1. CVPR 2008 June 24 – 26, 2008 Infrared camera:  Mid wave: 3.0-5.0 microns  Resolution: 640*512 pixels with 14bits  Frame rate: 30/60/115 fps  Sensitivity: about 25mK To explore contact-free heart rate and respiratory rate detection through measuring infrared light modulation emitted near superficial blood vessels or a nasal area. Ming Yang 2, Qiong Liu 1, Thea Turner 1, Ying Wu 2 1 FX Palo Alto Laboratory, Inc., 3400 Hillview Ave., Palo Alto, CA 94304 2 Dept. of EECS, Northwestern Univ., 2145 Sheridan Rd., Evanston, IL 60208 Vital Sign Estimation from Passive Thermal Video Experiments Overview of our approach Ground truth: ADI PowerLab 4/30 Test dataset: Age 20-60, F:8 and M:12  20 subjects for heart rate estimation  7 subjects for respiratory rate estimation  Accurate subject alignment for temporal signal extraction, e.g. involuntary muscular movements are inevitable.  Robust harmonic analysis with low signal-to-noise ratio (SNR) temperature modulation signal, e.g. modulation magnitude 0.1K vs. camera sensitivity 0.025K.  A novel contact-free vital sign measurement method.  Low risk of harm & convenience for quick deployment.  Potential applications: airport heath screening, long- term elder care, workplace preventive care, etc.  N. Sun, M.Garbey, A. Merla, I. Pavlidis. Imaging the cardiovascular pulse. CVPR 2005. (S)  S.Y. Chekmenev, A.A. Farag, E.A. Essock. Multiresolution approach for non-contact measurements of arterial pulse using thermal imaging. CVPR 2006 Workshop. Goal Motivations Challenges Pioneering work Automatic ROI segmentation and alignmentSignal enhancement and outlier removalRobust harmonic analysis  Region-of-interests segmentation by thresholding the isotherms and alignment by contour tracking.  Signal enhancement using a non-linear filter, and outlier removal by pixels-of-interests clustering.  Robust harmonic analysis by dominant frequency voting.  Perform N-point (N=1024/2048/4096) FFT of all temperature signals of all pixels using a sliding window:  Non-linear filtering by taking the point-by-point minimum of a rectangle window W r (t) and a Hamming window W h (t)  Cluster H(x j, f ) in the band of interest (40-100 bpm for heart rates, and 6-30 bpm for respiratory rates) using K-means, then select the largest cluster to estimate. Subject #fps# of framesGT bpmEst. bpmDiff. 46030001815.8-2.2 76030001715.1-1.9 1011550001111.8+0.8 1111550001716.8-0.2 1411550001613.9-2.1 1511550001513.1-1.9 1711550002018.5-1.5 1911550001615.2-0.8  Segment the initial ROI by selecting the isotherm with the sharpest gradient.  Align the ROI by tracking the contour  Extract the temporal signals for individual pixels inside the ROI and denote by  Respiratory rate estimation results  Heart rate estimation results  Point-by-point comparisons Subject #fps# of frames GT bpmEst. bpmDiff.RMSE 130200065.365.8+0.51.9 230200066.663.9-2.73.9 330175065.764.73.3 460300059.860.7+0.92.5 560350060.760.3-0.43.3 660250066.353.0-3.33.9 760300061.160.9-0.22.3 8115500064.065.0+1.03.8 9115500078.980.1+1.21.9 10115500065.264.4-0.81.7 11115500062.866.2+3.44.2 12115500063.562.4-1.13.2 13115500073.372.6-0.71.8 14115500086.687.9+1.34.9 15115500078.776.5-2.23.1 16115500075.374.7-0.71.9 17115500083.183.2+0.12.1 18115500067.268.21.3 19115500067.669.3+1.72.8 20115500068.770.1+1.42.9  The initial ROI segmentation results  Insensitive to initialization and robust to gentle subject movement and facial expressions.  More stable estimation results compared with the state-of-the-art methods. Conclusions