An Action Recognition Method Based On Improved ST-GCN And Applied In Upper Limb Motion Assessment

Human action recognition plays an important role in medical health,physical exercise,safety prevention and control,etc.However,the diversity of the human body and the freedom of limb movements make action recognition face major challenges.Traditional action recognition methods rely on manual feature extraction,which is not suitable for big data and has low robustness.In action recognition algorithms based on convolutional neural networks,image features are automatically extracted and learned,which greatly simplifies traditional methods.The process of manual extraction is used,but less attention is paid to the dynamic bones of the human body.Therefore,this thesis introduces attention mechanism and extended hierarchical temporal convolutional network in Spatial-Temporal Graph Convolutional Networks(ST-GCN),proposes an improved ST-GCN human action recognition method,and applies it to the upper limb motion assessment in the health field of medical treatment.The specific work is shown as follows:First,the natural connection of human joints and the continuity of action sequences are used to establish an undirected spatio-temporal graph of human action.Construct a Spatio-Temporal Graph Convolutional Kernel in the process of action recognition to capture the spatial and temporal features in the action.Secondly,aiming at ST-GCN ignoring the different degrees of influence of joint points on movement in human action recognition,it is proposed to introduce spatial attention mechanism and temporal attention mechanism in basic ST-GCN to assign different weights to human joint points.,And make dynamic adjustments in the process of action recognition to obtain the differences in time and space of each joint point.At the same time,aiming at the problem that the single-layer temporal convolution in the basic ST-GCN does not sufficiently extract the temporal features,in the feature extraction of the time dimension,the temporal convolution layer of the basic ST-GCN is extended and a residual mechanism is added.To extend the hierarchical temporal convolutional network,the extended temporal convolutional network can capture the long-term dependencies in the action,and the residual mechanism can capture the short-term dependencies in the action.Finally,the method proposed in this thesis is experimentally verified on the public large activity dataset NTU RGB+D,Kinetics and the upper limb motion score dataset.The experimental results show that the motion recognition rate of this model is significantly improved compared with 3DCNN and basic ST-GCN,especially the recognition rate is improved by 8.3% under the cross-object benchmark of the NTU RGB+D dataset.Significantly,it has certain advantages over other methods.