Improving Millimeter-Wave Radar Target Recognition Using Self-Attention and Multi-layer Feature Fusion Network

In this paper, we propose an improved method for enhancing millimeter wave radar target recognition through the utilization of self-attention and a multilayer feature fusion network. Traditional millimeter wave radar exhibits limitations in terms of accuracy in elevation angles and resolution in angular dimensions, thereby resulting in restricted efficacy in the detection of complex scenes. 4D millimeter wave radar is capable of providing elevation angle data, which can be utilized to generate 3D point cloud images. The point cloud produced by 4D radar exhibits sparsity, clutter, and noise, resulting in a disparity between the imaging results and the true target geometry. Traditional methods for point cloud recognition exhibit subpar performance. To address this issue, we employ deep learning methodologies to accomplish point cloud segmentation. Our method employs PointNet and a self-attention mechanism to extract global features, while also enhancing feature representation through the integration of a multilayer local-global feature fusion structure. Additionally, we introduces a point cloud denoising algorithm aimed at improving the quality of the input data. We collected a 4D radar target part segmentation dataset, and completed manual annotation. Experiments indicate an effectiveness of our method with overall accuracy of 88.6%, which is suitable for processing 4D radar point cloud.