PETer: a lightweight human action recognition network based on millimeter-wave radar sparse point clouds

Millimeter-wave radar, with its excellent environmental adaptability, high resolution, strong penetration capability, and privacy protection, is widely used in smart homes, elderly care, and security monitoring. However, due to the high sparsity of millimeter-wave point cloud data, it carries less information compared to video or LiDAR, posing significant challenges for precise and rapid human action recognition. This paper proposes a lightweight long- sequence spatiotemporal edge convolution Transformer network (PETer). The model first enhances the data, then separates the spatial and temporal information of the point cloud sequence, utilizes EdgeConv to capture the spatial geometric structure information of single-frame point clouds, and employs a Transformer to extract temporal features from the multi-frame spatial information. To validate the effectiveness of the model, this paper uses the IWR1443-ISKmillimeter-wave radar as the detection sensor to capture human actions and generate a 3D point cloud dataset. This dataset covers 12 types of human actions from 7 volunteers, including walking, waving, boxing, and standing. The proposed model achieves an average accuracy of 98.50% on the dataset and 100% accuracy on the MARS human skeleton dataset. Compared to various point cloud networks, the PETer network is the smallest in size, fastest in speed, and highest in accuracy, demonstrating its superior performance in millimeter-wave radar human action recognition tasks.