Research On The Action Recognition Method Of Human Skeletal Point By Integrating Attention Mechanism

With the rapid development of artificial intelligence technology,more and more researchers and research institutions begin to pay attention to human action recognition technology.Human behavior is very complex,and human action recognition technology can be widely used in public security,intelligent home,intelligent classroom and other fields.In human action recognition,both the input feature representation and the recognition algorithm model will affect the accuracy of human motion recognition.In order to improve the accuracy of motion recognition,this paper carried out research from the aspects of action feature representation,feature extraction method of recognition model,node update method,model fusion,etc.,designed the corresponding optimization strategy,and built an experimental platform to verify and analyze on the data set.The main work of this paper is as follows:(1)the bone point information captured by the Kinect sensor was used as the action feature representation to avoid the problems of poor interference ability of color video image and large amount of image information calculation.Depth image noise is removed by combined bilateral filtering to obtain more accurate bone point information.(2)by introducing the time convolution network into the continuous motion recognition,the feature extraction mode of the convolution kernel is changed,which solves the insufficiency that the convolutional neural network cannot process the input of timing bone information.At the same time,the feature extraction and classification recognition were carried out under the condition of retaining the joint features of various parts of the human body,which enhanced the expression power of the model.In this method,the residual unit is added to avoid the problem of gradient disappearance,so that the robustness of the deep network is higher,and the time-core of different scales can also take into account the action characteristics of different time lengths to cope with the complex action changes.(3)apply the concept of graph theory to motion recognition,conduct graph modeling for human body continuous motion,and introduce attention mechanism.The model will assign different attention weights to joints according to the context information of the action,and the attention weights of joints with great influence on the action will increase accordingly.Attention weight shows the contribution of different joints in human body to action recognition and enhances the ability to pay attention to local joints in action recognition.(4)the recognition results of the time-convolution network improved by introducing multi-scale convolution kernel and the graph neural network improved by introducing attention mechanism are fused in the later stage,so as to better show the spatio-temporal characteristics in actions.Experiments show that the accuracy of the model improved by introducing multi-scale convolution kernel and attention mechanism is higher than that before the improvement,and the accuracy of action recognition can be significantly improved by using the network integrating multiple models.Experimental data show that the fusion model with attention mechanism can improve the accuracy of human action recognition.