A Comprehensive Study Of Non-contact Heart Rate Detection During Exercise

There is a common framework for non-contact heart rate detection algorithms.The facial information of target is collected through the camera to finish the task of obtaining heart rate value without contacting the target under test.It helps infectious patients,patients who moving with limitation,also the elderly and children.It provides the evaluation of emotion classification with reference index,which is of great research value.However,mostly,the content of these similar works focus towards the situation of fretting and static.This paper aims to establish an effective framework for heart rate monitoring under intensive exercise.On the basis of the collection for original brightness information in ROI area,we focus on the data about the facial feature points,and propose a method for judging the intensive motion while using camera at fixed place.The increment in the vertical and horizontal direction give corresponding threshold judgments for different exercise stage.Warm-up(<3pixels),acceleration(10?20pixles),intensive exercise(25?40pixels),which provide the preparations for the subsequent research.In terms of movement state,a processing algorithm based on the enhanced SSA motion component separation algorithm is proposed.The SSA is used to separate the peak frequency component with a higher priority.Eliminate and obtain the frequency component corresponding to the heart rate on this basis.The data source comes from a self-developed database.The general structure of the database is a completed treadmill fitness exercise process.All indicators have been significantly improved.The result of moving exercise segment performs well,and the ?bpm?5 proportion is increased by 20%,and the optimal is 95.45%.Compared with the traditional frequency peak detection algorithm,the overall image performance has been significantly improved.As for deep learning framework,we establish a special frequency sequence map as the input feature map,apply RetinaFace facial recognition project to obtain the region of interest,pre-process the features,extract the frequency sequence map through the attention mechanism,and invest in the basic network of VGG16 to build the data model.The data source is belongs to the HR-VIPL database of the Chinese Academy of Sciences(CAS).For the selection of three regional block models,the comparison of the best regional experiments is obtained through data comparison,and on this basis,preliminary results are obtained MAE=4.87,RMSE=14.52.Compared with the original author's experimental results,the RMSE is quite different.After repairing and improving the classification categories and abnormal data of the model through data statistics,the results of the optimized model are improved to MAE=2.38,RMSE=9.84,and pass 10998(70%)According to the statistics of the data set,37 cases(out of 0.33%)of abnormal values were found.Through the correction of the abnormal data,the MAE decreased to 59%and the RMSE decreased to 68%.A preliminary attempt to solve motion video through the basic network of deep learning has been realized.