ZCY-POS: An Unsupervised Remote Photoplethysmography (rPPG) Algorithm Improved by POS: Leveraging Center Cropping and Optimized Projection Matrix

Remote Photoplethysmography (rPPG) is a non-contact biomedical sensing technology that extracts pulsatile signals from facial videos to compute physiological metrics like heart rate. Given the lower computational demands of unsupervised methods, we developed an improved algorithm based on the POS method, named ZCY-POS. This method optimizes video input through center cropping to focus on facial regions and reduce background noise. Through experiments, the optimal projection matrix was determined, enhancing the quality of RGB signals. These signals are transformed into normalized feature vectors, then mapped to a new chromatic space to highlight color changes caused by the pulse. After precise weighting and combination, the signals are detrended and filtered, and the resultant signal is computed for heart rate using Fast Fourier Transform (FFT). Finally, we compared our method with existing methods on two public datasets. In comparisons with unsupervised methods, ZCY-POS leads by at least 2.07 in MAE on the UBFC-rPPG dataset and performs excellently on the UBFC-PHYS dataset; in supervised comparisons, it leads by at least 0.49 in MAE on the UBFC-rPPG test set and by at least 0.97 on the UBFC-PHYS test set. These results indicate that ZCY-POS shows potential advantages over other methods.