Research On Non-Contact Heart Rate Measurement Method Based On Facial Video

Heart rate is very important for determining a person’s physical and mental state.Traditional heart rate measurement uses wearable devices.Although the heart rate can be accurately measured,long-term wearing will cause discomfort to the subjects,and some subjects are inconvenient to use the wearable devices.In recent years,researchers have proposed many methods for non-contact heart rate measurement based on face video,but in the process of extracting pulse wave signal from face video,it will be affected by video information redundancy,light changes,head movement and other factors,resulting in low accuracy of heart rate measurement.How to use non-contact method to measure heart rate more accurately has important research significance and application value.The main work and innovation of this paper are as follows:(1)Face video as a heart rate signal source,in addition to the face area related to the heart rate signal,there are background areas unrelated to the heart rate signal,and the background area contains a lot of noise of light changes,which is not conducive to the extraction of heart rate signals.In order to eliminate the interference of background noise,this paper improves the method of region of interest extraction,which is divided into two stages.In the first stage,the Retina Face face detection framework is used for face detection,and the detected face part is cropped out as the area of interest.Then,the Linknet segmentation algorithm is used to segment the face skin region to obtain the final region of interest.(2)In order to reduce the interference of redundant face video information,this paper designs a time-domain segmentation subnetwork composed of several subspace networks.In this paper,the learning process of time-domain segmentation subnetworks is divided into three stages: first,segmented videos are sent to subnetworks respectively to collect facial features,then important facial features are paid attention to through spatial attention mechanism,and time information is aggregated using average pooling aggregation function,and finally the feature graphs generated by each subnetwork are cascaded out.This method of gradually converting face video into feature map significantly reduces the input variance and improves the accuracy of heart rate measurement.Experimental verification was carried out on the public dataset UBFC-RPPG.The mean absolute error and standard deviation of the proposed method for measuring heart rate were 1.97 and 3.82,the mean absolute error and standard deviation were 5.80 and 8.53 on the PURE dataset,and the mean absolute error and standard deviation were 3.24 and 7.86 on the COHFACE dataset.(3)In order to counter the noise caused by light changes and head movements,this paper designs a r PPGNet network with attention module to extract pulse wave signals.r PPGNet can more accurately model the relationship between time series and space,and the attention module can help the model more effectively correlate the information in the time domain,improve the model’s anti-noise ability,and thus improve the accuracy of heart rate measurement.In this paper,the end-to-end network composed of face region of interest extraction,video information de-redundancy and r PPGNet concatenation is used as the final heart rate measurement model.The average absolute error on the UBFC-RPPG dataset is 1.45 with a standard deviation of 3.97,and the average absolute error on the PURE dataset is 2.88 with a standard deviation of 6.68.The absolute error of all COHFACE data sets is 5.80,and the standard deviation is 8.53.The experimental results show that the non-contact heart rate measurement method proposed in this paper has certain reliability and feasibility.