| Human action recognition aims to automatically analyze and judge human behaviors through computer intelligence.It is an important research topic in the field of computer vision and pattern recognition.Skeleton-based action Recognition is an important research direction in this field.Skeleton data is an abstract representation of human motion,consisting of continuous human skeletal joint points along the time axis.The behavior recognition task based on skeleton data has the characteristics of high robustness,low cost,and easy deployment.Its application fields include security monitoring,intelligent health care,virtual reality,autonomous driving,and so on.With the continuous exploration of researchers in this field,graph convolutional networks(GCN)for skeleton-based action recognition have made some research progress.However,it still faces some challenges at present.On the one hand,existing dynamic graph convolutional methods enhance feature extraction ability by learning sample-related topological matrices,but similar actions can produce similar topological relationships,which can interfere with the classification results of the model.On the other hand,conventional regularization methods are not suitable for skeleton data due to overfitting phenomena in skeleton graph convolution,and regularization methods need to be redesigned based on skeleton data.To address these problems,this paper proposes a new network named Topological Transformation Graph Convolutional Network(TT-GCN).The TT-GCN network has two advantages:(1)TT-GCN has a stronger classification ability for similar actions.The design of Topological Transformation Graph Convolution(TT-GC)is based on the principle that similar movements should have similarities within the same category,but differences between different categories.Therefore,TT-GC enhances the inter-class differences and intra-class similarities of actions to improve the model’s ability to learn action sequences.The TT-GC mainly consists of a node-enhanced topological relationship extraction module and a second aggregation module based on enhanced topology,which is used to obtain the enhanced topology structure of skeleton data and extract features of input data based on the enhanced topology.(2)TT-GCN can effectively reduce the risk of overfitting.This paper proposes a novel regularization method,DropSkeleton,and uses it in TT-GCN.This method is designed based on the skeleton graph structure.The information of the discarded node is prevented from flowing into other nodes through node level discarding and edge level discarding.At the same time,the adaptive selection discarding is realized based on the spatial-temporal characteristics,which further enhances the regularization effect.The TT-GCN model was evaluated by conducting ablation experiments on the NW-UCLA dataset and comparative experiments on the large-scale action recognition datasets NTU RGB+D 60 and NTU RGB+D 120.The experimental results showed that the proposed TTGCN network achieved better performance than other methods overall.In addition,this paper also conducted research on student action recognition on campus and achieved end-to-end recognition of student behavior.Due to the lack of public datasets,this paper also constructed a campus action dataset(Student Action in School Scene,SA-SS).The experimental results showed that TT-GCN had good performance on this dataset and had good generalization ability. |
