Estimate Heart Rate from Recovered Facial Signal

Heart rate is an important physiological indicator of the human body, reflecting a person's health and mental state. Remote photoplethysmography (rPPG) is a noncontact method for measuring cardiac signals from facial videos. This method uses facial video to analyze the subtle and instantaneous changes in skin color caused by the heartbeat that cannot be detected by the human eye and realizes the detection of physiological indicators such as heart rate, respiration rate, and heart rate variability. RPPG signals (also known as "blood volume pulse signals" or "BVP" signals) are the signals extracted by rPPG techniques. Heart rate can be obtained through BVP signals extracted by rPPG techniques. However, the quality of the signal is affected to varying degrees due to effects such as lighting and motion, so the accuracy of remotely measured heart rate is also affected. The existing methods rarely perform signal recovery processes for measuring heart rate, usually estimating heart rate by complex models. In this paper, we propose GBR-HR, a framework based on Gradient Boosting Regression (GBR) and Convolutional Neural Networks (CNN) for remote heart rate estimation from facial videos. The results show that the proposed method can effectively improve the quality of the BVP signal compared with the chrominance-based color space projection decomposition algorithm (CHROM), and the Pearson correlation coefficient is improved by an average of 24.5% on the UBFC-rPPG and UJN-rPPG datasets. For heart rate estimation, results demonstrate that GBR-HR outperforms most of the existing methods on the UBFC-rPPG dataset. This simple framework can quickly and accurately restore the BVP signal, which makes physiological indices such as heart rate estimation even more accurate.