AI-Based Remote Photoplethysmography: Benchmarking on Realistic Datasets.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Computer Vision

Background:

• Remote photoplethysmography (rPPG) enables non-contact heart rate monitoring.
• Existing rPPG models often perform poorly in realistic, unconstrained environments.
• There is a need for robust rPPG datasets that capture naturalistic subject behavior.

Purpose of the Study:

• To evaluate the Contrast-Phys AI model for remote photoplethysmography (rPPG) in realistic conditions.
• To introduce and utilize the novel CAMVISIM_LAB dataset for rPPG research.
• To improve the accuracy of heart rate estimation using AI models.

Main Methods:

• Development and use of the CAMVISIM_LAB dataset, featuring naturalistic participant movement and interaction.
• Evaluation of the Contrast-Phys model on CAMVISIM_LAB, UBFC-RPPG, and PURE datasets.
• Comparative analysis of model training strategies: from scratch, pre-trained, and fine-tuned.

Main Results:

• Initial Contrast-Phys training on CAMVISIM_LAB yielded a Mean Absolute Error (MAE) of 7.7 bpm.
• Using a pre-trained model improved MAE to 5.8 bpm.
• Fine-tuning with pre-trained weights and CAMVISIM_LAB data achieved a state-of-the-art MAE of 1.0 bpm, surpassing results on the PURE dataset.

Conclusions:

• The Contrast-Phys model, when fine-tuned on the CAMVISIM_LAB dataset, demonstrates superior performance in realistic rPPG applications.
• The CAMVISIM_LAB dataset provides a valuable resource for developing and validating robust rPPG algorithms.
• AI-driven rPPG holds significant potential for accurate, non-invasive heart rate monitoring in diverse settings.