Research On Dynamic Textures Segmention And Background Subtraction Based On Improved Mixtures Of Hidden Markov Model

In daily life,a large number of scenes that constitute our visual world are perceived as dynamic textures.Dynamic textures can be regarded as a certain repetitive motion pattern in time and space,which provides important information in image and video processing.Dynamic textures has a wide range of applications in many fields such as medicine,military,industry and environment such as auxiliary disease diagnosis,industrial product quality monitoring,detection and prevention of natural disasters.Dynamic texture segmentation is to segment a video sequence containing multiple textures into nonoverlapping regions,each region showing different uniformity and consistency.In previous studies,Hidden Markov Model(HMM)and Multivariate Hidden Markov Model have achieved good modeling for a single dynamic texture in video sequences.Considering the spatial constraints of dynamic textures,this paper explores the modeling of various dynamic textures in video sequences.The main research contents of this paper are as follows:(1)A dynamic texture segmentation method based on the improved mixtures of hidden Markov model is proposed.In view of the problem that the previous model can only model a single dynamic texture in the video sequence and does not consider the constraints of spatial pixels,then introduce an indicator variable considering space constraints to determine which type of dynamic texture a pixel belongs to and solve the three basic problems of improving the mixtures of hidden Markov model,realize the modeling of various dynamic textures in the video sequence,and verify the dynamic texture segmentation performance of the improved mixtures of hidden Markov model on the relevant data set.The experimental results show that the improved mixtures of hidden Markov model can effectively describe the appearance characteristics and motion patterns of dynamic textures,and improve the segmentation accuracy of related data sets.(2)A generalized SG model based on the improved mixtures of hidden Markov model is proposed.In view of the problem that the SG model does not perform well in scenes with spatiotemporal dynamics and the problem that the improved mixtures of hidden Markov model cannot process video in real time.This paper combines the improved mixtures of hidden Markov model with the SG model and pushes it to its online update method.The model can perform online estimation to adapt to long term changes,and can quickly include new background motion by adding mixed components and discard outdated information by discarding mixture components with small priors.Finally,the background subtraction performance is verified by simulation on relevant datasets.