Probabilistic Graphical Model And Its Application In Video Segmentation

Video Segmentation is the basis of many computer vision applications, such as location and recognition of moving objects. Segmentation becomes difficult when the background of video is complex because of the changing illumination and the moving objects. The main obstacle of video segmentation is how to establish a model for the changing background. This model is explicitely used to describe the change of the background and to evaluate the background state according to the video input.This paper proposes a unified framework for detecting and segmentation of foreground objects in complex scenes involving swaying trees, moving shadows and ocean waves etc. A mixture model, which includes hidden Markov models and undirected probabilistic graphic models, is used to describe the dynamic background. We suppose the background variations between neighboring pixels have strong correlation, which is also known as"co-occurrence". This property is used to initialize and update the mixture model. Forward algorithm and Bayesian belief propagation algorithm allow us to efficiently calculate the maximum of the posterior probability of background with input data. The result of the experiment has shown this method has effective and robust performances in video segmentation, especially with complex background.At the beginning, this paper introduces the appearance and development of the Bayesian network. Then it gives a brief review of correlated concepts and theories. Basing on these theories, methods of using Markov Random Fields and Mixed Probabilistic Models for video object segmentation are proposed. The algorithms are described in details.