Facial Video-Based Remote Physiological Measurement via Self-Supervised Learning.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Remote physiological measurement estimates vital signs from facial videos using remote photoplethysmography (rPPG).
• Current deep learning methods require extensive annotated data (facial videos with synchronous photoplethysmography signals), which is difficult to collect.
• There is a need for unsupervised or self-supervised methods to overcome data acquisition challenges.

Purpose of the Study:

• To develop a novel self-supervised framework for estimating rPPG signals from facial videos.
• To enable vital sign measurement (heart rate, respiration frequency) without requiring ground truth photoplethysmography signals.
• To improve the efficiency and accessibility of remote physiological monitoring.

Main Methods:

• A frequency-inspired self-supervised learning framework is proposed.
• Video samples are augmented into positive and negative samples with similar/dissimilar frequencies.
• A learnable frequency augmentation module and a local rPPG expert aggregation module are utilized.
• Frequency-inspired losses (contrastive, ratio consistency, cross-video agreement) are introduced for optimization.

Main Results:

• The proposed method successfully estimates rPPG signals without ground truth photoplethysmography data.
• Significant improvements were observed in heart rate, heart rate variability, and respiration frequency estimation across five benchmarks.
• The framework demonstrates state-of-the-art performance, outperforming existing methods by a considerable margin.

Conclusions:

• The novel self-supervised framework effectively learns rPPG signals from facial videos.
• This approach removes the dependency on synchronously recorded photoplethysmography signals, simplifying data collection.
• The method offers a promising solution for accurate and accessible remote vital sign monitoring.