Adaptive Gaussian Mixture Model Driven Level Set Segmentation for Remote Pulse Rate Detection.

This paper presents an approach for pulse rate extraction from videos. The core of the presented approach is a novel method to segment and track a suitable region of interest (ROI). The proposed method combines level sets with subject-individual Gaussian Mixture Models to yield a time varying ROI. The ROI builds up from multiple homogeneous skin areas under constraints regarding the area and contour length of the ROI. Together with state of the art signal processing methods our approach yields an Mean Average Error (MAE) of 2.3 bpm, 1.4 bpm and 2.7 bpm on own data, the PURE database and the UBFC-rPPG database, respectively. Therewith, our method performs equal or better compared to widely used approaches (e.g. the KLT tracker instead of the proposed image processing yields an MAE of 2.6 bpm, 2.6 bpm and 4.4 bpm). Such results and the 2nd place with a MAE of 7.92 bpm in the 1st Challenge on Remote Physiological Signal Sensing prove the applicability of the proposed method. The taken approach, however, bears further potential for optimization in the context of photoplethysmography imaging and should be transferable to other segmentation tasks as well.