Graphical models for video analysis

Research on video understanding and modeling is becoming an increasingly popular research topic as demand for algorithms that deal with video and computation power increases. To enable computers to achieve video understanding, we need to build robust computer vision algorithms that can operate without supervision. Numerous methods have been developed to achieve video understanding in computers. Although rich and complex information is embedded in video, there is also a large amount of repetition, noise and variability. This dissertation investigates new approaches to the modeling of video and to video analysis that utilize graphical models. A graphical model is a probabilistic model that can be represented by a finite graph, in which each node is a variable and each edge represents the probabilistic dependence between random variables. One of the major obstacles to the use of such complex graphical models is the computational complexity that they would require. Graphical models are useful for modeling images and video, as they provide principled solutions and enable easy encoding of spatial relationships and photometric and geometric properties of pixels, regions, interest points, and other features. Therefore, this work is concerned with the design of fast, robust graphical models capable of unsupervised performance. The validity, effectiveness, and potential uses of the proposed approaches are verified with experiments that use real video sequences in several important applications, such as stabilization, layering, video surveillance, clustering and light modeling.