Recognition And Evaluation Of Taichi Based On Graph Convolutional Neural Network

Rehabilitation after stroke is a cyclic process,and rehabilitation training is an important part of rehabilitation therapy.Tai chi,as a traditional Chinese medicine method,can promote the recovery of motor dysfunction in patients with hemiplegia after stroke.Action recognition is very important for understanding the semantic expression and intention inference of the actors.The existing methods pay more attention to the accuracy of action recognition in Tai chi,and how to establish the consistency evaluation of training actions is very important for rehabilitation training.This thesis takes Tai chi rehabilitation training as the research object,carries out the research on reconstruction of the missing motion capture data and action recognition and evaluation of Tai chi rehabilitation training,and designs and implements the data management and analysis platform of Tai chi interactive training.The main tasks of the research are as follows:1.Aiming at the problem of missing data collected for rehabilitation training,the Distance Likelihood Based Probabilistic Model（DLPMA）is proposed.The results of the four independent models are weighted to reconstruct the damaged marker tracks and linear correction was introduced to enhance the continuity of the recovered data.The test results show that the average recovery error of DLPMA was 10.92 mm,9.55 mm and8.86 mm under the change of sequence length,gap number and gap length,respectively.The marked track continuity after recovery is better,providing more authentic movement data for the motion recognition and evaluation of Tai chi rehabilitation training.2.View Point Transformation Adaptive Graph Convolutional Networks（VT-AGCN）is proposed for the features of complex movements such as body turning in rehabilitation training.The optimal consistent viewpoint of human skeleton spatio-temporal graph is determined by viewpoint transformation,and spatio-temporal feature extraction is carried out by introducing spatio-temporal convolution block,and human topological structure is learned adaptively.The experimental results show that VT-AGCN can reduce the viewpoint difference of "turning" motion skeleton and improve the recognition accuracy,which are 90.03% and 98.72% on NTU-RGB+D and UMONS-TAICHI datasets,respectively.3.In view of the lack of specific motion capture features in rehabilitation training motion evaluation methods,Maximum Mean Discrepancy with Neighbourhood Component Analysis based on self-attention mechanism and triple loss（SA＿MMD＿NCA）is proposed,which uses self-attention mechanism to score sequence pose and extracting more informative temporal features to improve the accuracy of motion evaluation.The results show that the false positive rate of SA＿MMD＿NCA in TPR-90,TPR-80 and TPR-70 on CMU was 32.84%,26.07% and 21.23%,respectively,and the average scores of NMI and F1 are 57.92% and 60.82%,respectively.4.Based on Spring Boot,My Batis and vue.js frameworks,B/S architecture mode,using Java language and My SQL database design and implementation of Tai chi training data information management;The scene interaction module is realized based on MFC,Open CV,Kinect SDK and other technologies,providing support for interactive rehabilitation training.The platform has the functions of user management,system management,expert and patient training data management and scene interactive training.