TSHNN: Temporal-Spatial Hybrid Neural Network for Cognitive Wireless Human Activity Recognition

Abstract

WiFi-based human activity recognition has gained ever-growing attention in the field of wireless sensor networks. As a promising technology, it has large application potential for smart homes and elderly activity monitoring. Existing WiFi-based activity recognition methods mainly exploit two-dimensional convolutional neural networks, which focus on extracting spatial information. In such WiFi-based methods, however, temporal information still needs to be fully utilized. To overcome the shortcoming, a Temporal-Spatial Hybrid Neural Network, named TSHNN, is proposed in this paper for the effective fusion of temporal and spatial information. Our proposed TSHNN utilizes channel state information (CSI) data from WiFi signals for activity recognition. More specifically, we convert the amplitude information in the CSI data into video data and then feed the data into a 3D neural network to extract temporal and spatial information. We further enhance the model to extract temporal features using a gated recursive unit (GRU) component. Our experiments show that the proposed TSHNN model achieves over 99% accuracy on both the coarse-grained WiAR and fine-grained SignFi datasets with high stability and reliability. Compared with the state-of-the-art methods, the TSHNN model improves accuracy by 0.68% and 0.3% while using real-world WiAR and SignFi datasets, respectively.