Research On Human Action Recognition Based On Graph Convolutional Neural Networks

As the main object of expression in images and videos,the study of allowing machines to automatically recognize human action behaviors from images and videos has very important research value,and it has also been a hot issue in the field of computer vision research.Human behavior recognition research has broad application prospects in video surveillance,medical care,smart home,and human-computer interaction.The human body's action performance is mainly completed by the mutual traction and cooperation between the skeleton and the joint points.Therefore,the human skeleton joint diagram contains rich action feature information,and the skeleton information is very robust to changes in scale,illumination,and perspective.Sex.In video streaming tasks,the time series dimension is relatively high,the modeling methods of time series skeleton features and spatial skeleton features have been immature,the model calculation is large and the recognition accuracy is not high.This paper conducts in-depth research on the above problems and proposes two graph neural network models.One is to introduce the hole convolution and spatial pyramid pooling module in the spatio-temporal graph convolutional neural network,and the other is the fusion graph convolutional neural network and long The two-stream network method2S-LSGCN of short-term memory neural network,the main work of thesis paper is as follows:(1)Summarize and analyze the human behavior recognition methods based on deep learning,as well as the common methods and data sets of human behavior recognition based on skeleton joint points,and perform certain classification processing.And for the acquisition of skeleton joint points in the video stream and the construction of skeleton joint diagrams,a general introduction is made.(2)A simple and lightweight model is proposed,which can effectively solve the problem of insufficient time features and insufficient recognition accuracy in the action joint diagram.First of all,the model in this paper introduces hole convolution on the basis of spatio-temporal graph convolutional neural network to replace the ordinary time convolution kernel in traditional models,and uses perforated convolution to extract timing features for behavior recognition to extract larger frames Timing characteristics of the time range.Secondly,the video of simple actions has a short time and lacks sufficient timing features,which makes the recognition accuracy rate low.Because the pyramid pooling has different sampling rates and multiple fields of view,it can capture the spatial and temporal characteristics of objects at multiple scales.This model introduces the spatial pyramid pooling module to compensate for the lack of timing features and improve the recognition accuracy of the model.(3)Propose a two-stream network method 2S-LSGCN that combines graph convolutional neural network and long-short-term memory neural network.This method extracts the spatial and temporal features of the action from the skeleton joint diagram composed of human joint points.And use GCN to extract the potential spatial information between the skeleton joint points,use LSTM to extract the time series features between before and after the human action as a supplement,and finally merge the prediction outputs of the two networks separately to solve the insufficient generalization ability of a single network.In order to solve the problem of space complexity and time complexity in complex actions.