Enhanced Dual-Stream Point Cloud Feature Extraction Network with Mask Improvement for Human Activity Recognition

Millimeter-wave (mmWave) radar technology is widely applied in the field of human health monitoring due to its capability to obtain 3D point cloud sequences for human activity recognition (HAR). The current single-stream models, influenced by the sparsity of point clouds, are unable to capture the multi-scale features of human activities. Therefore, this paper proposes a dual-stream point cloud feature extraction network with mask improvement to address the adverse effects on network pooling operations caused by point cloud completion, and introduces a multi-channel hybrid fusion architecture. A dataset involving 10 different actions using mmWave radar has been established. The experimental results demonstrate that the proposed dual-stream network outperforms existing single-stream networks in recognition accuracy on both proprietary and public datasets. The strategy of temporal feature fusion applied to the dual-stream network results in a 1% improvement in the accuracy of recognizing each action.