Skip to content
Book Open access

Radar-Camera 3D Object Detection via Density- and Intensity-Guided Refinement with Reliability-Aware Fusion

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 52 references

Abstract

Radar-camera 3D object detection has shown great promise for autonomous driving, yet the inherent sparsity and noise of radar point clouds continue to hinder the detection performance. Existing methods typically adopt single-branch radar enhancement, which struggles to balance feature completeness and suppress noisy responses. To address this issue, we propose a dual-branch radar feature refinement framework that enhances radar representations from complementary perspectives. Specifically, a density-guided branch employs a statistically estimated density prior for sparse feature enhancement, while an intensity-guided branch utilizes a physically derived intensity prior for noise suppression. To offer richer semantic context, we further employ a camera-prior mask generated by spatial attention to enrich radar foreground features. Since radar refinement can further amplify the inherent cross-modal inconsistency, we design a deformable bidirectional alignment fusion module for cross-modal guided geometric alignment and adaptive feature fusion through bidirectional learnable offsets. In addition, a reliability-aware module generates reliable scores based on low-level visual cues to ensure robustness across diverse imaging conditions. Extensive experiments on the nuScenes and View-of-Delft( i.e., VoD) datasets demonstrate the effectiveness of the proposed framework. Compared to the baseline, our method improves mAP by 5.6% and NDS by 6.6% on the nuScenes dataset.

Read PDF

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