Research On Human Pose Recognition In Rehabilitation Training Scene

Rehabilitation training is a necessary means for patients with limb motor dysfunction to regain their limb motor ability,so the study of rehabilitation medical technology is of great significance to society and family.However,traditional rehabilitation training techniques have some problems,such as high cost,need to wear equipment and need to receive rehabilitation treatment at designated points.In order to get rid of the limitations of traditional rehabilitation training such as cost and location,visual technology represented by Kinect depth camera has been gradually applied to rehabilitation training scenes in recent years.Existing Kinect devices are expensive and sensitive to environmental lighting and occlusion.With the continuous progress of deep learning technology,human pose estimation and human motion recognition technology has achieved good results,and has the possibility of application in practical scenes.However,while obtaining high precision,the complexity of these deep learning models is very high,which restricts their practical application to a certain extent.In view of the above requirements and the problems existing in the existing technology,under the guidance of comprehensively measuring the balance between the accuracy and complexity of the algorithm,this thesis proposes a lightweight centernet network human pose estimation algorithm and a lightweight motion recognition network based on human pose sequence.In this thesis,the human posture sequence of patients is extracted from the image sequence using lightweight centernet network.Then,based on the obtained human posture sequence,the rehabilitation training movement of patients is recognized by using lightweight action recognition network.Finally,this thesis based on lightweight human posture estimation and motion recognition algorithm,QT framework is used to build a rehabilitation training system based on human posture recognition,which can enable patients to complete rehabilitation training at home.The specific research contents of this thesis are as follows:（1）In view of the situation that the overall complexity of the centernet network is high and the actual reasoning speed is slow,in order to make the lightweight centernet network achieve a better compromise between accuracy and complexity,this thesis first analyzes the complexity of the centernet network.Analysis,and then the lightweight design of the centernet network is carried out based on the lightweight modular design.Specifically,this thesis realizes the lightweight design of the centernet network through methods such as lightweight basic residual blocks and upsampling modules.After design,this thesis finally obtains a lightweight centernet network.The parameter amount and computational complexity of the network are 3.56 M and 6.8GFLOPs,respectively,which are only 17.7% and 17.4% of the centernet network,respectively.In the end,this thesis achieves a large compression of the centernet network at the cost of losing a small amount of accuracy,and achieves a better balance between network accuracy and complexity.（2）This thesis proposes a lightweight action recognition network based on human pose.The network fuses information such as position-view invariance features of actions and multiscale motion features.The experimental results show that the network has achieved a recognition accuracy of 96.90% on the rehabilitation training action dataset containing 15 types of actions constructed in this thesis,and its parameter amount and computational complexity are only 0.41 M and 0.006 GFLOPs,respectively,achieving a lightweight efficiently identify the patient’s rehabilitation training movements.At the same time,on the public action recognition data set JHMDB,the network obtained an accuracy rate of 80.70%,and achieved a good recognition effect,which further verified the effectiveness of the network.（3）This thesis builds a rehabilitation training system based on human posture recognition based on the QT framework.The system not only uses the lightweight algorithm and cosine DTW algorithm proposed in this thesis to complete the identification and evaluation of the patient’s rehabilitation training actions,but also builds a somatosensory game function for rehabilitation training.In addition,the system has also built functions such as managing personal information and disease course records,allowing patients to perform rehabilitation training in a home environment,thereby improving the efficiency of rehabilitation training.