Heart Rate Estimation Based on Facial Image Sequence

Heart Rate Estimation from Facial Image Sequences of a Dual-Modality RGB-NIR Camera

“…The color conversion is performed based on two considerations. First, the superior performance of HR estimation in CIELab space over the RGB space has been considered in [18,28] for ICA-related algorithms, and the luminance component is less effective in extracting the BVP information [25]. Additionally, the decrease of signal size from 4 × T to 3 × T will help in saving computational loads and speed up the processing in realistic applications (such as fitting devices in gym or elderly caring center).…”

“…Inspired by Tulyakov's work [28] and the technique of RPCA (Robust Principal Component Analysis) [31], we would like to model our observations matrix M by M = L+ S (3) where L stands for a low-rank matrix and S is a complementary part. In this model, the low-rank matrix L is considered as a "background signal" which is present throughout the video and to be recovered from the highly corrupted measurements M ; the matrix S can be considered as a sparse outlier noise resulted from abrupt motion, illumination change, or region tracking errors.…”

“…Inspired by Tulyakov's work [28] and the technique of RPCA (Robust Principal Component Analysis) [31], we would like to model our observations matrix M by…”

“…The measured rPPG signal can be actually modeled as a combination of real HR signal and other irrelevant noises (e.g., from head motion, facial expression, or environmental illumination), denoted as E, which might also contribute to L in Equation (3). After denoising M (3 × T) and decomposing M(3 × T) to obtain L (3 × T), ICA was often adopted to separate L into the rPPG signal P and other irrelevant sources E. By referring to [9,28], the Joint Approximation Diagonalization of Eigen-matrices (JADE) algorithm [33] was adopted for ICA analysis to derive another 3-component output signals from L, which are then further processed by a customized selection algorithm to identify the best component as P for HR estimation. For each of the 3 output component signals, we locate all the peak positions and calculate the standard deviation (SD) of the peak-to-peak temporal distances.…”