Assessing Confidence in Video Magnification Heart Rate Measurement using Multiple ROIs

Heart Rate (HR) is an essential vital sign for assessing the health status of individuals and is clinically measured using electrocardiography (ECG) or Pulse Oximetry. However, there is a need for remote and contactless methods to measure HR where direct contact is not possible. Video Magnification (VM) is a non-contact method to measure HR by magnifying the subtle and minuscule changes in skin tone caused by blood flow. VM performance depends on several factors, including body motion, skin tone, and illumination. For a given region of interest (ROI), HR is identified as the frequency with the largest peak in the spectrum of skin tone modulations; however, spectra may vary across different ROIs and each ROI spectrum may have multiple peaks. In this paper, a confidence metric for VM-based HR is defined as the ratio of the first and second largest spectral peaks. This confidence metric is used to evaluate two different methods for combining disparate ROIs on the face: (1) frequency-domain averaging (FDA) and (2) time-domain averaging (TDA). These methods were tested on 19 subjects with the FDA method showing the correct HR as the largest spectral peak for 16 subjects, and the TDA method showing it correctly for 18 subjects. The average confidence metric was higher for the TDA method (2.9) than the FDA method (2.3) or a single forehead ROI case (2.3), and the confidence metric was generally lowest
$(< 2)$
for subjects with more body motion or darker skin tones.