Lightweight rPPG Framework for Real-Time Physiological Monitoring on AR Devices

Abstract

To address the accuracy–efficiency trade-off that limits current rPPG methods in dynamic environments, we propose rPPGViT, a multi-scale Vision Transformer designed for real-time and energy-efficient physiological monitoring on AR devices. The model reconstructs precise blood-flow rhythm signals from facial videos while maintaining robustness against illumination variation and motion interference. Furthermore, a lightweight re-parameterized variant, Light-rPPGViT, is developed to achieve low-power, low-latency inference suitable for edge deployment on wearable AR platforms. Experimental results show that rPPGViT achieves MAE of 0.27 BPM and 0.65 BPM on the UBFC and MMPD datasets, respectively, with a correlation coefficient $\rho$ of 0.99. Light-rPPGViT attains comparable accuracy while reducing computational cost by more than 65%. These results demonstrate the feasibility of deploying high-accuracy, energy-efficient rPPG models on wearable AR devices, paving the way for next-generation remote healthcare, psychological monitoring, and intelligent human–computer interaction.