A Machine Learning Method to Improve Non-Contact Heart Rate Monitoring Using RGB Camera

Abstract

Recording and monitoring vital signs is an essential aspect of home-based healthcare. Using contact sensors to record physiological signals can cause discomfort to patients, especially after prolonged use. Hence, remote physiological measurement approaches have attracted considerable attention as they do not require physical contact with the patient’s skin. Several studies proposed techniques to measure Heart Rate (HR) and Heart Rate Variability (HRV) by detecting the Blood Volume Pulse (BVP) from human facial video recordings while the subject is in a resting condition. In this thesis, we focus on the measurement of HR. We adopt an algorithm that uses the Independent Component Analysis (ICA) to separate the source (physiological) signal from noise in the RGB channels of a facial video. We generalize existing methods to support subject movement during video recording. When a subject is moving, the face may be turned away from the camera. We utilize multiple cameras to enable the algorithm to monitor the vital sign continuously, even if the subject leaves the frame or turns away from a subset of the system’s cameras. Furthermore, we improve the accuracy of existing methods by implementing a light equalization scheme to reduce the effect of shadows and unequal facial light on the HR estimation, a machine learning method to select the most accurate channel outputted by the ICA module, and a regression technique to adjust the initial HR estimate. We systematically test our method on eleven subjects using four cameras. The proposed method decreases the RMSE by 27% compared to the state of the art in the rest condition. When the subject is in motion, the proposed method achieves a RMSE of 1.12 bpm using a single camera and RMSE of 0.96 bpm using multiple camera.