Crowd Scene Analysis Based On Distance Dependent Nonparametric Bayesian Model

Crowd scene analysis is an important application field of video analysis.Learning motion patterns from crowd scene surveillance video is a major task in the crowd scene analysis.Because of the numerous objects,the diverse scene structures and the subjectivity of people,there are many complex and uncertain motion patterns.Therefore,learning motion patterns from these scenes is important and challenging.The distance dependent nonparametric Bayesian model is outstanding as a unsupervised learning algorithm for it has advantages on learning motion patterns automatically,increasing model size along with data scale and introducing prior information to guide clustering.We propose a new Distance Dependent Chinese Restaurant Franchise(DDCRF)mixture model,introduce a new framework to learning motion patterns from crowd scene based on the DDCRF mixture model,and achieve state-of-the-art results in the experiments.The main works of this thesis:Firstly,an overview of the researches about crowd scene analysis is presented,and the basic framework of learning motion patterns form crowd scene is introduced.Along the normal procedure of crowd scene analysis,related works are analyzed from three aspects: feature extraction,feature expression,and model construction.And a suitable framework is selected to learn motion patterns with richer semantic expression ability in our special task.Secondly,we propose a novel Distance Dependent Chinese Restaurant Franchise(DDCRF)mixture model,which is based on Distance Dependent Chinese Restaurant Process(DDCRP)and Hierarchical Dirichlet Process(HDP).In our thesis,we firstly detail the generating process and learning process of DDCRF,and then create some synthetic data to demonstrate the validity of our learning algorithm.Thirdly,we apply the DDCRF-based framework to many crowd scene datasets.Owe to the advantages of DDCRF,our framework can learn motion patterns from scene automatically and give a semantic expressions of these patterns.