Towards A Lightweight Deep Learning Framework for Real-Time Vital Sign Prediction with RGB Cameras

Download PDF
Open Webpage

Seong-Wook Kim, Hyeong-Yeong Park, In-Gyu Lee, Se-Na Jang, Seong-Hyun Yu, Jun-Young Oh, Gee-Myung Moon, Ji-Hoon Jeong

Published: 2025, Last Modified: 27 Jan 2026BCI 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Remote photoplethysmography (rPPG) is a technology that extracts blood volume pulse (BVP) signals by analyzing the wavelength differences of light reflected from a person's skin using devices such as RGB and infrared cameras. This technology enables the measurement and estimation of various physiological signals, including heart rate (HR), systolic blood pressure (SBP), diastolic blood pressure (DBP), respiratory rate, and oxygen saturation. In this study, we overcame the limitations of existing rPPG methods by constructing a custom RGB-rPPG dataset. The dataset was collected from 14 subjects using an RGB camera (a standard webcam), wearable device, and sphygmomanometer. The leave-one-subject-out (LOSO) approach was applied for model training to predict SBP and DBP, while HR was estimated using a fast Fourier transform (FFT). The average mean absolute error for SBP/DBP prediction and HR estimation were 4.55 (±2.7660) mmHg, 1.84 (±1.6064) mmHg, and 6.91 (±8.0319) bpm, respectively. These results demonstrate that the constructed RGB-rPPG dataset and the proposed method are effective for predicting physiological signs.