Wi-SensiNet: Through-Wall Human Activity Recognition Based on WiFi Sensing

Abstract With the advancement of Wi-Fi sensing technology, its significant benefits in convenient operation and privacy protection have become apparent, particularly in fields like smart homes, medical monitoring, and security surveillance, where the application prospects of Human Activity Recognition (HAR) technology are increasingly broad. This study focuses on a novel approach to HAR using Wi-Fi Channel State Information (CSI), especially under complex conditions such as Non-Line of Sight (NLoS) paths and through-wall transmissions. Traditionally, most research has concentrated on Line of Sight (LoS) path HAR, sensitive to environmental changes, while the NLoS path signals, especially through-wall signals, present unpredictability due to weak reflections caused by walls. Addressing this issue, we propose Wi-SensiNet, an innovative deep learning-based method that combines the spatial feature extraction capabilities of Convolutional Neural Networks (CNN) with the temporal sequence processing power of Bidirectional Long Short-Term Memory networks (BiLSTM). This method also incorporates an attention mechanism to enhance the accuracy of human activity recognition in complex environments. Wi-SensiNet is specially designed for through-wall settings, effectively handling the complexity of CSI data, and achieving accurate through-wall human activity detection. In our experiments, we collected a through-wall CSI dataset comprising seven common activities, including running, sitting, standing, squatting, falling, punching, and walking, and verified Wi-SensiNet's average accuracy exceeded 99\% on the original test set. These results not only demonstrate the model's robustness and high accuracy in handling HAR tasks in complex environments but also highlight the potential of CNN and BiLSTM working in tandem to enhance performance.