WiTS: A Wi-Fi-Based Human Action Recognition via Spatio-Temporal Hybrid Neural Network

Human action recognition has many applications in different scenarios. With the advancement of wireless sensing and the widespread deployment of Wi-Fi devices, the perception technology of Wi-Fi channel state information (CSI) has shown great potential. Related studies identified actions by capturing specific attenuation and distortion features caused by human posture on CSI. These methods are less susceptible to the effects of lighting and object occlusion. However, they have yet to adequately extract information within CSI. The challenge of enhancing model performance through the comprehensive utilization of information features within different dimensions remains an imperative area. To address this, a spatio-temporal hybrid neural network model named WiTS is proposed. It integrates the advantages of different neural networks, using CNN to extract spatial features, combining TCN and Bi-LSTM for dual temporal dimension modeling, and incorporating Transformer's global attention mechanism to achieve comprehensive extraction and multi-level fusion of spatio-temporal features. Additionally, this study further optimizes the original WiTS model from three aspects. The Experiment on WiAR and CSIAR datasets show that the model achieves average accuracy rates of 95.75% and 96.71%, respectively, with F1-scores exceeding 96%. The model has only 2.19 million parameters and less than 560 million FLOPs, offering significant advantages in terms of lightweight design, making it suitable for deployment on limited-computing edge terminals while meeting real-time requirements.