Robust Method for Non-Contact Vital Sign Measurement in Videos Acquired in Real-World Light Settings From Skin Less Affected by Blood Perfusion

Color fluctuations caused by blood perfusion on the skin are an essential tool for the success of imaging photoplethysmography (iPPG) methods. However, color fluctuation of the non-facial skin is less observable than that of the facial skin. Furthermore, some light sources may produce image artifacts obscuring vital signs that iPPG methods are aimed to observe. This work contributes to an iPPG domain in several ways. First, it systematically visualizes and evaluates the impact of light settings on the accuracy of heart-rate estimation. Three light settings were arranged, and Eulerian video magnification was applied to illustrate their underlying characteristics. This work showed that fluorescent tubes could produce stripe artifacts and failed existing iPPG methods. Second, techniques to strengthen blood perfusion signals and better estimate heart rates are proposed. They significantly improved the accuracy in three light settings and a wide range of skin types categorized by the Fitzpatrick standard. Lastly, this work presents a new public dataset acquired from the forearm skin of 83 human subjects. The proposed method utilizes the Savitzky–Golay filter to treat components obtained by independent component analysis (ICA). We apply normalized autocorrelation to evaluate signals’ period consistency and weigh ICA components accordingly. This method is a reliable way to accumulate weighted fast Fourier Transform spectra and compute the final frequency representing an estimated heart rate. In our experiments, the method significantly reduced the errors of prior methods in all of the three light settings, especially for ceiling fluorescent tubes where the errors were reduced by 83% on average.