ST-Phys: Unsupervised Spatio-Temporal Contrastive Remote Physiological Measurement.

Area of Science:

• Biomedical Engineering
• Computer Vision

Background:

• Remote photoplethysmography (rPPG) enables non-contact physiological measurements from facial videos.
• Current rPPG methods often rely on supervised learning with large labeled datasets.
• Existing unsupervised rPPG techniques struggle with spatio-temporal features and environmental challenges like low light and noise.

Purpose of the Study:

• To develop an unsupervised contrast learning approach for robust rPPG signal extraction.
• To address limitations of existing unsupervised methods in handling low-light and noisy environments.
• To improve the utilization of spatio-temporal features for accurate physiological parameter measurement.

Main Methods:

• Proposed ST-Phys, an unsupervised contrast learning framework for rPPG.
• Incorporated a low-light enhancement module, temporal dilated module, and spatial enhanced module.
• Introduced a circular margin loss to attract signals from identical videos and repel from distinct ones.

Main Results:

• ST-Phys demonstrated superior performance compared to state-of-the-art unsupervised rPPG methods across six datasets (RGB and NIR).
• The method shows significant improvements in handling random low-light conditions and long-term dependencies.
• Achieved advantages in parameter reduction and enhanced robustness against noise.

Conclusions:

• ST-Phys offers a promising unsupervised approach for accurate rPPG measurement, particularly in challenging environments.
• The proposed modules and loss function effectively enhance spatio-temporal feature extraction and signal quality.
• This method advances unsupervised rPPG by improving noise robustness and reducing computational parameters.