4D Radar-Driven Road User Detection based on Spatial Density Feature Fusion
Xinxian Huang, Huabiao Qin, Yixiang Xie, Xianda Guo · 2025
4D Millimeter-wave Radar (4D Radar) has demonstrated remarkable potential and broad application prospects in 3D object detection tasks because of its robustness in extreme environments and excellent velocity measurement capabilities. But, the severe sparsity and noise issues inherent in 4D Radar point clouds limit their application in 3D object detection. Existing LiDAR-based algorithms fail to effectively leverage the correlations between sparse points and Doppler information, resulting in performance that falls far short of expectations. To address this problem, a novel 3D object detection method is proposed—a 4D Radar point cloud network based on spatial density estimation. In the point branch, a Spatial Density Estimation (SDE) module is designed based on PV-RCNN. It captures the spatial density information of sparse point clouds by efficiently encoding features from the raw point cloud, which enhances the density of point cloud features and effectively mitigates information loss caused by sparsity. The point cloud features enriched with spatial density information enable better integration of key points with multi-scale voxelization features, thereby improving detection accuracy. Experimental results on the View-of-Delft dataset show that, compared with the baseline methods PV-RCNN, the proposed method improves 3D mAP by 0.73% across the entire annotated range and by 3.59% within the driving corridor, surpassing existing state-of-the-art methods in the driving corridor.