WiFi-Based Indoor Human Activity Recognition Using Bidirectional Time Convolutional Networks

Human activity recognition (HAR) has been an important aspect of research in the field of artificial intelligence, with a wide range of applications. While methods utilizing vision and sensors provide more valuable data, they also have constraints such as high costs, reliance on illumination, and privacy concerns. The ubiquitous presence of WiFi has sparked the interest of researchers, and WiFi-based sensing approaches have been widely used for human activity recognition due to their low cost, ubiquity, and versatility. In this study, we introduce a hybrid deep learning model named CNN-BiTCN-AttNet, which can automatically extract spatiotemporal dual features from time-series data like CSI, and effectively classify human activities. We conduct experiments using ten-fold cross-validation on a self-collected dataset and a widely recognized public datasets (CSI-HAR), and comparative analyses are carried out with five conventional deep learning models (RNN, LSTM, BiLSTM, GRU, and TCN). Experimental results indicate that the CNN-BiTCN-AttNet model achieves an average recognition rate of 97.33% on our self-collected dataset, over 95% in Line of Sight (LOS) conditions on public datasets.