The Home-based Remote Upper Limb Rehabilitation Motion Recognition System Based On Deep Learning

Purpose:China is the country with high incidence rate of stroke in the world.Among all the complications of stroke,the disability rate of hand motor dysfunction is as high as 100%.Therefore,the rehabilitation of upper limbs of stroke patients has become a difficult and hot topic in modern rehabilitation medicine.At present,most of the training programs for upper limb rehabilitation in China are based on professional rehabilitation hospitals,which bring economic and life pressure to patients and their families.In order to solve the problems of economy,life,psychology and rehabilitation effect of stroke patients in the process of traditional rehabilitation training,this paper designs a set of home-based upper limb rehabilitation system for patients.Methods:Firstly,The home-based remote upper limb rehabilitation motion recognition system developed in this project uses Open Pose algorithm to develop human tracking and recognition and joint angle program on Open CV platform in order to realize the joint measurement angle module.Secondly,developing the corresponding training game on Open CV platform.Thirdly,encapsulating the above modules with pyqt5 toolkit to develop the corresponding client.Finally,using Django and bootstrap framework to develop the required web modules.The data sharing among the above modules is completed by My SQL database management system to complete the configuration of the server module.In order to demonstrate the feasibility and error source of joint angle measurement module design,this paper designs the experiment of upper limb rehabilitation system on the measurement accuracy of joint range of motion.A healthy adult was selected as the subject.Two different measurement methods,Vicon infrared high-speed motion capture system and home-based upper limb rehabilitation system,were used to measure the left shoulder abduction and adduction,flexion and extension,and elbow flexion and extension simultaneously,and the experimental data were compared..Results:The system can realize the rehabilitation training of patients at home according to the game.The training effects such as reaction time,joint range of motion,and upper limb movement completion can be visually displayed in real time.At the same time,the training effect data can be recorded,saved and uploaded;training data can be stored for a long time and upload it to the cloud,so that doctors can view it in time on the Internet;it can clearly grasp the home rehabilitation progress of patients,so that online guidance is more scientific and timely.The intraclass correlation coefficients(ICC)of "shoulder abduction and adduction","shoulder flexion and extension" and "elbow flexion and extension" were 0.994,0.995 and 0.990.Bland-Altman consistency analysis shows that among 249 data points of shoulder abduction adduction sampling,6 samples were outside the 95% consistency limit(-13.2,7.5),accounting for 2.4%.The absolute mean value of the difference between the two measurement methods was 3.1%,and the error percentage(the ratio between the maximum error of the two measurement methods and the range of activity)is 10.4%;Among 201 data points of shoulder flexion and extension sampling,10 samples were outside the 95% consistency limit(-11.8,6.9),accounting for 4%,and the absolute mean value of the difference was 2.5%,and the error percentage is 8.9%;Of the 250 data points sampled for elbow flexion and extension,9 were beyond the 95%consistency limit(-4.2,9.5),accounting for 3.6%.The absolute value of the difference between the two was 2.6%,and the error percentage is 10.8%.Further analysis of the experimental data shows that the new system will produce system error in a certain direction according to different measurement items.This is due to the limitations of two-dimensional imaging in the process of joint rotation.The detailed reasons are as follows: 1.The limitations of automatic recognition algorithm lead to the instability of joint point recognition;2.During the process of joint rotation,the skin and muscle will have a certain degree of displacement relative to the bone in the rotation plane,so that the joint point after rotation is no longer the position before rotation.Conclusions:Based on the advantages of computer vision and portable hardware equipment,the home-based upper limb rehabilitation motion recognition system designed in this paper improves the shortcomings of traditional rehabilitation training,such as boring and monotonous training,tedious data recording,difficulty to save data,complex operation and so on,which has a wide application prospect.In the above three activities,the difference between the two measurement methods does not change with the change of the mean measurement angle,which means they are independent of each other,and the amount of data beyond the 95% consistency limit is less than 5%.The degree of consistency is well,which is within the recognized error range in the field of rehabilitation.