Towards WiFi-based Real-time Sensing Model Deployed on Low-power Devices

WiFi-based human activity recognition (HAR) is becoming popular due to its non-invasiveness and far-reaching applications such as security surveillance and health monitoring. Existing deep learning methods, while promising, consist of extremely large-scale neural network architectures with massive model parameters and high computational complexity. Thereby, real-time monitoring can be achieved only by leveraging power-hungry GPUs, making it difficult to recognize the performed activities timely. Towards this end, we propose LiST, a real-time and effective WiFi-based HAR model capable of quickly inferring an accurate human activity on a CPU, even on an embedded system. Specifically, a two-stream spatial-temporal structure is utilized to capture sub-carrier and time information from 2D data converted from 3D WiFi Channel State Information (CSI) readings, and uses the dilated convolution to expand the convolution perception range without increasing the number of parameters. To improve the model's generalization ability, we introduce a label smoothing technique to soft labels, enhancing the intra-class distance and separating the inter-class relationship. Extensive experiments demonstrate that LiST obtains excellent performance of over 95.80% with lower model parameters in three scenes and enables to infer activity in about 0.37 s on the Raspberry Pi 4, outperforming the state-of-the-art solutions.