Probabilistic modeling of texture transition for fast tracking and delineation

In this thesis a probabilistic approach to texture boundary detection for tracking applications is presented. We have developed a novel fast algorithm for Bayesian estimation of texture transition locations from a short sequence of pixels on a scanline that combines the desirable speed of edge-based line search and the sophistication of Bayesian texture analysis given a small set of observations. For the cases where the given observations are too few for reliable Bayesian estimation of probability of texture change we propose an innovative machine learning technique to generate a probabilistic texture transition model. This is achieved by considering a training dataset containing small patches of blending textures. By encompassing in the training set enough examples to accurately model texture transitions of interest we can construct a predictor that can be used for object boundary tracking that can deal with few observations and demanding cases of tracking of arbitrary textured objects against cluttered background. Object outlines are then obtained by combining the texture crossing probabilities across a set of scanlines. We show that a rigid geometric model of the object to be tracked or smoothness constraints in the absence of such a model can be used to coalesce the scanline texture crossing probabilities obtained using the methods mentioned above. We propose a Hidden Markov Model to aggregate robustly the sparse transition probabilities of scanlines sampled along the projected hypothesis model contour. As a result continuous object contours can be extracted using a posteriori maximization of texture transition probabilities. On the other hand, stronger geometric constraints such as available rigid models of the target are directly enforced by robust stochastic optimization. In addition to being fast, the allure of the proposed probabilistic framework is that it accommodates a unique infrastructure for tracking of heterogeneous objects which utilizes the machine learning-based predictor as well as the Bayesian estimator interchangeably in conjunction with robust optimization to extract object contours robustly. We apply the developed methods to tracking of textured and non textured rigid objects as well as deformable body outlines and monocular articulated human motion in challenging conditions. Finally, because it is fast, our method can also serve as an interactive texture segmentation tool.