A novel LSTM-based attention mechanism improves the accuracy of remote heart rate estimation by prioritizing high-quality rPPG signals and ignoring noise artifacts.
Heart rate estimation is very important for heart health monitoring. As a non-invasive optical technology, remote photoplethysmography (rPPG) has the advantages of noncontact, portability and low-price. However, motion and noise artifacts bring additional uncertainty to the results of heart rate estimation. Based on the signal quality assessment method, we propose a new remote heart estimation algorithm by signal quality attention mechanism and long short-term memory (LSTM) networks. The model consists of three parts: firstly, an LSTM network is used to estimate the heart rate sampling point by sampling point; secondly, a similar LSTM network predicts the signal quality; finally, an attention-based model uses the heart rates and quality scores predicted above to calculate the average heart rate of a period of time. The model allocates higher weights to the reliable heart rates estimated from high-quality signals, meanwhile, ignores unreliable results estimated from low-quality signals. Experiments show that LSTM with attention mechanism accurately estimates heart rate from corruption rPPG signal and it performs well on cross-subject tasks and cross-dataset tasks. The results also demonstrate that the scores predicted by the signal quality model is valuable to extract reliable heart rate.
Gao et al. (Wed,) studied this question.