Location-Free CSI-Based Lightweight Deeping Learning Model for Human Activity Recognition

In recent years, activity recognition based on channel state information(Csi) has a wide range of application scenarios in the field of human-computer interaction(HCI) due to its advantages such as no need to carry special equipment, no privacy disclosure, and no light intensity. Many approaches based on traditional machine learning and deep learning have encountered two challenges. One is that the problem of activity recognition limited to fixed locations and complex backgrounds remains unsolved. The second is that some studies require complex CSI processing, which increases the network parameters, significantly lengthens the recognition time and raises the deployment cost. To this end, this study develops a Wi-Fi-based location-independent lightweight recognition model. We propose a lightweight CSI processing strategy that is able to effectively extract the main relevant features while compressing the model size. We combine 3D convolution with spatio-temporal information selection gates to extract activity-related information from CSI data, and employ knowledge distillation techniques to migrate the features learned from this model to a simple model. Extensive experimental results show that our system outperforms other deep learning models in terms of efficiency, with recognition accuracy up to 98.6%.