Multiresolution Statistical Object Recognition

Model based recognition systems are designed to recognize the occurrence and pose of a known object in an input image that could contain zero or more occurrences of the object. A multiresolution statistical estimation approach to object recognition and pose estimation has been developed which incorporates models for the effects of background clutter, occlusion, and sensor noise. The resulting formulation is essentially an alignment approach in which separate methods are used to align, refine, and verify object poses. First, statistical models for the recognition problem are formulated. MAP estimation is then used to derive an objective function for pose, and the ExpectationMaximization (EM) algorithm is used to perform the non-linear optimization of the objective function. A number of starting values for the EM algorithm are generated by use of angle-pair indexing. Hypothesis testing is then performed on the resulting pose estimate to determine whether the object appears in the image at this pose. An asymptotic approximation to the generalized likelihood ratio test is utilized. In addition, the pose is estimated in a coarse-to-fine multiresolution progression. The pose is first estimated at the lowest resolution level and then used as the starting value for pose estimation at the next higher resolution level. The multiresolution processing results in improved computational efficiency and robustness. A number of experiments are described which demonstrate the multiresolution statistical object recognition system. Thesis Supervisor: Professor W. Eric L. Grimson Title: Associate Professor of Electrical Engineering and Computer Science