Research Of Skeleton Action Recognition Based On Central Difference Graph Convolution Network

Human action recognition is a challenging research topic in the field of computer vision.It integrates the research results of many disciplines such as computer vision,deep learning and pattern recognition,and has been widely applied in the fields of human-computer interaction,auxiliary medical treatment,video surveillance and so on.The human body can be regarded as a joint system,which is represented by the motion of joint nodes in three-dimensional space.With the development of depth sensors,skeleton-based action recognition has attracted extensive attention.With great representative ability in non-Euclidean space,graph convolutional network(GCN)has made great progress in skeleton-based action recognition.However,the existing GCN methods still have shortcomings.First,joint or bone data are widely used as input,which can only provide the inherent static information of human body.Second,the general graph convolution operation is directly used to aggregate the information of associated nodes in the topology graph,ignoring the local motion information between the central node and associated nodes.To solve the above problems,this paper improves the existing GCN methods.The work flow and research results of this paper are as follows:1)In this paper,a novel graph convolution operator,central difference graph convolution(CDGC)is proposed.The operator can not only capture the intensity-level semantic information of the associated nodes during feature aggregation,but also obtain the dynamic gradient information between the central node and its associated nodes.In addition,the proposed CDGC does not introduce any additional parameters,and can replace the vanilla graph convolution in any existing graph convolution networks to improve their performance.The experimental results on NTU RGB + D 60 dataset fully prove the effectiveness of CDGC.2)Although CDGC has great recognition performance,the computation is too complex.To solve this issue,a lightweight central difference graph convolution operator(Accelerated CDGC)is proposed in this paper.The operator combines spatial shift graph convolution with central difference,which not only achieves higher recognition performance,but also greatly improves the training speed of the network.It is a more expressive graph convolution operator.The experiments were carried out on two large benchmark datasets NTU RGB + D 60 and NTU RGB + D 120,with state-of-the-art performance achieved.