Research On Skeleton Action Recognition Based On Graph Neural Network

Human behavior recognition has always been an important research content of computer technology,and an important part of machines to understand human behavior.However,the traditional behavior recognition algorithm is mainly based on the RGBD information of the image sequence for feature extraction.The behavior recognition algorithm model has poor robustness,low recognition accuracy or high complexity.This article is based on graph convolutional neural network and human bone data,A fast behavior recognition algorithm of human skeleton graph based on graph convolutional neural network is proposed.The research content is as follows:First of all,in view of the uneven distribution and insufficient features of the original human bone data,which leads to low recognition accuracy,a behavior recognition algorithm based on Graph Convolutional Networks(GCN)based on data preprocessing is proposed.The coordinate information of the joint points in the original human skeleton relative to the world coordinate system is converted into the relative coordinate position relative to the center of the human body,so that the bone point data is evenly distributed.In order to extract feature information that can better characterize the human body’s movements,the speed characteristics of the human body’s joint point coordinates are increased by subtracting the corresponding joint point coordinates of two adjacent frames.The experimental results show that the proposed method can automatically capture the structure of the spatial joints of the human skeleton diagram,enhance the data of the human skeleton diagram,and improve the accuracy of the behavior recognition model.Secondly,in view of the problem that the spatial graph convolution cannot obtain the information of the adjacent nodes in the next frame,which leads to the problem that the model accuracy is not high,a behavior recognition method based on 3D graph convolution is proposed.Extend the 3D convolution method to 3D image convolution,and add time series convolution operations to the human skeleton space map convolution to ensure that each image convolution operation is not limited to the human body in each frame The skeleton diagram also contains the information of the next frame.In order to expand the range of the receptive field during feature extraction,the image convolution operation is performed in the way of hole convolution.At the same time,in order to better train the model,the SGDR method is used to update the learning rate when training the behavior recognition network model.The experimental results show that the 3D graph convolutional network model designed in this paper has higher accuracy than the previous behavior recognition model.Finally,in view of the high complexity of the model,a fast behavior recognition algorithm based on a lightweight network model is proposed.In order to simplify the complexity of the algorithm model,the model is optimized and simplified on the basis of knowledge distillation,and two loss functions composed of the real label and the predicted value and the teacher label and the predicted value are constructed,and the two loss functions are designed between the two loss functions.An appropriate weight is used as the final loss function.The experimental results show that the complexity of the proposed algorithm model has been greatly improved.Under the premise of ensuring the accuracy of the model,the model parameters and the amount of calculation are reduced,and the rapidity of behavior recognition is improved.