Human Action Recognition Using mmWave Radar with Velocity-Aware Graph Networks

This paper presents a human action recognition framework based solely on millimeter-wave (mmWave) radar, addressing the inherent sparsity and noise of radar point clouds. We introduce a 4D skeleton representation that augments 3D joint coordinates with per-joint velocity cues. This representation is processed by a spatio-temporal graph convolutional network (ST-GCN) enhanced with a velocity-aware attention mechanism to improve robustness to radar-induced skeletal instability. We capture human motions using a custom mmWave radar and convert them into 3D skeleton sequences for recognition. Evaluations on a ten-action dataset show that the proposed method achieves an accuracy of 97.84% and a macro F 1 -score of 97.45%. The results indicate consistent cross-subject generalization and stable performance under environmental variations.