Statistical object recognition

To be practical , recognition systems must deal with uncertainty. Positions of image features inscenes vary. Features sometimes fai l to appear because of unfavorable i l lumination. In this work, methods of statistical inference are combinedwith empirical models of uncertaintyinorder to evaluate andre ne hypotheses about the occurrence of a knownobject in a scene. Probabi l i stic models are used to characterize image features and their correspondences. A statistical approach is taken for the acquisi tion of object models from observations in images: Mean Edge Images are used to capture object features that are reasonably stable with respect to variations in i l lumination. The Al ignment approachto recognition, that has beendescribedbyHuttenlocher andUl lman, is used. The mechanisms that are employedto generate initial hypotheses are distinct fromthose that are used to veri fy (and re ne) them. In this work, posterior probabi l i ty and MaximumLikel ihood are the cri teria for evaluating and re ning hypotheses. The recognition strategy advocated in this work may be summarizedas Align Re ne Veri fy, whereby local searchinpose space is uti l izedto re ne hypotheses fromthe al ignment stage before veri cation is carried out. Two formulations of model -based object recognition are described. MAPModel Matching evaluates joint hypotheses of match and pose, whi le Posterior Marginal Pose Estimation evaluates the pose only. Local search in pose space is carried out with the Expectation{Maximization (EM) algorithm. Recognition experiments are describedwhere the EMalgorithmis used to re ne andevaluate pose hypotheses in2Dand3D. Initial hypotheses for the 2Dexperiments were generated by a simple indexing method: Angle Pair Indexing. The Linear Combination of Views method of Ul lman and Basri i s employed as the projection model in the 3Dexperiments. Thesis Supervisor: W. Eric L. Grimson Title: Associate Professor of Electrical Engineering andComputer Science 2