Adaptive Gaussian Mixture Model Driven Level Set Segmentation for Remote Pulse Rate Detection.

Area of Science:

• Biomedical Engineering
• Computer Vision
• Signal Processing

Background:

• Remote physiological monitoring is crucial for healthcare.
• Accurate pulse rate extraction from videos remains a challenge.
• Existing methods often struggle with varying conditions and subject movement.

Purpose of the Study:

• To develop a novel and robust method for pulse rate extraction from video.
• To improve the accuracy and reliability of remote photoplethysmography (rPPG).
• To create a time-varying region of interest (ROI) for enhanced signal capture.

Main Methods:

• A novel approach combining level sets with subject-individual Gaussian Mixture Models for ROI segmentation and tracking.
• Utilizing homogeneous skin areas within defined ROI constraints (area, contour length).
• Integration with state-of-the-art signal processing techniques for pulse rate calculation.

Main Results:

• Achieved Mean Average Errors (MAE) of 2.3 bpm (own data), 1.4 bpm (PURE database), and 2.7 bpm (UBFC-rPPG database).
• Outperformed the KLT tracker (MAE: 2.6, 2.6, 4.4 bpm on respective datasets).
• Secured 2nd place in the 1st Challenge on Remote Physiological Signal Sensing (MAE: 7.92 bpm).

Conclusions:

• The proposed method demonstrates high accuracy and applicability for video-based pulse rate extraction.
• The novel ROI segmentation and tracking technique offers significant advantages over existing approaches.
• The method shows potential for optimization and transferability to other segmentation tasks, including photoplethysmography imaging.