rPPG-NDCL: Unsupervised Remote Physiological Measurement Via Noise-Disentangled Contrastive Learning

Remote photoplethysmography (rPPG) utilizes cameras for noninvasive monitoring of vital signs and has attracted widespread attention. Unsupervised methods have become the primary focus of recent studies due to their advantage of not relying on labeled data. However, these methods lack specialized noise reduction designs and often fail when faced with dynamic noise from complex head motion and illumination variations. To address this issue, we propose a novel noise-disentangling unsupervised framework (rPPG-NDCL). First, we introduce a contrastive learning strategy specific for the signal and noise and incorporate pseudo labels as guidance to separately extract the rPPG signal, head motion and illumination noise. Additionally, we design a dependency reduction module to achieve noise disentangling. Experimental results demonstrate that our method outperforms both the state-of-the-art unsupervised methods and most supervised methods. We also train a robust rPPG estimation model using unlabeled video data not specifically created for rPPG, demonstrating the generalizability of our method to real-world scenarios. Our code is available at https://github.com/Tianyang-Dai/rPPG-NDCL.