Non-Contact Heart Rate Estimation for Moving Subjects via aDual-Trajectory Hybrid Deep Framework

Contact-free heart-rate monitoring with millimeter-wave Frequency-Modulated Continuous-Wave (FMCW) radar is promising for wearable-free healthcare, but accurate estimation remains difficult when subjects are moving. In dynamic scenarios, gross body motion causes range migration, phase distortion, and Doppler-related interference that can overwhelm the much weaker cardiogenic component. To address this problem, we propose a hybrid locate-then-reconstruct framework for robust heart-rate monitoring. In the first stage, a dual-trajectory fusion module with a customized Kalman filter determines a reliable phase-extraction path and retrieves an unquantized raw phase sequence from the original radar measurements. In the second stage, the recovered signal is processed by a Variational Mode Decomposition-enhanced Convolutional Neural Network--Bidirectional Long Short-Term Memory--Autoformer (VCLA) model to suppress residual motion artifacts and reconstruct a cleaner cardiogenic waveform. Template matching is then applied for final heart-rate estimation. Experiments in both single-person and multi-person scenarios show that the proposed method achieves a mean absolute error below 4 beats per minute and consistently outperforms representative baselines.