mmHands: Gesture Classification Based on Millimeter-Wave Radar

Millimeter-wave radar, with its robustness to lighting conditions and privacy-preserving characteristics, has become a promising sensing technology in the field of gesture recognition. However, due to the inherent sparsity and noise interference of millimeter-wave point cloud data, accurately classifying gestures involving fine hand structures, such as fingers, remains a challenging task. To address this issue, this paper proposes a millimeter-wave gesture classification framework based on deep point cloud learning techniques. Specifically, three-dimensional millimeter-wave point cloud data, including spatial coordinates and velocity information, are first collected. Then, velocity features are utilized to effectively filter out static background points, enabling robust segmentation of the moving hand point clouds. A point cloud upsampling module is introduced to solve the point cloud sparsity issue, reconstructing a dense hand point cloud. Based on the reconstructed data, a hybrid network combining PointNet++ and Transformer is designed to capture both local geometric features and global contextual relationships, supporting the gesture classification task. The hand gesture recognition task is formulated as a multi-class classification problem, outputting discrete gesture category labels. Experimental results on a self-collected millimeter-wave gesture dataset containing four typical gestures demonstrate that the proposed method achieves an overall classification accuracy of approximately 93.7%, maintaining stable performance across varying human–radar distances, thus validating its effectiveness in millimeter-wave gesture recognition tasks.