Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement Finding Heart Rate From Web-cam Images by Application of Independent Component Analysis Michael Mason Colorado School of Mines mimason@mines.edu December 1, 2014 1/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement Overview 1 Introduction 2 Descriprion of Independent Component Analysis The Problem Theory Practice 3 Application for Pulse Measurement 2/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement Remote Measurement of Pulse Goal To be able remotely detect a persons heart rate. Past work The remote detection of heart rate and other physiological indicators has been a subject of interest for some time. This technology is particularly suited for remote health-care applications due to the proliferation of the necessary hardware(a web-cam). This project was inspired by the work of the MIT CSAIL Laboratory and by Poh et al and their use of blind source separation for the afore mentioned task. 3/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement The Problem Theory Practice The Cocktail Problem 4/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement The Problem Theory Practice The Cocktail Problem specifics Same number of microphones as there are guests Could have more microphones than guests Square matrices easier to work with No echoes No time delay in recording 5/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement The Problem Theory Practice The Solution Independent Component Analysis 6/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement The Problem Theory Practice Assumptions Independent sources Justifies use of central limit theorem The sum of independent random variables is generally more Gaussian than any of the individual random variables Sources have non-Gaussian distribution(except 1) Allows ICA to go further than PCA Instantaneous, error-free measurement Not accounted for For Gaussian variables,being uncorrelated is equivalent to being independent If there is one more recording than source, the final source could be thought of as a Gaussian error ”source” 7/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement The Problem Theory Practice Assumptions cont. Recordings represent a linear combination of sources x = As A is an unknown mixing matrix, x contains your measurements at some point in time, s is the value of the sources 8/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement The Problem Theory Practice Blind Source Separation We have x in : x = As But, both s and A are unknown. We want to estimate an de-mixing matrix B: B = A−1 (1) (2) To do this, we pick the rows of B s.t. the signals Bx are maximally non-Gaussian 9/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement The Problem Theory Practice Measuring Non-Gaussianity Kurtosis Kurtosis, denoted γ2 , defined: µi = E[(X − µ)i ] µ4 γ2 = 2 − 3 µ2 (3) is zero for Gaussian distributions. Maximizing kurtosis maximizes non-Gaussianity. This is computationally cheap but very sensitive to outliers in the measured signal. 10/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement The Problem Theory Practice Measuring Non-Gaussianity cont. Entropy Entropy, defined as: X P(X = ai )log2 (P(X = ai )) (4) i is maximal for the Gaussian distribution, so minimizing entropy is equivalent to maximizing non-gaussianity. This is much more robust but it is computationally expensive 11/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement The Problem Theory Practice Limitations Can not determine amplitude(and sign) of signal Sources must be independent Cannot have more than one Gaussian source 12/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement The Problem Theory Practice ICA Algorithm 1 whiten data 2 choose number of independent components to find, m 3 choose initial mixing vector wj 4 optimize w for wjT x to be maximally non-Gaussian ensure wj is orthogonal to all previously found wi (e.g with Gram-Schmidt) at every step 5 repeat steps 3 and 4 until m components have been found 13/20 Michael Mason Finding Heart Rate With ICA Some Signals 14/20 Some “Recordings” 15/20 Comparison with Extracted Signals(right) 16/20 Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement Approach To obtain a pulse rate measurement follow these steps: 1 2 capture a color video of a face sum each color channel over all pixels these 3 sums serve as the recorded signals at that point in time 3 perform ICA on the recorded signals and look for an extracted source which is correlated with pulmonary signals 4 extract pulse from the periodogram of source 17/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement Results Poh et al Able to determine pulse rates very accurately and in real time. 99% accuracy when compared to a finger pulse monitor myself Able to find independent components in real time, but none of them had a strong pulmonary signal. Furthermore, the nature of the ICA algorithm does not guarantee an order to the sources, so there was little consistency between samples. 18/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement References A. Hyvärinen et al, Independent Component Analysis NY, John Wiley and Sons, 2001. Poh , M.-Z., McDuff, D.J., and Picard , R. W. 2010. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express 18, 10, 1076210774. 19/20 Michael Mason Finding Heart Rate With ICA Introduction Descriprion of Independent Component Analysis Application for Pulse Measurement Questions? 20/20 Michael Mason Finding Heart Rate With ICA