Real-Time Temporal Superpixels for Unsupervised Remote Photoplethysmography

Segmentation is a critical step for many computer vision applications. Among them, the remote photoplethysmography technique is significantly impacted by the quality of region of interest segmentation. With the heart-rate estimation accuracy, the processing time is obviously a key issue for real-time monitoring. Recent face detection algorithms can perform real-time processing, however for unsupervised algorithms, i.e. without any subject detection based on supervised learning, existing methods are not able to achieve real-time on regular platform. In this paper, we propose a new method to perform real-time unsupervised remote photoplethysmograhy based on efficient temporally propagated superpixels segmentation. The proposed method performs the segmentation step by implicitly identifying the superpixel boundaries. Hence, only a fraction of the image is used to perform the segmentation which reduces greatly the computational burden of the process. The segmentation quality remains comparable to state of the art methods while computational time is divided by a factor up to 8 times. The efficiency of the superpixel segmentation allow us to propose a real-time unsupervised rPPG algorithm considering frames of 640x480, RGB, at 25 frames per second on a single core platform. We obtained realtime processing for 93% of precision at 2.5 beat per minute using our inhouse video database.