Heart Rate Estimation from Facial Videos Using 2D Convolution Neural Network

Abstract

The severity of heart disease, one of the main causes of death worldwide, makes early detection and continuous monitoring of heart rate (HR) vitally important. Traditionally, heart rate is measured using contact devices which are not preferred in case of skin irritation or disease transmission. In this paper, a contactless deep neural network-based heart rate measurement system is proposed utilizing the relationship between facial image color intensity variations and the Blood Volume Pulse (BVP) signal. The proposed approach locates two regions of interest (ROI), the cheeks and forehead, and a 2D convolution neural network (CNN) is used to extract a vector of 2720 features from each region. The extracted features are added and fed into a set of fully connected layers to predict heart rate. The proposed model is trained on unprocessed frames without the necessity for data preprocessing to generalize it across other datasets. It outperforms state-of-the-art models on the Pulse Rate Detection (PURE) benchmark dataset in terms of the mean absolute error (MAE) of the predicted HR and achieves an average prediction time of 33 milliseconds, making it practical for real-time applications.