DenseRadar: Dense Radar Generation and Point Consistency Aggregation for 3D Object Detection
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Accurate environmental perception is critical for autonomous vehicles, typically achieved through multi-sensor fusion. However, existing camera-radar fusion methods often neglect effective feature aggregation across modalities, leading to suboptimal performance in real-world scenarios. To address this, we propose DenseRadar, a novel framework that integrates dense radar generation and point consistency aggregation for 3D object detection. Specifically, we design a Radar Dense Generation (RDG) module that leverages dense camera features to enrich sparse radar data through a Camera-Radar Interaction Refinement mechanism. Additionally, a Point Consistency Aggregation (PCA) module is introduced to guide sampling offsets toward object regions, enhancing feature focus on targets. Further, a Camera-Radar Feature Aggregation (CRFA) module employs a Mask-Guided Attention (MGA) mechanism, where a filtering enhancement mask extracts object heatmaps to refine cross-modal interactions in BEV space. Extensive experiments on the nuScenes dataset demonstrate that DenseRadar achieves competitive and robust performance, particularly in the overall detection score (NDS, 58.0%), with notable robustness under challenging conditions such as occlusion, adverse weather, and complex urban scenes. The proposed method provides a practical solution for high-precision, multi-modal perception in autonomous driving systems. Our code will be available inhttps://github.com/no-Name128/DenseRadar.git
Fields of Study & Subjects
Autonomous Vehicle Technology and Safety
Advanced Neural Network Applications
Radar Systems and Signal Processing
Publication details
- Journal / source
- IEEE Transactions on Multimedia
- Volume
- 28
- Pages
- 8957–8966
- Publication date
- 1 January 2026
- Language
- EN