Research Of The Algorithm For Remote Heart Rate Estimation Based On Face Video

Heart rate is one of the important indicators to measure the heart function state,and also is an important physiological parameter to reflect the change of emotional state.Real-time and convenient heart rate monitoring are great significance and value for assessing health of people and emotional state.Traditional heart rate estimation mainly uses the contact electrode sheet or smart bracelets,which requires skin contact with the tested person or guidance from a professional physician,so it is difficult to achieve the real-time and convenience.Remote Photoplethysmography(RPPG)uses an ordinary camera to capture the faint periodic color changes of the face caused by heartbeat.Through the analysis of the signals of color change,the blood volume pulse signals are extracted,and then the physiological parameters such as heart rate and respiration rate are estimated.Due to its characteristics of non-contact and realtime performance,in recent years,especially since the outbreak of the epidemic,non-contact heart rate estimation based on RPPG technology has become the frontier and leading point of international research,and has been widely used.However,practical application scenarios is complex and variable,and the accuracy of heart rate estimation is easily affected by the change of illumination,head movement,equipment resolution and other factors.Therefore,based on the RPPG method,this paper studies the effects of light changes and head movements on the accuracy of non-contact heart rate estimation.The main research contents are as follows:(1)In order to solve the problem of signal trend drift due to the change of transistor parameters caused by the change of equipment temperature,a noncontact heart rate estimation method based on sparse representation of structure is proposed based on the fact that the actual heart rate signals are the same between different regions of interest(ROI),which reduces the reconstruction error.The accuracy of heart rate estimation is improved.Firstly,a signal drift correction method based on adaptive iterative re-weighting punishment least squares is proposed,and then heart rate estimation is carried out.Specifically,the method mainly consists of four parts: face key point location,structure dictionary construction,sparse representation solution and heart rate signal reconstruction.Considering the periodicity and fluctuation characteristics of heart rate signal,a hybrid structure dictionary composed of cosine bases of different frequencies and wavelet bases of different scales is constructed,and the sparse representation regularization constraint of structure is designed.The sparse representation coefficient is obtained by alternating direction method of multipliers.Finally,the heart rate signal is reconstructed.Experimental results on two public data sets,UBFC and COHFACE,show that the mean absolute error of heart rate estimation is less than 5bpm.(2)Aiming at the influence of illumination variation on the accuracy of noncontact heart rate estimation,a signal fusion estimation method is proposed,which can effectively improve the accuracy and stability of estimation.Specifically,firstly,multiple ROI are obtained based on facial key points,and RPPG signals are extracted by POS and CHROM methods for each ROI.Then the signal-to-noise ratio of each RPPG signal is calculated to obtain the fusion weight.Finally,the final heart rate signal is obtained by the weighted fusion of POS and CHROM methods.In order to demonstrate the effectiveness of this algorithm,a real time heart rate estimation test demonstration system is designed.Experiments show that the average absolute error of heart rate estimation by this method is less than 5bpm compared with the contact pulse oximeter.