Human Action Recognition And Prediction Based On Deep Learning Of Graph Convolutional Networks

Manufacturing is the cornerstone of the national economy,to meet the needs of industrial production of small quantities and multiple varieties and personalized,machinery manufacturing industry ushered in industrial change,large quantities and a single variety of traditional production mode can no longer meet the requirements of the new era,flexible and efficient manufacturing hasbecome an inevitable trend.Human-machine collaboration ad opts the working mode of intelligent interaction between human and machine,and reatizes real-time and efficient human-machine collaboration with the help of human action recognition and prediction technologies,which greatly improves the flexibility and response speed ofmanufacturing system.Based on the action sequence data in the video,this paper establishes the human pose estimation model,human action recognition model,and human action prediction model to realize the recogntion and prediction of human action,and tests the performance of the human action recognition model and action prediction model by using datasets and actual scenes to provide theoretical basis and engineering practice reference for human-machine collaboration.A human pose estimation model based on the ELIIRNetwork is established to obtain the human skeleton information in images.Based on the HRNet human pose estimation model the LIIRNet model is established by lightening the Basicblock residual module which accounts for a relatively targe number of parameters in the HRNet model,the LIIRNet model has a lower number of parameters and computational volume compared with HRNet;by introducing the CBAM attention mechanism in the lightweight Basieblock residual module,more effective human keypoint features are obtained from the channel and spatial dimensions,and the ELIIRNet human pose estimation model is established to enrich the human keypoint feature information.The test results of the datasets and the actual scenes show that the established ELIIRNet can accuctely estimate the human pose in different scales and complex backgrounds with a low number of parameters and low computational effort,but the pose estimation performance needs further improvement for the presence of severe occlusion conditions.A human action recognition model based on the AST-RGCN network is established,and the human skeleton on each frame acquired by the human pose estimation model is modeled in time and space to realize human action recognition in the video sequence.Based on the ST-GCN human action recognition model,the AST-GCN model is constructed by introducing the spatiotemporal keypoint attention mechanism to obtain the keypoint features of human action sequences;the AST-RGCN human action recognition model is established by introducing the residual connection in the graph convolution module to retain the underlying features and enrich the input attention mechanism information.The dataset and actual scene test results show that the constructed AST-RGCN model can accurately identify video human action categories with obvious action features and slightly increase the number of parameters,but for the actions related to hand keypoints and actions crossed by double human keypoints,the model action recognition performance has room for improvement due to the problem of missing hand keypoint information and a large amount of occlusion.A human action prediction model based on the NSTS-GCN network is established to predict future pose sequences based on historical human action pose sequences based on action recognition.Based on the STS-GCN human action prediction model,the MSTS-GCN model is established by designing a multiscale temporal convolution decoder;a GRU-TCN decoder with the introduction of GRU is designed to improve the ability to model the spatio-temporal information of human action sequences,and the NSTS-GCN human action prediction model is established.The test results on the datasets show that the NSTS-GCN model has reduced the j oint prediction error of human action sequences for each time node and has superior action prediction performance than the original STS-GCN model.