Research On Contactless Heart Rate Extraction Technology Based On Deep Learning

As one of the main physiological information of the human body,heart rate can reflect the health status of many aspects of the human body,such as exercise,metabolism,nervous and cardiovascular systems,and is also an important indicator of the functional state of the heart and disease conditions.Traditional contact heart rate detection requires the human body to remain in contact with the detection device,and this approach has limitations.For special populations such as the elderly,critically ill,and newborns,traditional contact heart rate detection methods cannot meet the demand for real-time monitoring.For patients with injuries or burns,contact sensors can also aggravate pain or discomfort,which can affect the accurate measurement of heart rate.With the rapid development of deep learning,video-based noncontact heart rate detection has been applied and has advantages such as low cost,simplicity and convenience.However,changes in ambient lighting in the video,the movement of the subject’s head,changes in facial expressions and other factors may interfere with the capture and analysis of pulse wave signals,thus affecting the accuracy of heart rate measurement.In order to solve the above problems and obtain more accurate heart rate signals,a more effective non-contact heart rate extraction network model is proposed in this paper.The specific work in this paper is as follows:1.A contactless heart rate extraction method based on a 3D Multi-Scale Convolutional Attention(3D-MSCA)mechanism is proposed to address the noise problem caused by changes in ambient lighting,subject’s head movement and facial expressions.The 3D-MSCA attention mechanism uses depth convolution instead of normal convolution,and uses strip convolution to obtain contextual information,and also uses multi-scale convolution to obtain features at different scales,and finally obtains more accurate heart rate feature information by fusing the features at different scales.This method captures different detailed information in the input facial data by extracting features at multiple scales,thus reducing the noise caused by illumination changes and motion,and thus improving the accuracy of the model.2.To enable the proposed network model to better learn spatio-temporal feature information,the Self-Attention Convolution Long Short Term Memory(SA-Conv LSTM)is used in addition to the inclusion of the 3D-MSCA attention mechanism.This module combines the self-attention mechanism and the Convolutional Long Short Time Memory network(Conv LSTM),which can effectively learn the spatio-temporal features in time series data,so as to extract the spatio-temporal features with global dependence and local dependence.The heart rate extraction part was later optimized to use a physiological signal Multi-Task Learning module capable of extracting both r PPG signals and heart rate signals.This module is able to exploit the strong correlation between the r PPG signal and the heart rate signal to jointly train the network model,resulting in better performance.Finally,experiments were conducted on three publicly available datasets,and the final heart rate Pearson correlation coefficient values reached 0.86 in the compressed COHFACE dataset,0.98 in the PURE dataset,0.93 in the UBFC-r PPG dataset,and 0.88 in the r PPG signal.The results show that the proposed method can improve the accuracy of contactless heart rate extraction.