Dynamic Texture Segmentation Based on Improved Mixtures of Hidden Markov Model

Abstract

Dynamic texture (DT) can be regarded as a movement pattern with certain repetition in time and space. Dynamic texture segmentation focus on the problem of clustering multiple features and phenomena in time and space into groups, and for each group or region assign unique labels. Hidden Markov Model (HMM) can model visual processes that contain a single dynamic texture, with a major limitation being that they cannot account for the visual process containing multiple simultaneous dynamic textures. This paper explores the mixtures of HMM to model the visual process containing multiple dynamic textures and improves it. Finally we propose an improved mixtures of hidden Markov model (IMHMM). In this model, different dynamic textures are constructed by different HMM components. Considering the regionality of dynamic texture, a new indicator variable considering space constraints is introduced to discriminate the dynamic texture category to which a pixel belongs. Experimental results show that IMHMM is capable of model the visual process containing multiple simultaneous dynamic textures and achieves higher accuracy than existing dynamic texture segmentation methods.