Probability Models And Their Application To Video Object Processing

With the rapid development of computer science and media technology, the computervision technology has been of great concern in various fields. As one of the mostchallenging research focus in the field of computer vision, the video object processingrelate to multiple disciplines, such as computers, electronics, mathematics, artificialintelligence, information processing, and so on. It has been usded in industry, agriculture,military, meteorological and remote sensing widely. Therefore, the reseach of the videoobject processing has great theoretical and practical value. However, there are still anumber of puzzles in the domestic and international research of the video objct processing,which we need to do further work to solve. So we will do more in-depth studies on thevideo object processing with the tool of the probability model based on the others research.The main contents and innovative works in this thesis can be summarized as follows:(1) In order to overcome weaknesses of traditional video object segmentationalgorithms, this thesis proposes a segmentation algorithm of video moving objects based onthe Gaussian mixture model (GMM). Firstly, the video background and foreground will besegmented using the segmentation algorithm according to temporal properties of the video.Then, using the nearest neighbor principle, the video moving object will be segmentedagain based on the result of the first segmentation according to the spatial properties of thevideo. Finally, the moving object can be got accurately after removing noise and deletingthe isolated point in the video.(2) In order to solve the puzzles in the processing of the Gaussian mixture modelparameters estimating and model selecting, the research has been done as follows:1. Animproved EM algorithm has been proposed using the invers Wishart Prior as thepunishment of the likelihood function because of the defect that EM is easy to fall intolocal optimum of the solution space when we estimate parameters of GMM;2. A newGMM parameters estimation method based on differential evolution algorithm has beenproposed to estimate the GMM parameters when the moving object segmentation algorithmdose not need to be real-time;3. A GMM model selection criteria based on nonstationaritymeasure method has been proposed.(3) The research of video objects classification has been done in this thesis, thta is, aminimum-error-rate classification algorithm based on the bayesian rule has been proposedfor video objects. The main idea of this algorithm is that to find an orthogonal lineartransformation to map the d-dimensional feature space of the original video object into m-dimensional feature space, which can minimize the probability of misjudgment.(4) A video face recognition algorithm for certain situations has been proposed.Namely, Because of the defect of the principal component analysis algorithm, which can berun only when the data is a Gaussian distribution, the characteristic matrix of the video faceimage is divided into several sub-blocks to allow each sub-block data to more approximatethe Gaussian distribution. Then, the video face will be recognized with the tool of theminimum risk Bayes rule after main features of the face have been selected using theprincipal component analysis method.(5) A new video text extraction and recognition algorithm has been proposed in thethesis: Firstly, the tags field model and the features field model will be built after theoperation of the multi-resolution decomposition of the video text has been finished. Next,the tag field model will be trained using the Markov Chain Monte Carlo Estimation method.Using the trained model, the inference will be run to get the global optimization of labeldistribution utilizing the maximum a posteriori approach. Then the text in the video can beextracted effectively. Finally, a text recognition algorithm based on support vector machinehas been proposed for extracted texts.(6). In order to overcome defects of the traditional particle filter, this thesis proposed akernel particle filter tracking algorithm based on simulated annealing and multi-featurefusion under the framework of dynamic Bayesian networks. In the algorithm, the annealingproposal distribution of multiple features has been buit and the simulated annealingsampling is used to instead of the traditional priori probability-based sampling.Simultaneously, in the process of state approximation, the color, shape and texture featuresof the video object are used to generate the weighted likelihood function in differentannealing layer by weighting particles. Seen from experiments, we find that the algorithmcan effectively prevent the lack of particles and the target tracking is very accurate.(7) A video human behavior recognition algorithm based on the hidden Markov modelhas been designed in this treatise: Firstly, based on the genetic algorithm and quadraticdistortion measure, a vector quantization scheme is designed and used to deal with meshfeatures of human behaviors extracted from the video. Finaly, we can use the constructedhidden Markov model to reason and recognize human behaviors after the model has beentrained.