Fig. 7. Example video magnification process on a facial ROI for a subject that has limited motion except between the 40 and 50 second mark. (a) Average green pixel value; (c) Estimated heart rate based on the peak in each FFT; (b) Spectrogram and (d) FFTs for 10 second windows using 90% overlap. with red or blue wavelengths when using visible light [23], as shown in the figure. A time window for these signals is then analyzed to identify the frequency components in the color variation through frequency domain analysis such as Fast Fourier Transform (FFT) or Spectrograms where a peak in the spectrum indicates the heart rate. Methods typically use some form of sliding window method for the FFT analysis to allow for the tracking of rate variation [16]. Fig. 7 shows an example set of results for an ROI on the forehead of a subject and the processing of the green pixels within the ROI. Fig. 7a shows the average of the green pixels within the ROI for each frame of the captured video (30 frames/sec). Fig. 7b shows the resulting spectrogram and FFTs using a 10 second data window and 90% overlap between adjacent windows. Fig. 7d provides an alternative presentation of the spectrogram where each of the FFT results for the data windows is overlayed. One of the key challenges is the identification of a relevant frequency range for assessment. For example, heart rate can have a very wide range of values from at least 30 to 200 bpm, and this leads to a need to balance analysis across this wide 24 range while also being able to provide an accurate assessment. Fig. 7c shows the estimated heart rate by identifying the peak (highest local maxima) within the FFT for each window within the expected heart rate range. Another key challenge affecting the accuracy and performance of the algorithms is variation in the lighting such as light source flicker or subject motion relative to the light source causing additional sources of variation in the frame-to-frame image. Fig. 7b shows the effects of even limited subject motion that occurred in the 40 to 50 second period. The effect of light variance and methods to compensate for its effects have been studied. The effect of light source hue and intensity (and variability of these) on the performance of rPPG led to a skin reflection model [8]. This was extended [24] with the introduction of the concept of a dc signal from an image that combines the unchanging background parts of the image. The analysis calculates and estimates the variation in this background signal, leading to an estimate of the average lighting variation. ROIs are then analyzed based on their variation from the average. In another work, an auto-regressive model [25] was proposed to address the issues associated with lighting flicker. IEEE Instrumentation & Measurement Magazine February 2022