Human Activity Recognition with WiFi Channel State Information by Complex-Valued Neural Network

Many WiFi signal based deep learning approaches on human activity recognition (HAR) were proposed by researchers, for the activity of human in the WiFi environment will affect the Received Signal Strength Indicator (RSSI) and wireless Channel State Information (CSI). Conventional methods concentrated on RSSI or amplitude of CSI, which can't fully characterize the dynamics of the human activity hidden in the signal. For utilizing the full information in CSI with both amplitude and phase, we proposed a new HAR algorithm by using complex-value neural network (CVNN) combined with complex-value transformer (CVTransformer). In the proposed scheme, we treat the complex CSI information as a complex valued image, then we can use CVNN extract different activity features from the CSI “image”, and introduce complex attention mechanism by CVTransformer into the CVNN to enhance the capability to recognize key feature of the activity, thus increase the precision of the activity recognition. Experimental result shows that the proposed approach can effectively recognize most of the human activities in the datasets, with a precision of more than 96.24%, outperforms the related approach.