Retrieval Of Blood Volume Pulse Waveforms Using Multispectral Face Video Data

Jonathan Tyler, Ishtiaque Ahmed Khan, Surendrabikram Thapa, A. Lynn Abbott, Abhijit Sarkar

Abstract: This paper presents a new method for extracting blood volume pulse (BVP) signals from video of a person's face. Unlike most previous approaches, our method fuses data from spectral bands that extend beyond the visible wavelengths. The goal is to improve the performance of imaging photoplethysmography (iPPG), with a particular interest in good performance for diverse skin tones. Compared to visible wavelengths, near-infrared (NIR) illumination can penetrate deeper into the skin and provide information that is not present using visible light only. By combining standard RGB with NIR wavelengths ($850 \mathrm{~nm}, 940 \mathrm{~nm}$), our approach provides improvements in heart rate accuracy and BVP signal morphology. To test our approach, we collected multispectral data using three synchronized cameras from nine subjects who exhibit a range of skin tones. Our approach is based on an encoder-decoder neural network with temporal modeling, using image intensities extracted from preselected patches on the face. Compared to previous methods, our results show improvement in reducing bias that may result from differences in skin tone.

External IDs:dblp:conf/icip/TylerKTAS25