UC Santa Barbara

Due to the recent rapid growth of the number of wirelessly-connected devices, wireless signals are everywhere these days. These signals interact with our surroundings, e.g., by reflecting off of the objects/people in the environment, thereby carrying crucial information about them. Extraction of such information then facilitates many applications in surveillance, security, and smart homes, to name but a few.

In this dissertation, we investigate the possibilities of utilizing wireless signals of off-the-shelf devices (e.g., WiFi) to enable various novel sensing applications. In the first part of the dissertation, we develop a tool that converts video footage to WiFi signals. More specifically, given a video footage of a person engaged in some activity, our proposed tool generates/simulates the WiFi signals that would have been measured if a WiFi device was placed near the person in the video. This opens up the door to a wide range of new possibilities in RF sensing. We showcase the importance of our proposed tool in two sample applications. First, we develop a through-wall person identification system that can use the WiFi power measurements when a person is walking behind wall in a WiFi-covered area to determine if this person is the same as the one walking in a given video clip. Second, we show that our proposed tool can greatly reduce the training burden of learning-based sensing systems, by generating a training dataset from the vast available video data, and without the need for the laborious and time consuming process of real training WiFi data collection. As a case study, we train a gym activity classifier, for the first time, solely based on videos of people performing different gym exercises, and without collecting any real training data.

In the second part of the dissertation, we discuss how WiFi signals can be used for nocturnal seizure detection in a robust, fast, and contactless manner, which will provide much better support for epilepsy patients and their caregivers. We develop a new mathematical characterization for the received WiFi signals during different types of nocturnal sleep motions: breathing, normal movements, and seizures. We then propose a WiFi processing pipeline that detects all non-breathing motions and classifies whether they are seizures or not.

Finally, we discuss how to count a stationary (seated) crowd using off-the-shelf WiFi, based on the natural in-place motions that people naturally engage in while seated (also called fidgets), such as changing their pose or crossing their legs. We develop a mathematical model, inspired from Queuing Theory, that relates the fidgeting statistics of a crowd to the number of people in this crowd. Based on this modeling, we propose a WiFi processing pipeline that extracts the fidgets of the crowd and estimates the number of people accordingly. We experimentally validate all of our developed algorithms with several test subjects in several different environments with different levels of clutter.