Synthetic Data for Multi-Parameter Camera-Based Physiological Sensing.

Area of Science:

• Computer Vision
• Biomedical Engineering
• Artificial Intelligence

Background:

• Deep learning algorithms require extensive data for training.
• Camera-based physiological sensing methods are data-intensive but underutilize synthetic data.
• Existing methods lack diverse training datasets, impacting performance across different demographics.

Purpose of the Study:

• To develop and evaluate a high-fidelity synthetic data pipeline for camera-based cardiopulmonary sensing.
• To investigate the impact of synthetic data volume and diversity on the accuracy of physiological measurements.
• To assess the influence of synthetic avatar skin types on model performance.

Main Methods:

• Generation of synthetic videos featuring realistic facial blood flow and breathing patterns.
• Systematic experiments training deep learning models for multi-parameter cardiopulmonary sensing using synthetic data.
• Comparative analysis of model performance with varying numbers of synthetic avatars and different skin tones.

Main Results:

• Increased number of synthetic avatars in training sets led to higher heart and breathing rate measurement accuracy.
• Models trained with synthetic avatars representing darker skin types demonstrated superior overall performance.
• Physiologically-grounded synthetic data significantly enhances camera-based physiological sensing capabilities.

Conclusions:

• Synthetic data is crucial for advancing camera-based physiological sensing, particularly for deep learning applications.
• Diversity in synthetic training data, including representation of darker skin tones, is essential for equitable and robust performance.
• Further development of synthetic data generation techniques is needed to overcome current limitations in the field.