Remote Heart Rate Estimation in Intense Interference Scenarios: A White-box Framework

Remote photoplethysmography (rPPG) offers a non-contact approach to measuring heart rate (HR) based on video. However, practical situations often contain motion and illumination changes that may corrupt the rPPG signal, making extracting heartbeat information challenging. To solve these problems, this paper proposes a white-box framework for remote HR estimation, which consists of four explainable stages: baseline drift (BD) removal, signal decomposition, blind source separation, and signal selection. Initially, an adaptive BD removal method based on variational mode extraction (VME) is proposed. This method dynamically determines the bandwidth for removing BD according to the possible heartbeat frequency, making the heartbeat components among multiple regions of interest (ROIs) more correlated and the noise more distinct. Subsequently, a novel decomposition method customized for the rPPG signal called pulse VME (PVME) is proposed. PVME determines the bandwidth of each extracted mode according to the signal quality, and it determines the convergence criterion of the decomposition by considering both the quality and energy of the signal. Compared to conventional methods, PVME is more adaptive and has improved efficiency and accuracy in decomposition. The rPPG signals obtained from multiple ROIs are decomposed into multiple signal sets. The multiset canonical correlation analysis (MCCA) algorithm is applied to extract the common heartbeat information from these signal sets. Finally, a pulse selection index incorporating both time-domain and frequency-domain features is introduced. This index effectively identifies the real pulse signal amidst multiple candidate signals. The results on private and public datasets validate the robustness and generalization of the proposed algorithm.