ERC-Fusion: Fusing Enhanced 4D Radar and Camera for 3D Object Detection

In autonomous driving, the fusion of camera and 4D millimeter-wave radar point clouds significantly enhances perception robustness and accuracy in harsh environments and off-road scenarios by combining high-resolution vision with precise radar data. This improves target detection and environment modeling capabilities, thereby ensuring the vehicle's safe and efficient operation. However, the sparsity of point clouds presents significant challenges for multimodal fusion in terms of data density and resolution matching, affecting the accuracy of environmental perception. To address this issue, we have built a framework called ERC-Fusion, which specifically includes: we designed a KNN-based point cloud augmentation algorithm to enhance the density of sparse point clouds and increase the amount of information. Additionally, we developed a double-voxel feature fusion enhancement module, which constructs sparse voxels at different scales from the raw data to capture detailed target features and reduce data noise. We validated the performance of ERC-Fusion on the View-of-Delft dataset. Experimental results show that ERC-Fusion effectively alleviates the sparsity issue of radar point clouds and successfully extracts detailed features, thereby improving 3D object detection performance.