Title: Facial Video based Physiological Variables Estimation in Dark Environments
Author: Ankit Gupta
Location: Room 0.57 at the University of Madeira and Zoom: https://videoconf-colibri.zoom.us/j/96967272561?pwd=nCpXXoVD9PK4egFdIYpaKAgvQrYBNE.1 Pass: 853030
Date/Time: 27/jun/2024, 9h00
Abstract
Physiological parameter estimations play a significant role in determining an individual’s health status. Among these parameters, heart rate and oxygen saturation have been extensively used for health monitoring during medical checkups, surgery, sleep disorders diagnosis, and intensive care units. The gold standard techniques for estimating these parameters are electrocardiography and photoplethysmography. Both are contact-based techniques and, therefore, can cause discomfort to the subject in scenarios such as prolonged monitoring and sensitive or burnt skin. Thus, remote photoplethysmography was introduced as a non-contact variant of photoplethysmography. It extracts the blood volume pulse signal from the spatiotemporal sequences of the region of interest, followed by heart rate estimation. On the other hand, oxygen saturation estimations are being performed using the ratio-of-ratios method using red and blue channels. Existing non-contact methods were designed for ambient light conditions. A few methods developed for dark environments used infrared cameras, which are expensive, and the resulting spectra have poorer pulsatile strength than visible spectra. Therefore, this thesis investigates the potential of visible spectra for physiological measurements in dark environments (illuminance ≤ 0.5 lux). Specifically, this thesis has three key contributions: first, a novel heart rate estimation method based on undercomplete independent component analysis, which was developed and tested under different real-time conditions, and second, a "Dark-Video" dataset encompassing participants of different ethnicities and finally a novel deep learning architecture for dark image enhancement that was also proposed to facilitate physiological measurements in the above mentioned dark environments (i.e., estimation methods cascaded by image enhancement). Diverse experiments conducted for the performance analysis using critically selected performance metrics not only proved the superiority of the developed methods but also exhibited their potential of being clinically viable. The future direction of this research aims to implement these methods for scenarios such as non-contact sleep monitoring or monitoring during nighttime driving.
Keywords: Blind Source Separation, Blood Volume Pulse, Deep Learning, Non-contact Approaches, Physiological Parameters Estimation.