CPulse: Heart Rate Estimation From RGB Videos Under Realistic Conditions
Arya Deo Mehta, Hemant Sharma
Abstract
The challenges obscuring contactless heart rate (HR) measurement using an RGB camera to emerge as a reliable alternative to the existing contact-based methods are the reliability and inconsistencies of performance in natural conditions resulting from illumination variations, varying skin tone, and motion artifacts. To address these problems, this paper proposes a robust end-to-end automated framework (called CPulse) for HR measurement using facial RGB videos. Face being the region-of-interest (ROI) is split into a number of small sub-ROIs of uniform size and a pool of quality sub-ROIs is created to deal with spatial uneven illumination variation and motion artifacts. The color signals from each sub-ROI undergo homomorphic filtering to mitigate the effect of temporal illumination variation. Two distinct color channels, namely green (G) and chrominance (A) from RGB and LAB color spaces, respectively, are selected and processed using the empirical wavelet transform followed by principal component analysis to obtain a pulsating signal describing the volumetric changes in the blood for HR computation. The experimental analysis is carried out over several recordings from 3 publicly available datasets (PURE, UBFC, and COHFACE) and our own ASIPL dataset. The framework achieved the mean absolute error of 0.98 beats-per-minute (bpm), 1.06 bpm, 1.99 bpm, and 1.81 bpm for PURE, UBFC, COHFACE, and ASIPL datasets, respectively, depicting the HR measurement accuracy well above the threshold of clinical acceptance and outperforming the state-of-art methods used for comparisons. Notably, the experimental analysis reveals substantial tolerance of the technique against varying skin tones which is a critical performance influencing factor. The performance results advocate the suitability of the framework in practical situations.