Unsupervised Evaluation and Selection of ROIs for Remote Photoplethysmography

Abstract

In this paper, we have proposed an unsupervised method for the evaluation and selection of regions of interest (ROIs) in remote photoplethysmography (rPPG). Our approach involves several key steps: (1) face detection and tracking, (2) segmentation of the face into sub-regions designated as ROIs, (3) extraction of pulse signals from each region and computation of property scores based on sliding-window analysis and statistical assessment using signal-to-noise ratio (SNR) and accuracy metrics, and (4) integration of the selected pulse signals to estimate the final pulse signal. We compared our method against traditional techniques such as Green, CHROM, and POS, demonstrating substantial improvements in SNR and accuracy. Our method achieved a minimum SNR of 4.55, approximately 50% higher than the best-performing traditional method (POS, 1.95), with an average SNR of 7.11, outperforming POS and CHROM by 40% and 43%, respectively.In terms of accuracy, our method achieved a minimum of 95.5%, exceeding existing methods by 3.8% to 4.2%.The average accuracy of our method (96.85%) shows a clear improvement over traditional methods (Green, CHROM, POS), enhancing the reliability of heart rate estimation, especially in low SNR environments. The findings underscored our method's potential as a reliable and precise solution for heart rate estimation, especially in low SNR environments, which is critical for remote health monitoring technologies. The results not only highlighted the advancements in rPPG signal extraction compared to traditional methods, but also indicated substantial benefits for applications such as telemedicine, rPPG video compression, etc. where accurate monitoring of vital signs and physiological signal preservation are essential.