Skip to content
Book Open access

SPRFusion: Semantic and Physics-driven Robust Radar-Camera Fusion for 3D Object Detection

Nov 2026 · 0 citations

Abstract

Radar-camera fusion combines visual semantics with reliable radial depth and Doppler velocity for 3D object detection in autonomous driving. However, sparse and noisy radar measurements, cross-frame misalignment of dynamic objects, and cross-modal semantic discrepancies and reliability variations hinder robust fusion. We propose SPRFusion, a query-based radar-camera fusion framework guided by semantic and physical priors, comprising three complementary components. The Cross-Modal Temporal BEV Alignment (CMTA) module uses radar Doppler cues to estimate a dense displacement field and warp historical camera BEV features, improving temporal consistency for dynamic objects. The Difficulty-Aware Circular Distribution (DACD) strategy initializes object queries from a distance-dependent difficulty field, allocating more queries to regions affected by occlusion, image degradation, and radar sparsity. The Dual-Robust Cross-Modal Decoder calibrates radar-guided query representations using a static LLM-derived physical-semantic codebook and adaptively suppresses features from unreliable modalities through uncertainty-aware gated fusion. Extensive experiments on nuScenes and View-of-Delft (VoD) demonstrate the effectiveness and robustness of SPRFusion. With a ResNet-50 backbone, SPRFusion achieves 62.4% NDS and 55.4% mAP on the nuScenes validation set. With a V2-99 backbone, it achieves 68.2% NDS and 62.9% mAP on the official test set without future frames, test-time augmentation, or model ensembling, demonstrating its effectiveness as a robust multimodal 3D detector.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.