Dynamic texture classification based on bag-of-models with mixture of student’s t-hidden Markov models

Abstract

Hidden Markov model (HMM) is a popular statistical approach for modeling sequential data comprising continuous attributes, and it has been applied in finance, machine vision and other fields. Due to higher tolerance to outliers than Gaussian distribution, the multivariate student’s t-distribution is used to serve as the observation emission distribution in continuous HMM (SHMM), which has been exploited to improve the performance for modeling sequential data. In this paper, a bag-of-models method based on the mixture of SHMMs is proposed to describe dynamic texture for dynamic texture classification, in which codebook is constructed with SHMMs. Specifically, a novel mixture model, mixture of SHMMs, is proposed to model and cluster the observation sequences of the dynamic texture. The parameter learning method is derived by using the expectation maximization (EM) algorithm. Then all components of the mixture of SHMMs are assembled to obtain the codewords and the bag-of-models based features are extracted for describing the dynamic texture. Finally, we demonstrate the effectiveness of the bag-of-models based on the mixture of SHMMs for dynamic texture classification on different benchmark datasets. The experimental results show the potential performance of the proposed approach by comparing with some existing dynamic texture classification methods.