An WiFi-Based Human Activity Recognition System Under Multi-source Interference

WiFi-based human activity recognition in simple scenes has made exciting progress driven by deep learning methods, but current applications are focused on recognition without interference. When the channel state information(CSI) matrix of the receiver contains both the features of the target activities and other interference, the neural network often needs a deeper model structure if deep features of the activities are desired. But a deep network model is often difficult to converge, resulting in a decline in accuracy. And the model size is too large to be deployed in the real world. In this study, an ultra-lightweight neural network recognition system with a group communication(GC) named GC-LSTM is proposed. This design can easily convert a large model into a lightweight counterpart and improve network performance under multi-source interference via reducing network size and complexity. The experimental results show that the optimal recognition rate of the proposed method is 98.6% in the classification of four kinds of activities under six different interferences. By further adjusting the parameters, the model size is reduced to 4.1% of that of plain Long Short-Term Memory(LSTM), while the identification accuracy remains at 96.4%.