Enhanced DeepPhys: Leveraging Deep Learning for Heart Rate Detection from Facial Videos

Abstract

The recent advancements in computer vision have made remote photoplethysmography (rPPG) an appealing method for monitoring heart rate (HR), providing a more comfortable option compared to traditional HR measurement techniques. The contactless nature of rPPG is expected to gain increasing importance in the future. However, the accuracy of rPPG is often affected by head movement and changes in lighting conditions. To tackle these challenges, we have introduced an enhanced version of the DeepPhys model, a new deep-learning architecture aimed at improving accuracy. Our approach involves two main steps. Firstly, we use MediaPipe FaceMesh to detect and monitor the region of interest (ROI) on the face. Secondly, we apply our model to extract the rPPG signal. To validate the effectiveness of our model, we conducted experiments using two public datasets: the PURE dataset and the UBFC-Phys dataset. Our model has demonstrated outstanding performance, achieving a mean absolute error (MAE) of 0.99, a root mean square error (RMSE) of 1.45, and a Pearson's correlation coefficient (R) of 0.997 on the PURE dataset.