Based On Data-driven Markov Chain Monte Carlo Model Of The Dynamic Texture Analysis

Dynamic textures are sequences of images of moving scenes that exhibit certainstationary properties in time. For example, flowing water, dancing flag, joggling leaves.This paper uses a generative model to characterize dynamic texture, and describes how touse Data-Driven MCMC to learn and inference for the model. To check the validation ofthis model, the paper presents the method of classifying, tracking and editing of dynamictexture based upon this model. The main works are as follows:(1) Modeling of dynamic texture: By Analyzing the advantages and disadvantages ofvarious models of dynamic texture, a generative model is adopted. The parameters arelearned and inferred by Data-Driven MCMC.(2) Editing of dynamic texture: By learning the parameters of the model, any desiredlength of sequences of images can be synthesized, and the statistics of new dynamictexture are equivalent to the original, this shows that the generative model captures theessence of dynamic texture.(3) Classifying of dynamic texture: Based on the generative model, differentparameters corresponding for different dynamic texture are learned. Then the similarity ofhidden graph between these patterns is defined to classify them.(4) Tracking of moving object using dynamic texture: The interested object isconsidered as a dynamic texture. By learning the model parameter of the dynamic texture, the object can be tracked. A sample of human tracking is introduced to demonstrate theapplication of dynamic texture. This simple case shows the ability of the generative modelin tackling tracking problem.