VA-Net: 3D Object Detection with 4D Radar Based on Self-Attention
Yaping Zou, Jianlang Li, Ling Chen, Bo Yang · 2024
Abstract: 4D millimeter-wave radar has all-weather sensing capability and low cost.Existing 3D object detection methods often fail to effectively leverage the correlations among sparse points in 4D radar point clouds, leading to lower detection accuracy.To improve the detection accuracy and enrich the feature encoding of 4D radar point clouds, we introduce an object detection approach, VA-Net, utilizing 4D millimeter-wave radar data, which employs a voxel-based self-attention mechanism. Building upon SECOND, this method employs a self-attention mechanism to comprehensively seize relationships between point clouds for more accurate feature representation. The proposed approach is evaluated on the publicly available 4D millimeter-wave radar dataset VoD. Experimental results demonstrate an improvement in detection accuracy compared to other point cloud-based object detection approach. Specifically, when compared with the baseline method SECOND, the detection accuracy of car, pedestrian, and cyclist increased by 2.75%, 2.5%, and 2.6%, respectively, confirming the efficiency of the VA-Net.