Abstract

This paper presents a non-contact framework for pulse rate variability (PRV) estimation using remote photoplethysmography (rPPG) from facial video. The proposed method integrates Plane-Orthogonal-to-Skin (POS) transformation for initial pulse extraction, Singular Spectrum Analysis (SSA) for signal enhancement, and a two-channel one dimensional convolutional neural network (1D CNN) for PRV estimation. The model utilises both the enhanced pulse waveform and a peak-based inter-beat interval tachogram to capture temporal variability. Experiments conducted on the UBFC-Phys dataset demonstrate strong agreement with reference signals and achieves low prediction error and good correlation. The results highlight the potential of combining signal processing and deep learning for accurate, unobtrusive cardiovascular monitoring in remote and naturalistic settings.