Performance limits of ICA-based heart rate identification techniques in imaging photoplethysmography

Abstract

Imaging photoplethysmography is a relatively new technique for extracting biometric information from video images of faces. This is useful in non-invasive monitoring of patients including neonates or the aged, with respect to sudden infant death syndrome, sleep apnoea, pulmonary disease, physical or mental stress and other cardio-vascular conditions. In this paper, we investigate the limits of detection of the heart rate (HR) while reducing the video quality. We compare the performance of three independent component analysis (ICA) methods (JADE, FastICA, RADICAL), autocorrelation with signal conditioning techniques and identify the most robust approach. We discuss sources of increasing error and other limiting conditions in three situations of reduced signal-to-noise ratio: one where the area of the analyzed face is decreased from 100 to 5%, another where the face area is progressively re-sampled down to a single RGB pixel and one where the HR signal is severely reduced with respect to the boundary noise. In most cases, the cardiac pulse rate can be reliably and accurately detected from videos containing only 5% facial area or from a face occupying just 4 pixels or containing only 5% of the facial HR modulation.

Effects of frame rate and image resolution on pulse rate measured using multiple camera imaging photoplethysmography

“…Efforts in the area of image processing have been focused on targeted region of interest (ROI) selection 37 , color manipulations to achieve chrominance-based iPPG 38 , and motion artifact reduction 9,39,40 . In the area of signal processing, methods for detecting individual pulse waves 41,42 and blind source separation (BSS), particularly Independent Component Analysis (ICA) 7,8,43,44 and Principle Component Analysis (PCA) 45 for recovering an iPPG signal. Further efforts have been devoted toward modified collection methods including using an imager with additional color bands (red, green, blue, cyan, orange) to increase channel space inputs for blind source separation 15,46 or using near infrared light sources and a near infrared imager to successfully collect iPPG signals when highly variable, visible lighting conditions exist 47 .…”

A Survey on Computer Vision for Assistive Medical Diagnosis From Faces

“…pulse). The study of cardio-vascular pulse waves traveling through the body and the monitoring of blood flow and its velocity, provide indications in the diagnostic of vascular disease [41], but also can help in the monitoring of conditions related to sudden infant death syndrome, sleep apnoea or pulmonary disease [42]. While different methods have been developed in the past to detect pulse [43], alternative non-contact solutions based on photoplethysmography or thermal imaging, able to capture the pulsatile peripheral blood flow, have been promoted.…”

Reliability of Pulse Measurements in Videoplethysmography

“…FastICA, RADICAL) gave better results than other (JADE, Robust ICA) [4,5]. However, in [5] the authors showed that sometimes there were problems with ICA algorithms, for example, FastICA failed to properly initialize. It was also shown that using a camera embedded in a laptop computer the use of green trace, bypassing the ICA stage, gave better results.…”

“…The results showed that some algorithms (i.e. FastICA, RADICAL) gave better results than other (JADE, Robust ICA) [4,5]. However, in [5] the authors showed that sometimes there were problems with ICA algorithms, for example, FastICA failed to properly initialize.…”