Multichannel Radar-Based Heartbeat Sensing With Deep Learning in a Monitoring Area

Abstract

This article presents a deep-learning-based approach for robust noncontact heartbeat sensing using multichannel frequency-modulated continuous-wave (FMCW) radar. The conventional methods typically rely on the phase signal of a single predefined or strongest range bin, which makes them vulnerable to subject displacement, multipath propagation, and interference from respiration or body motion. To overcome these limitations, we propose an end-to-end two-stage convolutional neural network (CNN) that processes a radar tensor spanning the monitoring area, enabling the adaptive integration of informative channels while suppressing noise. The proposed architecture, which comprises a feature extraction stage and a feature fusion stage, directly maps radar signals to synchronized photoplethysmography (PPG) waveforms. Validation experiments are conducted using a 79-GHz radar system equipped with array antennas, whereby data are collected from 50 participants in standing and seated postures at three different distances between 0.5 and 0.9 m, producing 300 recordings of 30 s each. The results demonstrate that under the evaluated conditions, the proposed method substantially improves accuracy and robustness compared with single-bin and strongest-bin baselines, particularly for standing postures. Although the evaluation data included minor body-motion variations, further studies involving larger body movements are warranted to confirm the method’s robustness in more dynamic environments.