Method of Remote Photoplethysmography Robust to Interference in Video Registration of Human Facial Skin

Abstract

Monitoring of heart rate (HR) and its dynamics is necessary in ambulatory and telemedicine for diagnosis and treatment of diseases. Remote photoplethysmography (rPPG) allows for HR and its dynamics monitoring by video analysis of facial skin blood flow, and is ofparticular importance for patients with delicate skin such as infants, the elderly, or those with severe burn injuries. Unlike other contactless HR measurement methods, rPPG does not require special hardware, but allows to monitor HR on the basis of analyzing a sequence of video images of a person's face. rPPG involves a multi-step process including signal extraction, dimensionality reduction to estimate the photoplethysmographic (PPG) signal, and subsequent HR estimation. However, the presence of high-amplitude spikes due to subject movement, facial expressions, lighting fluctuations, video compression artifacts, ROI tracking errors, among others, can interfere the useful PPG signal, leading to inaccurate HR estimations. A method has been developed that allows to increase the accuracy of HR estimation in the rPPG problem due to its robustness to interferences inevitable during video recording. The proposed approach tackles the rPPG challenge by applying a pre-processing approximation of the signal finite difference using a single-layer neural network with radial basis function (RBF) inner layer. Transitioning to the signal finite difference helps reduce the amplitudes of irrelevant low-frequency peaks within the HR search range, thus avoiding their masking effect on the HR-related spectral peaks. The neural network's RBF approximation further diminishes irrelevant high-frequency spectral peaks when the number of RBF nodes is less than half the signal sample count. The correctness of the solutions is confirmed by numerical experiments carried out on the Mahnob-HCI public database.