mmWave-Based Relay Reflector Reconstruction for LiDAR-Free Around-Corner Human Sensing

mmWave radar enables mobile platforms (e.g., mobile robots) to perceive conditions of around-corner human via multi-bounce signal reflection through relay reflector. Existing methods rely on LiDAR to reconstruct relay reflector geometry, which helps map mirror radar point cloud back to its actual position. However, LiDAR is not universally available and is unable to reconstruct transparent relay reflectors. This paper presents mmRec, an mmWave-based relay reflector reconstruction method which avoids LiDAR assistance for the first time. LiDAR absence poses two challenges: generating dense radar point cloud for recovering relay reflector geometry, and separating radar point clouds of human and relay reflector solely based on radar point cloud features. mmRec addresses such challenges with two key designs, including (i) a point cloud generation component that generates dense point cloud of relay reflector by leveraging diffuse reflection and performing peak detection in the range-angle domain, and (ii) a semantic segmentation component that performs accurate segmentation of radar point cloud via feature augmentation, and neural network design which enables learnable graph construction and multi-scale feature fusion. Extensive experiments show that mmRec-reconstructed relay reflectors achieve an average post-mapping location error of 6.25cm.