Hidden Conditional Random Fields Based Human Action Recognition In Wide-area Videos

With the development of pattern recognition and artificial intelligence,the human activity analysis as an important research field in computer vision has made consider-able progress.The researches have developed to analyze sophisticated human activities in wide-area videos from simple figure silhouette in binary images.The main challenge in the task is to correctly recognize human action types in a wide-area video with complex large area scene.Wide-area surveillance video is the data of imaging a wide area.Wide-area surveillance video has characteristics of large scene,small target,low resolution and not obvious feature etc.Using a variety of useful information contained in the video image as much as possible is an important way to improve the analysis performance of human activity in wide-area surveillance videos.Hidden conditional random field model?HCR-F?is a probabilistic graph model,which is researched with focus in recent years and show the ability to effectively model and use the temporal and spatial contextual information.The hidden variable layer contained makes the HCRF have strong expression ability,and meanwhile the HCRF has the ability to model the observation image and context infor-mation in labels and observations in a unified framework.Therefore,HCRF model is an effective way to solve the problem of human activity analysis in the wide-area surveil-lance video.This thesis mainly focuses on the research of HCRF for modeling spatial and temporal contextual information in the wide-area surveillance video.The following three aspects of research work have been carried out:Firstly,HCRF model is introduced into the human activity recognition in the wide-area surveillance video.The methods of model construction,model training and mod-el inference are analyzed in detail.Typical wide-area surveillance video databases of UT-TOWER and Rooftop are used for the comprehensive test of the research methods.Through experiments,the factors affecting the expression and recognition ability of the model involved in HCRF model are analyzed in detail and compared with the existing methods with excellent performance.Experimental results show that the HCRF model achieves the results not only comparable to that of SVM without using the contextual in-formation,but also better than HMM models using contextual information.The research demonstrates the feasibility of using HCRF for the human activity recognition in the wide-range surveillance video.Secondly,a l_1/2regularization HCRF model is proposed for human activity analy-sis.The importance of HCRF model training with regularization is analyzed firstly and the general framework of HCRF model training with regularization is given.Then,using l₂and l₁ regularization in HCRF model training as examples,the thesis gives the anal-ysis of the performance of the HCRF model obtained by regularization training to solve the overfitting and obtain the model sparsity.Based On the analysis,a l_1/2regulariza-tion training method is introduced and a l_1/2regularization HCRF model is proposed for human activity recognization.The comprehensive experiments demonstrate that the pro-posed l_1/2regularization HCRF model can further improve the sparsity and recognition performance.Finally,through analyzing difficulties in human activity recognition in the wide-area surveillance video,the usual HCRF training method could make many nodes in the hidden layer of the HCRF model perform similarly.This could leads to the redundancy of the model and thus reduce the depiction ability and recognition performance of HCRF model.For this reason,in the basic framework of regularization training,this thesis will introduce a diverse promoting distribution of model parameters and then propose a new diversified HCRF model for human activity recognition.The proposed model will further improves the depiction and recognition performance.