Research And Design Of Human Behavior Recognition Based On Video Stream

With the development of the research of artificial intelligence technology and theory,computer vision has become one of the key research directions for the majority of scientific researchers.This field involves intelligent hardware,intelligent control,electronic information,virtual reality and so on.In this paper,human behavior recognition as the main research object,the human behavior recognition research can greatly promote the development of intelligent monitoring industry,using intelligent algorithm instead of human work,can greatly reduce the work burden,effectively prevent the crime.In addition,the research results of human behavior recognition have broad prospect and application value in security,virtual reality,athletes'auxiliary training,biological medical treatment,human-computer interaction and other fields.Firstly,the research status and significance of human behavior recognition and target detection are introduced in detail.Secondly,according to the traditional research process,target recognition and tracking methods are analyzed one by one,and some methods are compared and verified.The Hu invariant moments and Fu Liye descriptors for human body feature extraction are introduced in detail.The pattern of classical classifier combined with behavior recognition algorithm has very important significance in the field of behavior recognition.Therefore,the article focuses on the analysis of SVM,K-means,Softmax classifier and HOG,SIFT,Haar feature extraction methods,and puts forward the solution of human motion information description and feature extraction.In the end,the paper gives a brief introduction to the deep learning model and its key ideas and principles,and focuses on the analysis of the CNN depth learning model.In this paper,the 3D convolution neural network model and LSTM are used to test the UCF-Sports,KTH and Weizmann data sets.The experimental results verify that the convolution neural network model can detect the human behavior very effectively.Based on the application and understanding of CNN structure,a kind of VGG deep convolution neural network model structure is proposed.First,the Keras rapid development framework and Tensorflow depth learning platform are built in the experiment.Secondly,the training model and prediction model are established for the three data sets of human behavior recognition respectively.Each model automatically generates visual network structure through tool methods.Finally,data preprocessing,data training,and data testing and analysis are performed on three standard human behavior recognition databases.The experimental results show that the method can train a large number of self-learning video streaming human behavior characteristics through data.It can be seen from the characteristic channel graph of the inner coiling layer of these deep networks.The class VGG depth network can identify the characteristics of the human body and the background accurately,compare with the methods of traditional handmade features combined with all kinds of classifier,having more expressiveness and robustness.The experimental data analysis curve shows that the method can converge at a faster speed and achieve higher accuracy.