Var-HR: Noncontact Heart Rate Measurement Using an RGB Camera Based on Adaptive Region Selection With Singular Value Decomposition

Monitoring heart rate (HR) from facial RGB videos possesses several pre- valent challenges related to facial motions and lighting conditions, affecting its reliability and accuracy in realistic scenarios. To this end, this letter suggests a novel technique (called Var-HR) for stably estimating HR from facial videos. The face, being the region of interest (ROI), is divided into small sub-ROIs, and an effective strategy is presented to select the quality sub-ROIs to alleviate the impact of spatial uneven illumination and facial motions. The color signals from the quality sub-ROIs are filtered using the singular value decomposition (SVD) to isolate the pulsating component. The HR signal is further consolidated via RGB to LAB color transformation and a linear combination of RGB signals. The technique showed consistent performance yielding lower mean absolute errors of 0.84, 1.05, 1.65, 1.43, and 4.52 beats per minute over the PURE, UBFC, COHFACE, ASIPL, and MANHOB-HCI datasets, respectively, outperforming the existing methods. The Var-HR showed notable tolerance against the performance influencing factors, making it a reliable approach for HR extraction from videos.