Physiological Signal Preserving Video Compression for Remote Photoplethysmography

• Published in: IEEE sensors journal Vol. 19; no. 12; pp. 4537 - 4548
• Main Authors: Zhao, Changchen, Chen, Weihai, Lin, Chun-Liang, Wu, Xingming
• Format: Journal Article
• Language: English
• Published: New York IEEE 15.06.2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Biomedical monitoring; Clustering; Digital cameras; Feature extraction; Light sources; Luminous intensity; Patient monitoring; Pattern clustering; Photoplethysmography; Physiological signal preservation; Physiology; pulse region detection; remote photoplethysmography; Remote sensors; ROI-based video compression; Signal monitoring; Subspaces; Video coding; Video compression; Video data
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2019.2899102

The consumer-level digital camera has become a physiological signal monitoring sensor due to the rapid growth of the remote photoplethysmography (rPPG) technique. However, rPPG suffers from the artifacts caused by video compression, a technique that widely exists in nearly all video-related applications. This limitation greatly narrows the application range of the rPPG. In this paper, a novel physiological signal preserving video compression algorithm called POSSC is proposed such that the existing rPPG signal extraction approaches can be applied directly on the compressed video without modification. The proposed approach consists of three main steps: 1) rPPG feature extraction; 2) sparse subspace clustering; and 3) region of interest (ROI)-based video compression. This paper models the skin/non-skin feature classification problem as sparse subspace clustering. Physiological signals are preserved by allocating more (fewer) bits to the ROI (non-ROI) regions. A self-collected benchmark dataset is established to evaluate the performance of POSSC in terms of body part, ROI size, light source, illumination intensity, and multiple subjects in an image. The results demonstrate that POSCC is effective in preserving physiological signals for facial videos under normal light intensity, insensitive to ROI size, shape, and the number of subjects. This paper also demonstrates the effectiveness of sparse subspace clustering for pulse region detection.