| In the field of computer vision,human action recognition has been widely used in major scenarios such as video surveillance,virtual reality,and human-computer interaction.In recent years,although the research on human action recognition has made rapid progress,it is still a very challenging task due to the complexity and diversity of human actions.Skeleton-based action recognition has attracted a large number of researchers’ interest due to its robustness in dynamic environments and high recognition rate in complex scenes.In recent years,the continuous development of Open Pose,Deep Cut and other human body pose recognition algorithms has reduced the difficulty of obtaining skeleton data.At the same time,the increasing development of graph convolutional networks has greatly improved the accuracy of human skeleton action recognition.With the introduction of the classic human skeleton action recognition(Spatial-Temporal Graph Convolution Neural Network,ST-GCN)algorithm based on spatio-temporal graph convolution neural network,human skeleton action recognition has achieved a qualitative leap,but there are still some problems: Firstly,the features of key nodes are underutilized in space,and the ability to recognize actions with large motion differences needs to be improved.Secondly,due to the limitations of graph convolutional networks,the relationship modeling between non-adjacent nodes in space and time is not considered.The problem,the final model complexity is high and the predefined adjacency matrix is not flexible,which makes the training of the network too difficult.Therefore,in response to the above problems,this thesis launched research on the skeleton action recognition algorithm based on the spatio-temporal attention mechanism,aiming to improve the accuracy of human skeleton action recognition.The main research work is as follows:(1)Aiming at the insufficient utilization of key node features in ST-GCN,this thesis uses the attention mechanism and spatiotemporal graph convolution network to construct a spatiotemporal attention graph convolution network(Channel-Spatiotemporal Attention Graph Convolution Network,CSA-GCN).In the process of feature extraction,spatial features are first extracted through spatio-temporal graph convolution,and then the Channel-Spatiotemporal Attention(CSA)module is embedded to strengthen the key node features to obtain the spatiotemporal features of the human skeleton.It makes up for the problem of insufficient utilization of key nodes in the spatio-temporal graph convolutional network.Compared with ST-GCN,the final experimental results show that the accuracy of the algorithm in this thesis has increased by 2.8% and 2.9%,respectively.(2)Aiming at the problem that the relationship modeling between non-adjacent nodes is not considered in the existing methods,this thesis proposes Joint Spatiotemporal Collaborative relationship modeling(JSCRM).Based on the original algorithm,this thesis uses the query and key vector of each node to construct the relationship feature matrix between nodes,and realizes the feature capture and reorganization of nodes in the space-time dimension.The relationship between non-adjacent nodes in space and time is constructed through the query and key feature vectors in the attention mechanism,and then the key features are weighted to construct a learnable adjacency matrix.Compared with ST-GCN,the final experimental results have improved the accuracy of the algorithm by 6.3% and 7.3%,respectively,which verifies the superiority of the algorithm.(3)Aiming at the high computational cost of attention-based networks and the noise impact caused by the relationship modeling between individual joints,this thesis proposes the SpatialTemporal Cooperative Awareness Block(SCA Block).This thesis converts the skeleton diagram from the joint level to the component level,which reduces the complexity of the model while realizing the reorganization of the feature matrix relationship,and enhances the positioning and mining of the skeleton receptive field and the key receptive field diagram.In the SCA Block,this article uses the Part-joint Spatiotemporal Awareness Mining(PSAM)architecture as the basic framework,and embeds the above-mentioned JSCRM into it.Realize the conversion of skeleton-level data to part-level data through the operation of graph pooling,and then realize the generation time attention mechanism of the self-learning network receptive field during the construction of the spatial adjacency matrix and the weight process.Finally,experiments in this thesis on multiple benchmark datasets verify the superior performance of the proposed algorithm.Compared with ST-GCN,the accuracy of this algorithm is improved by 7.4% and 7.8%,respectively. |
