Research On Skeleton Action Recognition Algorithm Based On Attention Mechanism

With the development of modern information technology,video growth is very rapid,has become one of the mainstream media today.The understanding of video content has become a hotspot in the field of computer vision.The development of action recognition has changed from the traditional manual feature extraction to the graph convolution network model.The methods used for research are becoming more and more efficient,the model is becoming more and more automatic,and the results obtained are becoming more and more significant.Most of the proposed models are for obtaining better video representation in the inherent data set and can model the spatial static and temporal dynamic information in the video.However,due to the use of three-dimensional motion information in human action recognition,there are still many problems in obtaining spatial-temporal information and recognizing human action quickly and accurately in real video.In order to solve these problems,this paper studies from the aspects of human skeleton,image convolution feature,video spatial-temporal convolution,and uses graph convolution for action recognition.The main work of this paper is as follows:In order to increase the accuracy of acquiring characteristic skeleton motion information,an Attention Enhanced Direct Graph Convolutional Network(ADGCN)is proposed.Since the application of skeleton motion recognition in daily life usually takes the information of bones and joints as input data and uniformly extracts the features of the input data,the input data obtained in this way does not take into account the motion information between joints and bones as well as the key actions or joints in the action video.Therefore,the ADGCN model proposed in this paper firstly adopts a four-stream framework to input four kinds of information joint information,bone information,joint motion information and bone motion information into the four-stream framework.Then,the input data is transmitted to the directed graph convolutional network model to extract the dependence between joints and bones.Then,the proposed spatialtemporal channel attention network is used to enhance the temporal,spatial and channel information of key frames in each video or key joints in each frame.The model was trained,validated and tested on NTU-RGB+D and Kinetics Skeleton datasets,and ablation experiments were performed on NTU-RGB+D datasets.The results show that the proposed ADGCN model has certain advantages over existing models.A Central Difference Transformer Graph Network(CDTG)is proposed to make the skeleton model can recognize the spatial-temporal information of feature extraction.Since most skeleton action recognition models focus on node information and obtain local spatial-temporal dependence,the information obtained by this method does not realize the importance of gradient information between nodes and global spatial-temporal context for action recognition.And CDTG model proposed in this paper,first of all,transfer the motion of the joints and bones information input to the model,after processing the data transferred to the proposed network in the network are focused on the gradient information of the nodes and between nodes,increasing network said ability and generalization ability,then the proposed converter figure convolution of spatial and temporal,It is to extract the spatial-temporal features of the central difference network,and the spatial and temporal dependence relations marked by joints.Finally,the CDTG model was applied to NTU-RGB +D and Kinetics Skeleton data sets,and the results show that the CDTG model has better recognition performance than other existing models.