Dual-Path Model With Fresnel Zone-Based Voting For Human Activity Recognition Using WI-FI

In Wi-Fi sensing, the spatial diversity granted by using multiple access points (AP) can significantly improve the performance of human activity recognition (HAR), due to the importance of the user location and orientation. Existing approaches leverage this fact by assuming channel state information (CSI) data is transmitted from the different APs to a central server for further processing. This approach is impractical due to the communication overhead of constantly transmitting CSI. In this paper, we propose a distributed model architecture, where dual-path light-weight models run on each AP, and the two paths are used for feature extraction along the time and frequency axes, respectively. When gestures take place, model outputs are combined centrally using a Fresnel zone-based voting scheme, making efficient use of spatial diversity to improve performance, while eliminating the costly CSI transmissions. Experiments on two datasets demonstrate the practicality and high accuracy of our approach for HAR.