Robust Distant Small Object Detection Using Radar-Vision Fusion and Prior-Enhanced Query Generation in Autonomous Driving

Detecting distant small objects under adverse visual conditions such as rain, fog, or low light remains a critical challenge in autonomous driving. To address this issue, we propose a novel radar-vision fusion detection network that incorporates three key components. First, the Spatially Partitioned Radar Projection (SPRP) decomposes radar point clouds into far-field and near-field domains through bidirectional sampling, mitigating information loss for distant objects in single-frame projection during perception range extension. Second, the Cross-modal Weighted Fusion (CMWF) module implements competitive sensor weighting to adaptively fuse radar and visual features across various environments, prioritizing robust modality cues under sensor degradation while enhancing spatial discriminability. Third, the Prior-Query Generation Module (PQGM) leverages radar-camera fused features to initialize detection queries enriched with radar-derived spatial priors for distant small object focusing. This module incorporates an integrated auxiliary branch employing one-to-many matching to refine the quality of generated queries during training. Extensive experiments demonstrate state-of-the-art performance of 35.2% mAP on nuScenes and 45.0% mAP on a custom dataset, achieving improvements of up to 1.8% mAP compared to existing fusion methods.