Deep-HR: Fast heart rate estimation from face video under realistic conditions

This paper presents a novel method for remote heart rate (HR) estimation. Recent studies have proved that blood pumping by the is highly correlated to intense color of face pixels, and surprisingly can be utilized HR Researchers successfully proposed several methods this task, but making it work in realistic situations still challenging problem computer vision community. Furthermore, learning solve such complex task on dataset with very limited annotated samples not reasonable. Consequently, researchers do prefer use deep approaches problem. In paper, we propose simple yet efficient approach benefit advantages Deep Neural Network (DNN) simplifying estimation from representation HR. Inspired previous work, learn component called Front-End (FE) provide discriminative videos, afterward light regression auto-encoder as Back-End (BE) learned map FE Regression informative could efficiently training samples. Beside this, more accurate well low-quality two encoder–decoder networks are trained refine output FE. We also introduce (HR-D) show our conditions. Experimental results HR-D MAHNOB datasets confirm run real-time estimate average better than state-of-the-art ones.