Diffusion-Phys: noise-robust heart rate estimation from facial videos via diffusion models.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Remote photoplethysmography (rPPG) enables non-contact physiological measurement from facial videos for health and emotional analysis.
• Noise in facial videos significantly limits the accuracy and generalizability of existing rPPG methods.
• Robust heart rate (HR) estimation is crucial for reliable remote health monitoring.

Purpose of the Study:

• To introduce Diffusion-Phys, a novel framework for robust heart rate estimation from facial videos using diffusion models.
• To address the challenge of noise in rPPG signals through a diffusion-based approach.
• To improve the reliability and generalizability of non-contact physiological measurement systems.

Main Methods:

• Developed Diffusion-Phys, a framework leveraging diffusion models for HR estimation.
• Utilized Multi-scale Spatial-Temporal Maps (MSTmaps) for data preprocessing.
• Simulated real-world noise by introducing Gaussian noise during training.
• Trained a denoising network for accurate HR estimation.

Main Results:

• Diffusion-Phys demonstrated comparable or superior performance to state-of-the-art methods on VIPL-HR, UBFC-rPPG, and PURE datasets.
• The proposed method achieved lower computational complexity compared to existing approaches.
• Explicitly addressing noise via diffusion modeling proved effective in enhancing rPPG reliability.

Conclusions:

• Diffusion-Phys offers a robust and computationally efficient solution for heart rate estimation from facial videos.
• The framework's ability to handle noise significantly improves the generalizability of non-contact physiological monitoring.
• This work highlights the potential of diffusion models in advancing remote health monitoring technologies.