Robust ML Model for Human Counting Using Ambient WiFi Traffic from Multiple Sources

Abstract

Although recent literature has demonstrated the ability to use WiFi Channel State Information (CSI) for sensing physical areas, this method has only been conducted in ideal scenarios. CSI based sensing has yet to been achieved using normal user network traffic, and in the presence of multiple transmitters, and these research gaps currently prevent deployment in the real world. This paper shows that the accuracy of counting human occupancy levels using CSI is dependent on the underlying traffic type, and that there is a degradation in performance when the classifier encounters an unseen traffic type. We experimentally demonstrate that using ambient traffic from end user devices has better accuracy outcomes than leveraging WiFi Access Points (AP). This paper further shows that multiple spatially diverse streams of WiFi CSI can be used for sensing to an accuracy of 99%, and we demonstrate that these multiple streams would reduce the accuracy of CSI sensing systems in literature which are trained for one transmission source. Finally, we show that by training an SVM with diverse traffic types, it can become robust to unseen traffic during the evaluation phase.