Low Illumination Non-contact Heart Rate Monitoring Based On Deep Learning

With the development and advancement of technology,people hope to be able to monitor their heart rate changes at any time in order to prevent and detect cardiovascular diseases.Traditional heart rate monitoring devices are mostly contact-based,requiring them to be carried around and having many disadvantages.Remote photoplethysmography(r PPG)technology,with its monitoring comfort and convenience,has effectively solved the drawbacks of contactbased heart rate monitoring.However,significant achievements have been made in various fields by deep learning algorithms,which have been replaced by traditional methods,and noncontact heart rate monitoring based on deep learning is gradually becoming the mainstream method.Videos collected in low-light conditions will have a lot of noise,which affects the analysis and acquisition of heart rate signals.Solving the problem of inadequate lighting is of great significance for promoting the development of future monitoring field.(1)To address the issue of noise in videos collected under low-illumination conditions,a video brightness enhancement preprocessing scheme is used to improve the accuracy of heart rate monitoring.Through experimental comparison and analysis of existing image/video brightness processing algorithms,a low-illumination enhancement preprocessing algorithm based on ZERO-DCE is utilized to establish the relationship between the three color adjustment channels and construct a loss function.Through experimental comparison,the video preprocessed by this method shows a significant improvement in subsequent heart rate monitoring accuracy.(2)To ensure the temporal continuity of the r PPG signal after video analysis,a low-light non-contact heart rate monitoring method based on spatiotemporal convolutional neural network(CNN)is proposed.The spatiotemporal CNN model adopts a hybrid network structure of C3 D and R(2+1)D to better obtain spatial and temporal information,ensuring the continuity of the signal.An attention mechanism is added to assign different weights to different skin areas to obtain a more accurate r PPG signal and avoid incomplete feature extraction caused by monitoring a single region of interest.Partition constraints are applied to avoid overfitting during model training and improve generalization ability.Through heart rate monitoring experiments compared with mainstream heart rate monitoring models,the results show that the proposed method has certain advantages compared with other mainstream heart rate monitoring methods and has a higher heart rate monitoring accuracy.(3)In daily monitoring,there are skin color differences between the subjects and the samples in the training dataset,which can affect the accuracy of heart rate monitoring.Retraining the model parameters after collecting data samples from the subjects requires a significant amount of time.To address the aforementioned issues,a low-illumination noncontact heart rate monitoring approach based on feature transfer with spatio-temporal convolutional networks is proposed.By transferring the key information of the obtained r PPG(remote photoplethysmography)signal features and fine-tuning the network parameters with a small amount of new dataset,the weights of key information features in acquiring the r PPG signal are increased while reducing the weights of non-key information features.Experimental comparisons before and after feature transfer demonstrate that the proposed method effectively resolves the issue of low heart rate monitoring accuracy caused by skin color differences.