Research On Abnormal Behavior Recognition Method Based On Graph Convolutional Network And Dynamic Mode Decomposition

With the development of hardware computing and the popularization of many deep learning methods,behavior recognition algorithms are gradually applied to reality,and bone based behavior recognition algorithm is a research field in behavior recognition,which attracts the attention of many researchers.As we all know,traditional methods can quickly detect abnormal behavior in the case of a few samples,but the accuracy is insufficient.The deep learning method can get high accuracy,but it needs lots of samples for training and a large amount of calculation.This inspires us to combine traditional methods with deep learning methods,so that the advantages of traditional methods and deep learning methods complement each other.Therefore,this thesis proposes a traditional method based on dynamic mode decomposition and a deep learning method based on graph convolution network and Transformer,and uses the traditional method to assist the deep learning method in the case of a few samples.The main work of this thesis is as follows:（1）In the aspect of traditional methods,a behavior recognition method based on dynamic mode decomposition is proposed.The algorithm uses dynamic mode decomposition to extract action features from structured bone data,and uses one class support vector machine（OCSVM）to classify sample into positive sample and negative sample by the extracted features.Compared with deep learning methods,traditional methods have great advantages in speed,and can achieve 90.8 % accuracy by using only a few samples.（2）In the aspect of deep learning methods,an algorithm based on graph convolution network and Transformer is proposed.Firstly,referring to the Transformer encoder,a behavior recognition algorithm based on Transformer is proposed and class token is introduced.In order to make full use of human bone information and further improve the accuracy,the graph convolution network is introduced.The graph convolution network is used to replace the embedding layer in the Transformer encoder,and the GCN（Graph Convolutional Network）-Transformer algorithm is proposed.Then,the algorithm is optimized,and the graph convolution network and Transformer encoder structure are constructed into a block.When the number of samples is small,the dynamic mode decomposition algorithm is used to assist the deep learning method in this thesis.The experimental results show that the accuracy of GCN Transformer algorithm on the NTU-RGB+D dataset is 85.4 %.Combined with the auxiliary training of the Dynamic Mode Decomposition（DMD）algorithm,the accuracy is improved by about 1 % on small-sized dataset.（3）The algorithm is applied to the real-world application.Firstly,the algorithm model in this thesis is pretrained by using a large-sized dataset,and fine tuned in the elevator scene dataset.Then,the elevator video abnormal behavior identification security system is designed.The trained algorithm is applied to the system.The interface is designed by using the Tkinter package.The camera parameters are controlled by the ONVIF protocol.The video stream is obtained by the RTSP protocol.The video stream is processed and displayed by Open CV.The experimental results show that the algorithm can achieve 92.7 % accuracy when it is applied to the elevator scene dataset.In the system design,the delay of the algorithm module is 155 milliseconds,which can meet the real-time requirement.