Human Action Recognition Research Based On Deep Learning

In the field of computer vision,video human behavior recognition has always been a popular research direction.Traditional human behavior recognition methods are difficult to make major breakthroughs in practical applications due to inherent flaws.With the rapid development of computer software and hardware,the deep learning theory proposed in 2006 has been put into application,and video human behavior recognition has once again been paid attention.Nowadays,video human action recognition is widely used in video monitor,smart home,somatosensory interaction and other fields,it plays an important role.Starting from the theory of deep learning,this article combines the latest research results to conduct experimental research on the following two points.On the basis of the 3D network,add dense and residual structure,and propose the dense residual network(R-Densenet).Compared with the 2D convolution kernel,the 3D convolution kernel adds a depth dimension to extract the timing information,which can effectively extract the timing information of the action,and has a good ability to recognize some similar actions.The complex network structure of the 3D network results in a large amount of calculations,slow operation,and the network is difficult to reach very deep depths,and can not give full play to the advantages of deep learning.Therefore,a dense residual structure is proposed on the basis of the 3D network,and the dense residual network structure is adopted.To solve the problems of gradient disappearance and network degradation,at the same time,jumping links can effectively reduce the parameters of the network,making the network more portable and conducive to reaching a deeper depth.Experiments show that the improved 3D convolutional network can effectively extract timing information,and through the dense network,the structure of the residual network further speeds up the operation.Research on video human behavior recognition based on graph convolutional network.Different wiht RGB based recognition,the input of graph convolutional is characteristic root.the graph is constructed with a topological map of human bones.In the previous graph convolution method,the division strategy of skeletal nodes is relatively simple,focusing on the overall position of the nodes in the skeleton and ignoring the importance of specific parts in behavior recognition.This paper proposes a new partitioning strategy for skeletal nodes,processing nodes in different regions separately,and conducting experiments on a data set to prove the effect of the network.The experimental results show that the network can effectively extract the local features of the bone node image,avoid the interference between similar actions,and improve the accuracy of the model.