quanyi Liu, Shenghua Xiong, Zerong Zhao, Tianbin Huang, Haibin Wang · IEEE Access · 2026
3D object detection is a fundamental task in autonomous driving. Although multi-view camera-based BEV methods have achieved strong performance, they remain limited in long-range scenes, adverse illumination, and motion-aware perception. Millimeter-wave radar provides direct range and velocity measurements and is more robust under challenging weather conditions, making radar-camera fusion in the BEV space a promising solution. In this paper, we propose R3-BEV, a unified radar-camera BEV 3D detector that jointly addresses radar encoding, radar BEV formation, and cross-modal fusion. Specifically, R3-BEV introduces a physics-aware radar routing encoder for voxel representation, an adaptive patch-splatting rasterization module for radar BEV generation with an evidence density map, and a reliability-aware fusion module for multimodal BEV aggregation. Experiments on the nuScenes dataset show that R3-BEV achieves a favorable NDS and box-quality trade-off, with competitive radar-camera BEV detection performance. These results demonstrate the effectiveness of explicitly modeling radar physical cues and reliability-aware evidence-guided fusion for radar-camera BEV perception.