Skeleton-based Action Recognition For Figure Skating

As the frontier direction of the computer field,human motion recognition has very important research significance.Related algorithms can be applied to many fields,such as motion content analysis,human-computer interaction,video synthesis,video retrieval,etc.In particular,the rapid development of short video platforms in recent years has resulted in more and more relevant human action videos.Research on the processing of these video information needs to be supported by efficient and accurate recognition algorithms.Therefore,research on human action recognition algorithms in the computer field is imperative.In the task of human action recognition,there are many classic video classification datasets,such as UCF101,HMDB51,etc.Many studies have achieved good results on these datasets.However,most of the sample content in these data sets has a certain correlation with the background,which can be classified by identifying the elements in the background,such as the recognition of musical instrument performance,cycling and other sports.This leads to some algorithms that are not enough to pay attention to the characteristics of human motion itself.This work chooses a more fine-grained figure skating sports dataset FSD-10 to study the human body motion recognition algorithm,discusses the advantages of bone point features in the study of human motion recognition,and proposes a new graph convolutional neural network structure——DSTG(Dense Spatial Temporal Graph Network)to improve the accuracy of human motion algorithms.The main contributions of this article are as follows:(1)Extract bone point features from the fine-grained skating data set FSD-10,establish a skeleton spatio-temporal map,and reproduce the existing graph convolutional human action recognition network on this dataset.This work proposes solutions to problems such as the unbalanced distribution of FSD-10 sample duration and large fluctuations in the characteristics of the skeleton points,and experiments have verified that the preprocessing scheme can effectively improve the recognition accuracy of the existing network model.The preprocessing strategy improves the recognition accuracy of STGCN network on FSD-10 from 75.03% to84.24%.(2)Combining the design ideas of Dense Net and STGCN,propose a new graph convolutional network structure DSTG module.The DSTG module is connected in the form of dense connection,which effectively utilizes the characteristics of the shallow layer of the network,reduces the redundancy of network parameters,and makes the DSTG network model parameters 5 times less than STGCN when the recognition rate on the FSD-10 dataset reaches86.82%.(3)Analyze the limitations of the existing graph convolutional network for human action recognition in the establishment of the adjacency matrix division strategy,according to the characteristics of skating motion,define the adjacency matrix based on the convex hull,which can be used as a module to the existing network model.Supplement,especially after combining with DSTG network,it can reach an accuracy of 88.47%,which improves the recognition effect of action recognition in figure skating.