Robust Heart Rate Measurement by Adaptive ROI Selection for Head-Rotation Conditions

Region-of-interest (ROI) selection is a key process for the remote heart rate (HR) measurement approach. Previous studies have shown that the selected ROI could easily be lost under head-rotation circumstances and does not perform well under drastic changes in illumination. To solve the problems, this work proposed a novel adaptive ROI selection method. First, we collected a dataset for modeling the signal-to-noise ratio (SNR) evaluated by illumination and reflection angle. Second, we established an SNR-estimated model by studying the mathematical correlations between light intensity, reflected angle, and the SNR of the regions on the face. Then, we proposed a robust ROI selection method based on the SNR-estimated model for adaptive ROI selection. Finally, the ROI change denoising method is proposed to remove the ROI change noise for synthesizing the imaging photoplethysmography (iPPG) signal and calculating the HR. Performances of the proposed method were tested on three datasets, including our collected dataset and two public datasets. The experimental results show that on the MR-NIRP CAR dataset, the percentage of the time that the HR error is less than 6 bpm (PET6) value is higher than 7.3% and 4.6%, respectively, compared with the state-of-the-art (SOTA) methods. The SNR, standard deviation error (SD), mean absolute error (MAE), and root mean squared error (RMSE) metrics are also improved on the Head-rotation and PURE datasets. Therefore, our method is robust and can potentially be applied in various scenarios.