Robust Federated Video-based Remote Physiological Measurement for Heterogeneous Multi-source Data

Remote photoplethysmography (rPPG) is a technology that uses facial video to capture physiological signals, namely photoplethysmography (PPG) signals, enabling convenient and low-cost non-contact physiolog ical measurement. Although current rPPG methods have achieved remarkable performance through extensive data training, they often ignore the transmission cost and privacy concerns of rPPG data. Federated learning (FL) pro vides a promising solution to these issues. However, effective federated training remains challenging due to the cross-domain heterogeneity that arises when using multi source rPPG data. To address these challenges, we characterize the heterogeneity of multi-source rPPG data from both the input and output domain perspectives. Motivated by this analysis, we introduce pseudo-labeling techniques and propose a novel FL framework, FedGRC. Specifically, we incorporate automatic gradient regularization calibration to mitigate the impact of input domain discrepancies on model performance. Furthermore, we leverage pseudo labels generated by handcrafted rPPG methods to align the output domain and reduce label inconsistencies across datasets. Our approach is validated across six publicly available datasets, demonstrating significant advantages over other approaches and effectively addressing the challenges of privacy protection and heterogeneity of multisource data.