Dynamic Texture Classification Using Student's t-distribution based Hidden Markov Model

Continuous hidden Markov model with Gaussian mixture model as the emission distribution is prominent in modelling sequential data. It is well known that Gaussian distribution confronts the bottleneck of sensitivity to outliers. However, student’s t-distribution is heavily tailed and more robust. Therefore, the student's t-distribution based hidden Markov model (SHMM) is useful for analyzing the sequential data with outliers. In this paper, the dynamic texture is considered as a set of the observed data sequences generated with an SHMM model. The motion and appearance properties of the dynamic texture are characterized with the hidden state sequence and the emission distribution, respectively. Then the parameters of the model SHMM can serve as the features to describe the dynamic texture. Finally, we use the maximum likelihood criterion to categorize the dynamic texture to be classified. The classification experiments on DynTex dataset and UCLA dataset demonstrate the superiority of the proposed method.