Core Concepts
Video-based remote photoplethysmography (rPPG) offers a promising non-contact approach for vital sign monitoring, but faces significant challenges in real-world scenarios. This study systematically evaluates the impact of various spatial, temporal, and visual artifacts on rPPG signal quality, and proposes practical mitigation strategies to enhance the reliability and resilience of video-based rPPG systems.
Abstract
This paper provides a comprehensive evaluation of video-based remote photoplethysmography (rPPG) methods in challenging real-world environments. The key highlights are:
Systematic assessment of the impact of spatial factors (facial region resolution, color depth reduction, image degradation), visual occlusions, and temporal variations (frame rate changes, random frame dropping) on the accuracy of rPPG-based heart rate estimation.
Evaluation of both non-learning-based (OMIT, CHROM, POS) and deep learning-based (EfficientPhys, ContrastPhys, PhysFormer, MTTS-CAN) rPPG methods across multiple public datasets.
Proposal and validation of mitigation strategies, including denoising techniques (Non-local Means, NAFNet), occlusion handling methods (occlusion segmentation, GAN-based inpainting), and frame reconstruction approaches (frame rate recalculation, timestamp-based interpolation) to improve rPPG signal quality and heart rate estimation accuracy under challenging conditions.
Comprehensive analysis of the results, highlighting the strengths and limitations of the evaluated rPPG methods, and providing insights for the design and integration of robust remote vital sign monitoring technologies.
The findings demonstrate that while non-learning-based rPPG methods generally outperform deep learning approaches in handling spatial and temporal degradations, deep learning methods like ContrastPhys exhibit greater resilience to noise. The proposed mitigation strategies, such as denoising and occlusion handling, can effectively enhance rPPG performance in real-world scenarios, contributing to the development of more reliable and adaptable remote vital sign monitoring systems.
Stats
Reducing the color bit depth from 8-bit to 6-bit has a minimal impact on heart rate estimation accuracy, suggesting potential for optimization in resource-constrained environments.
Noise has a more pronounced effect on rPPG performance compared to blur, with deep learning methods like ContrastPhys showing greater resilience to noise.
Random frame dropping significantly degrades heart rate estimation, but the proposed mitigation strategies, such as frame rate recalculation and timestamp-based interpolation, can effectively counteract these effects.
Quotes
"Video-based remote photoplethysmography (rPPG) offers a promising non-contact approach for vital sign monitoring, but faces significant challenges in real-world scenarios."
"Addressing these challenges is imperative for advancing and broadening the use of rPPG technology."
"The findings demonstrate that while non-learning-based rPPG methods generally outperform deep learning approaches in handling spatial and temporal degradations, deep learning methods like ContrastPhys exhibit greater resilience to noise."