Non-contact Heart Rate Detection Using Thermal Infrared Video Based On Multi-region Analysis

With the improvement of people's living standard,more and more people pay attention to their physical and mental health.Heart rate(HR)is an importan indicator of physiological and pathological state of human body,the detectionof which is beneficial to the prevention,diagnosis and treatment of cardiovascular diseases.Traditional HR detection techniques adopt detection devices to directly contact with the skin of subjects for reliable HR measurement,which will cause a lot of inconvenience and uncomfort.In recent years,non-contact HR detection methods have drawn an increasing attention.Among them,the new emerging technology of detecting HR based on image photoplethysmography(i PPG)is a hot topic,due to the fact that the cameras employed by i PPG technique is wide spread,low cost,and has no need to contact with the skin of subjects.IPPG technology is generally equipped with color cameras to shoot video images.However,the color cameras will not suitable for dark or dim environments.Thermal infrared(TIR)cameras instead can be an option,since it can record the skin-area pixel values periodically changing along with the temperature changes caused by the cardiac cycles.Besides,TIR is less sensitive to illumination changes.In this thesis,we proposed to measure HR remotely based on TIR technique.The main contributions are as follows:1)A new database involving 30 healthy volunteers(24 males and 6 females)is established for this study.Each subject keeps still in a relatively stable environment and conducts the experiment only once.Thereby,there are 30 videos in total.2)The multi-region analysis is proposed to measure HR using TIR videos.First,three facial regions of interest(ROIs)are manually determined for the first TIR image.Then,Kanade-Lucas-Tomasi(KLT)tracking algorithm is used to track ROIs for other images.Afterword,the mean pixel value within each ROI is calculated and concatenated to form a time sequence.Second,each time sequence is pre-processed for further multi-region analyis.Third,two methods,multi-region independent component analysis(termed as MRICA)and multivariate empirical mode decomposition(termed as MEMD)are used to extract independent components(ICs)and intrinsic mode functions(IMFs).Finally,the IC candidates and the IMF candidates are determined by power spectrum analysis,the dominant frequency of which is within the range from 0.75 Hz to 2.5 Hz.The one with the corresponding frequency having the highestsignal-to-noise ratio(SNR)is determined as the target IC or target IMF,and this frequency will be determined as the HR frequency.The algorithm performance will be evaluated by the proposed four metric.Besides,the performance of using multi-region analysis is also compared with that of using single-region analysis.Experimental results on the database demonstrated the feasibility and superiority of our proposed method.Among them,MRICA achieves the best results,with an average absolute error of 3.17 beat per minute(bpm),a root means square error of 2.93 bpm,a standard deviation of 4.3 bpm and a correlation of 0.87.