Radio frequency (RF)-based fall detection systems have been actively studied over the past decade due to their low privacy concerns and convenient non-wearable nature. In particular, WiFi channel state information (CSI)-based solutions offer the addit...
Radio frequency (RF)-based fall detection systems have been actively studied over the past decade due to their low privacy concerns and convenient non-wearable nature. In particular, WiFi channel state information (CSI)-based solutions offer the additional benefit of requiring no specific hardware due to the widespread deployment of WiFi. However, WiFi CSI-based systems typically suffer from severe performance degradation in cross-domain scenarios, non-line-of-sight (NLOS) and/or through-the-wall (TTW) environments. For successful real-world deployment, a fall detection system must be able to provide reasonably good performance even in such poor conditions. This thesis proposes two new WiFi CSI-based fall detection systems: one based on CSI amplitude, called AmFall, and another based on CSI phase, called PhaFall. These systems effectively address the aforementioned challenges.
To cope with NLOS/TTW situations, AmFall selectively uses the subcarriers and principal components containing valuable information. Considering that the continuous wavelet transform (CWT) is suitable for time-frequency analysis of non-stationary signals, such as those generated during falls, AmFall generates a scalogram by applying CWT to the resulting CSI signal with a novel denoising algorithm and extracts the speed information for segmentation. The segmented scalograms are fed into a deep learning-based classifier. This thesis also suggests a simple yet effective dataset augmentation method to generate multiple scalograms from a single CSI observation. AmFall, implemented on commercial WiFi devices, achieves a high fall detection accuracy of 95.28% to 99.67% in TTW and cross-domain environments, outperforming state-of-the-art WiFi-based methods.
PhaFall is designed under the assumption of harsher real-world conditions, specifically targeting fall detection through a thick concrete wall—referred to as through-the-concrete-wall (TTCW) scenarios. To achieve this goal using CSI phase, the thesis proposes a method to extract motion-induced CSI phase information by removing various phase offsets, and then utilizes the cleaned phase data to detect falls using techniques developed in AmFall. As a result, PhaFall achieves up to 93.1% fall detection accuracy in the TTCW environment, where AmFall struggles to perform effectively.