WiMUAR: Wi-Fi-Based Multi-User Human Activity Recognition Using Hybrid Spatio-Temporal Neural Network

Wi-Fi-based human activity recognition (HAR) technology is a promising human sensing technique for human-computer interaction in various scenarios due to its robust privacy-preserving and non-Line-of-Sight (NLOS) abilities. However, the simultaneous movements of multiple individuals in multi-user HAR (MU-HAR) task result in much more complex fluctuations in Channel State Information (CSI) signals than a single-user HAR (SU-HAR) case, making it extremely challenging to effectively extract the implicit spatio-temporal correlation features of CSI for MU-HAR tasks. To address this issue, we propose WiMUAR model for MU-HAR tasks, which employs an innovative hybrid spatio-temporal neural network (HSTNN) to improve the recognition accuracy of MU-HAR tasks. HSTNN integrates multi-scale dilated convolution downsampling structure (MDC) and Attention-enhanced Bidirectional Gated Recurrent Unit with Multi-branch Perceptron structure(AGM) to effectively extract the spatio-temporal features of CSI fluctuations of human activities. Additionally, to preserve the implicit spatio-temporal correlated features in CSI signals, we introduce CSI scaling and dynamic component separation strategy. Finally, we employ Assisted Online Knowledge Distillation (AOKD) to enhance the generalization of WiMUAR. Experiment results demonstrate that WiMUAR outperforms other algorithm by an accuracy of 69.61%. These results illustrate not only its superior robustness in multi-user scenarios but also its potential to advance Wi-Fi multi-user sensing technologies. We will further explore lightweight model designs and more complex scenarios.