Skin Segmentation Using Active Contours and Gaussian Mixture Models for Heart Rate Detection in Videos

Alexander Woyczyk, Vincent Fleischhauer, Sebastian Zaunseder; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 312-313

Current research focuses on non-contact means to capture physiological signals like the heart rate. One promising approach uses videos (imaging PPG, iPPG). The common procedure to derive the heart rate by iPPG comprises three steps: segmentation of a region of interest, usage of colour information from that region to yield a pulse signal and analysis of that signal to estimate the heart rate. This contribution proposes a novel approach to yield a region of interest using a Gaussian mixture model based level set formulation. The proposed method aims to segment a homogeneous region on an individual basis. To that end, we model the probability distributions for the pixel skin and non-skin class by two separate Gaussian mixture models. The proportion of the posterior probabilities are then included in the formulation of the level set function. The procedure yields a region of interest, which is used to derive a pulse signal from its average intensity or additional processing steps. We tested the method on own data and data of the 1st Challenge on Remote Physiological Signal Sensing. It is shown that the proposed method can improve the results for heart rate estimation on moving subjects. The potential of our approach is underlined by the promising result in the challenge.