WarnFi: non-invasive wifi-based abnormal activity sensing using non-parametric model

Abnormal activity sensing has attracted increasing research attention in military surveillance, patient monitoring, and health care of children and elderly, etc. Researchers have exploited the characteristics of wireless signals to sense 'keystrokes' and 'human talks', relieving the privacy invasion concern caused by mounting the surveillance cameras or wearing the smart devices. However, existing technologies usually require some specialized hardware, and can only sense a fixed set of pre-defined activities through a supervised learning from those wireless signals patterns. In this paper, we propose WarnFi, a non-invasive abnormal activity sensing system with only two commodity off-the-shelf (COTS) WiFi devices. The intuition of WarnFi is that whenever the human body occludes the wireless signal transmitting from the access point to the receiver, the time-series of Channel State Information (CSI) will experience a unique variation. By using a non-parametric model, WarnFi can dynamically cluster the human body activities for abnormal sensing. Extensive experiments in various scenarios demonstrate the satisfactory performance of WarnFi.