Title:Leveraging Multiple Legacy Wi-Fi Links for Human Behavior Sensing

Abstract:Taking advantage of the rich information provided by Wi-Fi measurement setups, Wi-Fi-based human behavior sensing leveraging Channel State Information (CSI) measurements has received a lot of research attention in recent years. The CSI-based human sensing algorithms typically either rely on an explicit channel propagation model or, more recently, adopt machine learning so as to robustify feature extraction. In most related work, the considered CSI is extracted from a single dedicated Access Point (AP) communication setup. In this paper, we consider a more realistic setting where a legacy network of multiple APs is already deployed for communications purposes and leveraged for sensing benefits using machine learning. The use of legacy network presents challenges and opportunities as many Wi-Fi links can present with richer yet unequally useful data sets. In order to break the curse of dimensionality associated with training over a too large dimensional CSI, we propose a link selection mechanism based on Reinforcement Learning (RL) which allows for dimension reduction while preserving the data that is most relevant for human behavior sensing. The method is based on a sequential state decision-making process in which the CSI is modeled as a part of the state. From actual experiment results, our method is shown to perform better than state-of-the-art approaches in a scenario with multiple available Wi-Fi links.