Publication
Wireless communication systems constantly estimate and equalize the channel to ensure reliable information transmission. The channel estimation not only facilitates effective communication, but also provides an opportunity to infer the physical environment. By analyzing changes in the channel estimate, it becomes possible to uncover corresponding physical changes in the surrounding environment. This link between the channel variations and environmental changes offers opportunities for an entirely new class of applications: wireless sensing. Wireless sensing leverages the interaction between pervasive wireless signals, such as WiFi, ultra-wideband (UWB), Bluetooth, or cellular signals, and the environment to extract meaningful information, such as motion, presence, gestures, or even vital signs, without requiring physical contact. This thesis explores various wireless sensing applications, including single- and multi-target breathing rate estimation, device positioning, device-free movement tracking, gesture recognition, and human presence detection, using omnipresent communication signal sources such as WiFi, UWB, and 4G long-term evolution (4G-LTE).
While wireless sensing has made significant progress over the past decade, many challenges and unresolved issues persist. One fundamental question that appears across the board is whether the frequency domain or delay domain representation of the channel estimate is more suitable for wireless sensing applications. Additionally, the lack of synchronization among sensing devices introduces distortions in the channel estimates, leading to detection errors such as misinterpreting timing offsets and phase shifts as variations in the channel caused by the sensing target. To simplify and improve the efficiency of sensing models, it is crucial to preprocess channel estimates to extract relevant features. However, given various sensing applications, the question of which features are optimal for specific sensing tasks remains largely unaddressed.
In this thesis, we explore these fundamental yet underexamined questions. First, we compare the frequency domain channel frequency response (CFR) and the delay domain channel impulse response (CIR), analyzing their suitability for different sensing tasks. We also propose calibration methods tailored to different channel estimate representations and communication technologies, as well as dimension reduction techniques to enhance sensing performance by selecting relevant components or combining multiple components of channel estimates. To address the complexity of physical propagation channels, we develop learning-based approaches to map channel estimates to sensing outcomes. Additionally, we tackle the challenges of feature extraction and selection, emphasizing the importance of leveraging prior knowledge of the physical channel to pre-select features rather than directly inputting raw channel measurements into learning models. Through these contributions, this thes