Comprehensive optimization of physiological Doppler radar sensing: Models, demodulation techniques, sedentary state classification & joint communication and sensing

Abstract

Over the past fifty years, the field of radar-based human physiological sensing has advanced markedly in both hardware architectures and signal-processing methodologies. This dissertation begins with an exhaustive survey of Doppler radar configurations, comparing continuous-wave (CW) radars against a variety of modulated waveform implementations. Signal-analysis approaches are systematically classified into “domain-expert” and “data-driven” paradigms, with specific emphasis on their application across diverse physiological monitoring contexts.Building upon this conceptual framework, the central contribution of this dissertation is a rigorous evaluation of demodulation strategies applied to raw CW Doppler returns. Despite the existence of both time-domain and spectral-domain algorithms, a unified performance assessment under identical experimental conditions has been lacking. Accordingly, four time-domain methods, namely, Arctangent Demodulation (AD), Extended Divide-and-Cross-Multiply (EDACM), Modified Divide-and-Cross-Multiply (MDACM), and Linear Demodulation (LD), and two spectral-domain techniques, the Polyphase Basis Discrete Cosine Transform (PB-DCT) and the Quadrature Cosine Transform (QCT), are implemented and benchmarked. Employing a collection of human and mechanical mover datasets captured under controlled motion profiles, each algorithm is evaluated in terms of micro-Doppler sensitivity, phase-wrapping immunity, computational burden, and resilience to variations in signal-to-noise ratio (SNR). Critical insights into method-specific trade-offs are articulated: AD’s performance is contingent on accurate center fitting and DC offset removal; in high-SNR regimes where AD center fitting degrades, EDACM and MDACM maintains rate accuracy; LD provides reliable estimates of torso displacement and velocity when IQ-arc lengths remain minimal; spectral-domain approaches yield enhanced frequency resolution relative to FFT-based methods. Subsequently, the dissertation addresses the challenge of discerning sedentary from non-sedentary activities in extended monitoring scenarios, including inpatient surveillance, autonomous elderly care, and occupational stress assessment. In these contexts, locomotion and extraneous body motion may obscure physiological signatures and trigger false-positive detections. Initially, hand-crafted based on time-series features are assessed for non-sedentary detection. Thereafter, spectral-domain feature extraction replaces traditional time-domain metrics to refine classification boundaries. The efficacy of machine learning classifiers trained on DCT spectrograms is then demonstrated, followed by RNN-based architectures (e.g., long short-term memory and gated recurrent unit networks) applied directly to raw time-series inputs. Cross-validation analyses quantify precision, recall, and F1-score, establishing a comparative framework against handcrafted and spectral-domain ML models. Finally, a novel dual-function modality is presented, wherein a commercial OFDM-based 28 GHz, 52-channel radar communication system is repurposed for simultaneous physiological sensing and data exchange. By engineering sidelobe beam configurations for target detection while preserving main-lobe integrity for high-throughput communication, the system achieves vital-sign monitoring with no more than 35% degradation in link-layer performance. Empirical validation using both mechanical and human subject datasets underscores the potential of integrated radar-communication platforms for future ubiquitous monitoring applications.