CoDPN: An Unsupervised Collaborative Dual-path Network For Contactless Remote Physiological Measurement.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Remote photoplethysmography (rPPG) offers contactless physiological measurement for health monitoring.
• Supervised methods require costly labeled data, while unsupervised methods struggle with illumination and motion variations.
• Existing unsupervised contrastive learning methods face limitations in complex real-world scenarios.

Purpose of the Study:

• To develop a novel unsupervised end-to-end framework, CoDPN, for robust physiological measurement using rPPG.
• To overcome the limitations of current supervised and unsupervised rPPG methods.
• To reduce the dependency on labeled datasets for accurate physiological signal extraction.

Main Methods:

• Proposed a dual-path architecture: Short-term Information Extraction Path (SIEP) and Long-term Information Extraction Path (LIEP).
• Implemented a collaborative learning strategy to integrate features from SIEP and LIEP for complementary information exchange.
• Utilized an unsupervised learning strategy leveraging both frequency and time domains of rPPG signals.

Main Results:

• CoDPN demonstrated superior performance in physiological measurements, including heart rate (HR), heart rate variability (HRV), and respiratory frequency (RF).
• The framework achieved state-of-the-art results on benchmark datasets (UBFC-rPPG, PURE, UBFC-phys) under complex conditions.
• Successfully reduced the reliance on labeled data for accurate rPPG signal extraction.

Conclusions:

• CoDPN provides a robust and efficient unsupervised solution for contactless physiological measurement.
• The dual-path architecture and collaborative learning strategy enhance the extraction of vital physiological signals from rPPG.
• This framework holds significant promise for daily health monitoring and clinical applications.