Research On Human Skeleton Action Recognition Based On GNN And Attention Mechanism

Human skeleton action recognition is to identify human action in human skeleton sequence data.The action recognition model based on manual features calculates the statistical information of skeleton joint points,ignoring the correlation between bone joints.The existing deep learning model treats the action sequence equally and cannot give different attention to the skeleton data frame.In addition,the deep learning techniques(such as convolution operations)used in the current related research cannot model the long-distance correlation of the skeleton data in the spatiotemporal dimension,and there is a problem of insufficient feature mining.In order to solve the above problems,the following two models of human skeleton action recognition are proposed based on graph neural network(GNN)and attention mechanism.First,to address the problems of insufficient modeling of the correlation between skeletal joints and failing to give different attention to skeleton data frames,a recognition model that integrates spatial reasoning and context-aware attention is proposed: SRCA.The model uses GNN to model the correlation between joint points in the skeleton for spatial structure inference,and obtains the spatial structure features of the skeleton.At the same time,context-aware attention is used to give different attention to the skeleton data through the actional context features and time series features of the skeleton data.The model solves the problem of insufficient modeling of the correlation between bone joint points and the problem of different attention to the skeleton data frame.Then,aiming at the problem of insufficient modeling of long-distance correlation in skeleton data,a human skeleton action recognition model based on multi-attention mechanism is proposed: HAN.The model uses spatial self-attention to model and learn the long-distance correlation between bone joints in the skeleton data,obtains the structural features of the skeleton,and then uses the temporal self-attention module to analyze the long-distance correlation between data frames and frames.The temporal features are obtained by modeling and learning,and finally the context-aware attention is used to give the temporal features different degrees of attention.The model can model different distance correlations in the spatiotemporal dimension of skeleton action,solve the problem of insufficient correlation modeling,and improve the accuracy of skeleton action recognition.Finally,the proposed SRCA and HAN human skeleton action recognition models are tested and analyzed on real data sets to verify the effectiveness of the proposed models.