mmWave Radar for Human Sensing and Fall Risk Assessment in Healthcare

Abstract

Falls and mobility impairments are leading causes of injury, hospitalization, and loss of independence among older adults. Preventing falls in an aging population requires continuous, objective, and privacy-preserving assessment of fall risk. Camera-based systems face occlusion and privacy constraints; wearables suffer from adherence and placement variability. Millimeter-wave (mmWave) radar is non-contact, illumination-invariant, and inherently privacy-preserving, but practical solution is limited by sparse/noisy point clouds, temporal jitter in pose estimates, a lack of biomedical constraints in learning models, and non-standardized datasets and evaluation. This dissertation delivers an end-to-end pipeline that converts raw mmWave radar into temporally consistent human motion representations and healthcare interpretable biomarkers. First, a lightweight dynamic model suppresses inter-frame jitter while preserving spatial accuracy, yielding 40–60% reductions in instability metrics. Second, mmPose-FK integrates Conv3D+LSTM spatiotemporal features with a differentiable forward kinematics (FK) layer that enforces fixed bone lengths and joint-rotation limits during training, reducing bone-length violations and pose jitter and improving position accuracy. Third, a structured survey systematizes two decades of radar-based fall-detection and risk assessment research by sensing modality, data representation, algorithm, and identifies gaps in practice. Fourth, the pipeline is validated on Sit-to-Stand (STS) and gait movements by collecting a dataset of 78 participants across four environments. Radar-derived features closely match with Kinect and wearables, with high intraclass correlation coefficients (ICCs; ≥ 0.90) and narrow Bland–Altman limits of agreement. These contributions establish a camera-free pathway from radar signals to validated mobility biomarkers suitable for continuous, at-home monitoring. The findings lay a foundation for privacy-preserving fall-risk assessment and rehabilitation support, and motivate future work on multi-sensor fusion, resource-efficient edge inference, and benchmarks.