Robust Remote Photoplethysmography Estimation With Environmental Noise Disentanglement.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Remote photoplethysmography (rPPG) is gaining traction for applications like fitness tracking and medical monitoring.
• Deep learning methods currently lead in rPPG estimation but often integrate preprocessing, limiting generalization.
• Existing end-to-end networks may struggle with varied input skin regions and environmental noise.

Purpose of the Study:

• To develop a robust and lightweight rPPG estimation network.
• To improve the generalization and accuracy of rPPG estimation across diverse scenarios.
• To disentangle environmental noise from intrinsic rPPG signals for enhanced performance.

Main Methods:

• Designed a lightweight rPPG estimation network utilizing spatiotemporal convolution.
• Proposed Noise-Disentangled DeeprPPG (ND-DeeprPPG) incorporating adversarial canonical correlation analysis.
• Employed background regions for self-supervised noise disentangling.

Main Results:

• ND-DeeprPPG achieved state-of-the-art performance in heart rate estimation.
• Demonstrated significant robustness in cross-skin-region and cross-dataset experiments.
• Showcased improved performance in other rPPG-based tasks.

Conclusions:

• The proposed ND-DeeprPPG network offers a robust and effective solution for rPPG estimation.
• Noise disentanglement is crucial for improving rPPG accuracy and generalization.
• The self-supervised noise disentangling strategy enhances model adaptability.