An Online Diagnose And Rehabilitation System Based On Human Pose Estimate

With the rapid development of society,the pace of people’s life is speeding up.When mild symptoms occur,they may be tired of going to the hospital,which will lead to the aggravation of the condition.In addition,the rehabilitation of chronic diseases(such as osteoarthritis)is a long process for patient,if can’t occupy the recovery,you need to often go to a hospital or rehabilitation institutions,increasing the hospital accepts the pressure at the same time the patient trouble.In recent years,many online interrogation systems have emerged,which can meet the needs of people with mild discomfort.It enables people with mild discomfort to have a preliminary understanding of their condition in advance,and then doctors can issue rehabilitation suggestions.But they still need to rely on special rehabilitation equipment and rehabilitation places.To solve this problem,this thesis proposes an online consultation and home-based rehabilitation system based on computer vision,and implements it with human pose estimation technology.Patients do not need to wear any equipment or purchase any machinery,and can use ordinary mobile phones or computer cameras to complete the training,thus greatly reducing the equipment complexity and location restrictions of traditional rehabilitation training.In recent years,the rapid development of artificial intelligence has greatly promoted the research and related technologies in this field.This thesis mainly completes the lightweight improvement of the network around the human pose estimation technology and completes the development of home-based rehabilitation system.The specific work is as follows:1)The top-down model can greatly improve the accuracy of prediction,and its performance is significantly better than that of bottom-up algorithm in most scenarios.However,the obvious problem lies in top-down algorithm,which usually has problems of large number of parameters,high equipment requirements and slow operation speed.In order to solve this problem,we improved the HRNet algorithm,proposed the shuffle module to replace the residual module,and added the channel weighting method to fuse the branch features,so as to reduce the number of parameters and computation to the maximum extent,and improve the information utilization rate and accuracy.COCO2017 is selected for the training and testing data set.After 200 rounds of iteration,the accuracy of the original algorithm is 76%,while the accuracy of the improved algorithm is more than 74%,but the recognition speed is increased from about 10 frames per second to 20 frames per second,which saves computing resources and speeds up the analysis without losing too much accuracy.2)When the patient performs rehabilitation training,the system will provide action examples and real-time pictures completed at the same time,and analyzes the patient’s skeleton model with a human pose estimation algorithm.Patients can compare the difference between himself and the standard action easily.The system will detect the completion of users’ actions and remind them when the standard is not met,so that patients can adjust in time.In addition,in order to ensure the effect of rehabilitation and provide channels for other users to check their health status.The system realizes the function of health detection,which can reduce the number of patients going to the hospital for review.3)Since most users of the system have physical discomfort,they are more likely to fall down accidentally than healthy people.In order to protect the safety of our users,we adds the function of fall detection.By analyzing the position information of human body’s key points,it can determine whether a person falls down and inform the emergency contact person in time.In this thesis,the lightweight improvement of HRNet has been completed.Through experimental verification,the number of frames identified per second has been increased from 10 to 20.In addition,the development of four main modules of online consultation and rehabilitation system based on human posture recognition technology has been completed.