Hidden Markov Model Based Dynamic Texture Classification

Dynamic textures are video sequences of complex dynamical objects,which exhibit certain stationary properties in time.They provide very important visual cues for various video processing problems.Dynamic texture classification is to identify the types of regions or objects using the dynamic texture properties,which is a very important link in dynamic textures analysis.Despite the effort,dynamic texture classification is still an interesting and challenging research field.In traditional research methods,the dynamic texture is often considered to be the output of a linear dynamical system(LDS).Although this approach can combine motion and appearance properties of a dynamic texture,the LDS with continuous latent and observed variables is a linear-Gaussian model,which is inconsistent with the actual performance of the dynamic texture.Fortunately,hidden Markov model(HMM)has the same structure with LDS.Furthermore,the latent variables are discrete but with arbitrary emission probability distribution.Therefore,the HMM will be more suitable for dynamic texture description.The main work contents of this paper can be concluded as follows:1.To address the numerical underflow problem appears in higher-order model parameter reestimation process,the scaling factor method belongs to classical theory of HMM is extended to the higher-order HMM.Meanwhile,a brief discussion on the model decoding problem is given.A general Viterbi algorithm that is applicable to higher-order HMM is proposed in terms of the first-order and third-order Viterbi algorithm theory.2.This paper proposes a novel dynamic texture classification method based on maximum likelihood(ML)criterion.Specifically,the pixel intensity sequence along time of a dynamic texture video is modeled with a HMM that encodes the appearance information of the dynamic texture with the observed variables,and the dynamic properties over time with the hidden states.The rationale behind this model is based on the observation that the arbitrary emission probability distribution and the higher-order dependence of hidden states of a higher-order HMM will result in better representation of a dynamic texture.Subsequently,the ML classification criterion is applied to classify the category of a new dynamic texture sample.At last,the classification performance is analyzed by comparing with the LDS based method.3.This paper proposes a similarity-based dynamic texture classification scheme.According to the idea of bag-of-systems framework,the K-Means and K-medoids clustering algorithms are employed to obtain representative models aftering using HMMs to model dynamic textures,and then the definiton of similarity between different sequences is given for extracting feature vectors to describe dynamic textures.The K-nearest neighbor classifier and support vector machine classifier are adopted to classify the testing samples,respectively.Finally,the effectiveness of the classification scheme is verified by simulation experiments.