Research On Human Action Recognition Method Based On Decoupling Graph Convolutional Network

In the field of computer vision,human action recognition is a very challenging problem.In recent years,with the development of mobile network technology and the popularization of video acquisition equipment,human action recognition technology has a wider range of application scenarios and greater economic value,and has gradually received more and more attention.The input data type of the action recognition method can be divided into two types:image sequence data and skeleton sequence data.Compared with image sequence data,skeleton sequence data has the advantages of small data volume and high robustness in complex scenes.Early skeleton-based action recognition methods organize skeleton sequences into joint coordinates or pseudo-images,and predict action labels through recurrent neural networks or convolutional neural networks.However,because the skeletal data is in non-Euclidean space,such features cannot repeatedly reflect the topological structure of the human body,which ignores the internal connections between joints.In recent years of research,researchers have introduced graph convolutional networks into skeleton-based action recognition tasks,and extracted action features in sequence images through graph convolution and temporal convolution.These works extend the convolution from images to graphs of any size and shape,and have achieved fruitful results in the field of action recognition.This article uses an advanced decoupling graph convolutional network for action recognition,instead of the traditional coupled graph convolutional network method.In this article,a STC attention module is added before the temporal convolutional layer of the network.This module connects the three attention modules of spatial,temporal,and channel,so that the network pays more attention to learning the key information in the skeleton feature.After the decoupling graph convolutional layer of the network,this article proposes an adaptive normalization module to replace the original batch normalization layer.The adaptive normalization method combines batch normalization and layer normalization.Through training,the importance weights of the two types of normalization methods are adjusted to achieve a balance between learning and generalization capabilities,and play a better normalization effect.In order to prove the effectiveness and advancement of the improved network in this article,this article conducts experiments on the NTU-RGBD and NTU-RGBD-120 datasets.The experimental results prove that the improved network in this article improves the effect of action recognition,and it is competitive compared with the recent advanced action recognition methods.