Research On Application To Video Scene Analysis Of Hidden Markov Model

Computer vision mainly studys on how to make the computer apperceive and understand a single 2D image or image sequences formed by various 2D images transformed from 3D physical space. Because a single image often does not reflect more information, analysis and processing of video composed of image sequences has been a hot issue of research in computer vision.A video scene describes a complete logical story element, so scene analysis has a greater practical significance. In the video scene, we often pay more attention to the foreground activity in the scene, which can been understood by a combination of the background and foreground information which can be extracted through some ways. This paper mainly includes the following subjects:Firstly, as the theoretical basis of the analysis and processing of video scenes, this paper reviews the probabilistic graphical model which is widely used in this field and then gives a detail introduction to the hidden Markov model and related algorithms.Secondly, taking into account of the computational complexity, the accuracy of modelling analysis and the quality of result images obtained after video analysis, this paper preprocesses each frame image of the video scene by using algorithms related to image processing and (such as image scale transformation and image denoising).Lastly, this article models the time sequences composed by the pixel gray level in each position by using a forest of HMMs and extracts the static information and the dynamic information of the video scene by analyzing these models. The static information represents the background model of the scene and the dynamic information is the area of the video sequence which is affected by foreground activity.The video background model is segmented by a clustering algorithm based on Hidden Markov model in this paper, distinguishing the areas which are the same in terms of space and different in the aspect of time. The activity of objects are detected by using the weighted entropy strategy based on the stationary probability distribution. This paper highlights constituting a unified probabilistic framework for the scene analysis and understanding, and the algorithm only using the pixel level temporal behavior of each location in the scene sequcence images.