A Confidence Framework Through Temporal Averaging for Heart Rate Estimation in Video Magnification

Heart rate (HR) is an essential indicator of cardiovascular disorders as well as other internal physiological complications. Normally, HR tends to vary over time with larger variations associated with external factors like stress or exercise. Video Magnification (VM) can remotely measure HR from video data acquired with an RGB/Thermal camera. Incorporating VM into digital health monitoring systems could be essential in supporting and maintaining well-being and health, especially for an aging population. Previous work has explored factors affecting VM performance such as skin tone, body motion, and light conditions as well as different facial regions of interest (ROI). The current work extends a confidence model for VM methods by proposing temporal combination of partially overlapping consecutive windows. Results from 19 subjects show that temporal averaging enhances VM performance and improves overall HR accuracy from 61.1% to 68.4%. HR errors are associated with motion artifacts and/or darker skin tones. The confidence metric is generally proportional to the accuracy of the HR estimate, providing an indication of when VM-based HR measurements are reliable.