Estimation of Walking Body Center of Mass Velocity by Means of Microwave Radars and Deep Learning.

Abstract

This study introduces an innovative framework combining microwave Doppler radar networks with deep learning to estimate three-dimensional body center of mass (BCoM) velocity. We evaluated the system's ability to derive clinical gait-quality indices, aiming for a privacy-preserving alternative to laboratory-based motion capture. Sixty healthy adults performed treadmill walking at 2, 4, and 6 km/h, including a simulated hemiplegic gait. Three radars captured micro-Doppler spectrograms in anterior-posterior (AP), medio lateral (ML), and cranio-caudal (CC) directions. Recurrent neural networks, optimized via Bayesian techniques, estimated 3D BCoM velocities, which were validated against gold-standard motion capture. Three parameters were derived to assess gait smoothness (LDLJ), symmetry (iHR), and stability (RMS acceleration). The system reconstructed BCoM velocity waveforms with high fidelity and minimal bias. In particular, LDLJ achieved remarkable accuracy in the AP and CC directions, with errors below 6.1%. While medio-lateral and acceleration-derived metrics (particularly RMS) proved more sensitive to estimation noise- peaking at 22.3% error during simulated hemiplegia-the framework successfully preserved key gait-quality patterns at the group level. Integrating radar technology with sequence learning effectively recovers complex BCoM dynamics. Despite inherent challenges in acceleration-based metrics, this approach captures nuanced gait characteristics that go beyond traditional spatiotemporal parameters, maintaining clinical relevance in a non-invasive format. This work represents a significant step toward unobtrusive, markerless gait monitoring. It offers a scalable, privacy-preserving solution for continuous assessment in clinical and home settings, bridging the gap between laboratory research and real-world applications.