Spectrum-aware Human-Centric Sensing (HCS) using mmWave radars

This thesis explores the field of Human-Centric Sensing (HCS) using Millimeter Wave (mmWave) Frequency-Modulated Continuous Wave (FMCW) radars, proposing methods based upon Artificial Intelligence (AI) and Machine Learning (ML) tailored for data acquired from the radars. The mmWave FMCW-based HCS applications are divided into three categories of coarse-grained (such as localization), medium-grained (such as gesturerecognition for Human-Computer Interaction (HCI)), and fine-grained (such as water quality estimation) based on their requirements which influence the approaches needed for them. A set of novel approaches are proposed for each, based on their requirements. The performance of these approaches is assessed through extensive experimental evaluations, providing valuable insights into the potential of mmWave radar technologies for real-world HCS applications. In the field of fine-grained sensing, a neural-network-based model is proposed for water quality estimation, which can identify the source of water with 100% accuracy and detect metallic contaminants even at low concentrations. In medium-grained sensing, the thesis presents extensive research on utilizing mmWave FMCW radars for gesture recognition. Three models are proposed for gesture recognition using constrained devices without the need for GPU. Among them, the proposed Tesla-Rapture system achieves an accuracy of 97% on a set of 21 classes of gestures, pushing the limits of medium-grained HCI systems in terms of both accuracy and computational complexity. The thesis further extends the work to multiple radars for recognizing gestures in occluded and cluttered environments, and shows that the system can still be used with reasonable accuracy for zero-shot gesture-recognition. Moreover, the thesis presents research on coarse-grained applications such as multi-people localization, where a clustering-based approach and a neural-network-based approach are proposed and compared to state-of-theart technologies in the field of localization. The radar-based approach can achieve an average error of 3 cm, outperforming the mechanism based on Large Intelligent Surface (LIS). Given the congestion in the Radio Frequency (RF) spectrum, the thesis proposes an RF sensing approach integrated into the 5G standard using the sidelink mechanism. Furthermore, an approach based on Reinforcement Learning (RL) is proposed to efficiently utilize the unused part of the spectrum for communication. In conclusion, this thesis provides a comprehensive investigation of the potential of mmWave radars for HCS proposing new approaches based on AI and ML for each category of applications. The proposed methods can serve as a valuable resource for researchers in these fields, and the study's findings can assist in the development of new and improved technologies for real-world applications.