Advancing Generalizable Remote Physiological Measurement Through the Integration of Explicit and Implicit Prior...

Area of Science:

• Physiological signal processing
• Computer vision
• Biomedical engineering

Background:

• Remote photoplethysmography (rPPG) captures physiological signals from facial videos, with applications in healthcare and biometrics.
• Current rPPG research focuses on cross-dataset generalization, but existing methods lack effective prior knowledge integration.
• This limitation hinders robust performance across diverse data sources like varying cameras, lighting, skin types, and motion.

Purpose of the Study:

• To develop a novel framework for rPPG that incorporates explicit and implicit prior knowledge.
• To enhance the domain generalization capabilities of rPPG methods.
• To address the limitations of current approaches in cross-dataset rPPG evaluation.

Main Methods:

• Systematic analysis of noise sources (camera, lighting, skin, motion) across domains.
• Embedding analyzed prior knowledge into network architecture design.
• Utilizing a two-branch network to disentangle physiological features from noise via implicit label correlation.

Main Results:

• The proposed framework significantly outperforms state-of-the-art methods in RGB cross-dataset evaluations.
• Demonstrated strong generalization capabilities from RGB to Near-Infrared (NIR) datasets.
• The approach effectively addresses noise variations and improves rPPG accuracy.

Conclusions:

• Integrating explicit and implicit prior knowledge is crucial for robust rPPG domain generalization.
• The novel framework offers a significant advancement in cross-dataset rPPG performance.
• This work paves the way for more reliable rPPG applications in real-world, diverse environments.