Abstract
Remote photoplethysmography (rPPG) is an emerging contactless physiological parameter detection method utilizing cameras, showing great promise as a forefront technology for remote health assessment. While traditional rPPG methods have substantially contributed to the development of affordable camera-based health monitoring systems, their dependence on specific assumptions may lead to performance degradation when confronted with real-time dynamic interferences. The article presents a comprehensive overview of CAP-rPPG, an innovative method for remote physiological measurement through video analysis. This method employs a multi-scale deep learning architecture with a Gaussian pyramid to capture facial features at different scales that are often overlooked in prior work. A channel attention module further emphasizes rPPG-rich channels, mitigating the issue of feature dilution caused by excessive channel depth and enhancing the accuracy of physiological signal extraction from facial videos. The uniqueness of CAP-rPPG lies in its hybrid loss function, adeptly balancing the short-term characteristics, long-term characteristics of the signal and the correlation between the predicted HR and the ground-truth HR. CAP-rPPG demonstrates outstanding robustness under various challenging conditions, such as varying lighting environments and physical motion. It consistently outperforms most state-of-the-art methods on both the UBFC-rPPG and PURE datasets. Its capability to non-invasively capture subtle physiological changes from video data represents a significant leap forward in the realm of remote health monitoring technologies.