Deep Learning-based Human Activity Recognition using Wi-Fi Signals

Nowadays wireless signals are everywhere facilitating our daily communication. It turns out that they are not only the carrier of information but also an effective tool for sensing and recognition tasks such as Human Activity Recognition (HAR) and gesture recognition. Since wireless channels are extremely sensitive to environmental changes, even a tiny movement can cause signal fluctuation. However, activity-caused signal fluctuation can be buried in all kinds of environmental noises, which challenges wireless-based HAR. Wireless-based HAR have multiple advantages over traditional video-based or sensor-based HAR as it is not limited to line of sight, doesn’t require extra sensing equipment, and maintains better privacy. By utilizing state-of-the-art deep learning algorithms to differentiate the features in the variation of Channel State Information (CSI) of wireless signals, we can precisely identify human activities. In this paper, we design an end-to-end deep learning-based HAR system which contains Wi-Fi CSI preprocessing module, feature extraction module and classification module. Hampel filter and Discrete Wavelet Transform (DWT) preprocess the CSI signal to remove outliners and unwanted noises. Independent Component Analysis (ICA) analyzes subtle changes in WiFi CSI on continuous time series and Bidirectional Long Short-Term Memory (BiLSTM) classifies human activities. Extensive experiments show that the system can achieve an overall accuracy of 88.7%, outperforming comparison methods.