SNR-based signal selection framework for heart rate estimation using rPPG signals extracted from facial videos

Heart Rate (HR) estimation using remote photoplethysmography (rPPG) signals extracted from facial videos has recently attracted significant research interest. Despite all the advantages of HR estimation using rPPG signals, it is essential to mention that these signals are highly susceptible to noise sources such as illumination variations and motion artifacts and can easily be polluted. To address the mentioned issues, we propose a Signal-to-Noise Ratio (SNR)-based signal selection framework for HR estimation using the rPPG signals. To this end, in each estimation time window, three estimator signals are obtained from facial regions using green channel variations information, its Intrinsic Mode Functions (IMFs), and the Independent Component Analysis (ICA) approach applied to the extracted facial RGB signals. Finally, among all three estimator signals, one with the highest SNR is selected as the best rPPG signal to estimate the HR value. The proposed framework evaluation was conducted on the publicly available UBFC-rPPG dataset, demonstrating the effectiveness of the proposed framework. The proposed framework could achieve a Mean Absolute Error (MAE) of 2.40, a Root Mean Square Error (RMSE) of 4.36, and a Pearson Correlation Coefficient (PCC) of 0.97, which outperforms many of the existing state-of-the-art methods.