Consecutive-frame latent space normal estimation under sparse point clouds for 4D millimeter-wave radar

Aiming at the sparsity of point cloud data and the low accuracy of spatial alignment exhibited by millimeter-wave frequency-modulated continuous-wave (FMCW) radar in outdoor motion scenarios, a lightweight model for spatial alignment was proposed. This method was specifically tailored for point cloud processing across consecutive multi-frames in outdoor motion scenes captured by millimeter-wave radar. Leveraging spatio-temporal graph neural networks (ST-GNNs), it accurately estimated the hidden spatial normals of adjacent multi-frame point clouds, eliminating the need for position sensors. By transforming radar point cloud data from each frame into a unified observation coordinate system, the method facilitated multi-frame fusion of 4D point clouds and ensured precise scene alignment. Experimental results demonstrated that the proposed approach not only accurately assessed the spatial attitude of 4D point clouds but also effectively corrected and fused the coordinates of each point cloud frame. This enabled precise coordinate alignment during motion and vibration. Furthermore, the algorithm significantly enhanced point cloud imaging density, improved image accuracy and readability, and was capable of imaging both static and dynamic targets. It provided robust support for the application of millimeter-wave radar in outdoor motion scenes.