Transformer-based methods for remote photoplethysmography in non-contact physiological measurement: A review

Abstract

Traditional contact-based physiological monitoring techniques are limited by their comfort and convenience, making them unsuitable for continuous and unobtrusive monitoring in medical environments and daily healthcare applications. Remote photoplethysmography (rPPG) is a non-contact physiological sensing technique that estimates vital physiological signals, such as heart rate and respiration rate, from subtle skin color variations captured in facial videos. However, the practical application of rPPG remains challenging because the physiological signals extracted from facial regions of interest are weak and can be easily overwhelmed by non-stationary noise, such as changes in ambient illumination and subject head motion, resulting in a low signal-to-noise ratio. Recently, Transformer architectures have demonstrated strong capabilities in global dependency modeling through self-attention mechanism, enabling effective extraction of long-range temporal physiological features while suppressing global interference. These advantages provide a new paradigm to overcome the performance bottlenecks of conventional rPPG methods. This review provides a comprehensive overview of recent progress on Transformer-based rPPG research. Specifically, it covers multiple research directions, including pure Transformer end-to-end model, CNN-Transformer hybrid model, multi-stream information fusion frameworks, lightweight models for edge deployment, self-supervised learning frameworks, and extended applications in physiological monitoring. Additionally, this review discusses current limitations and challenges of Transformer-based rPPG systems and highlights future research opportunities for developing new non-contact physiological monitoring technologies.