For this purpose, MIT's Eulerian Video Magnification is next to impossible when you have both color changes and large motions (it is better suited to magnifying small motions, or subtle color changes in unmoving patches). NASA Glenn's thearn has developed another approach which has the same problems (https://github.com/thearn/webcam-pulse-detector). We developed our own solution, involving (among other things) pixel-accurate facial feature detection and filtering methods not subject to FFT's regularity assumption, based on the synchrosqueezing transform
