Statistical approaches for partially occluded object recognition

Occlusion is a pervasive and difficult problem in object recognition. It arises in many problems due to the interference between objects in a scene, such as the overlay of trees and vehicles. It degrades the recognition performance and makes the object recognition task significantly more complex because of the unknown nature of occlusion.; In this dissertation, we develop a statistical theory and corresponding algorithms for recognition of partially occluded objects, without explicit recognition of occluding objects. In feature-based object recognition, scenes to be analyzed are represented by scene features, and object models are described by sets of object features. We propose a statistical framework for representing the relationship between the scene features and the underlying object. This framework models occlusion, noise and clutter statistically. We introduce two occlusion models: an independent model and a Markov Random Field (MRF) model.; Based on the statistical framework, object recognition is posed as a maximum likelihood classification problem. This problem is complicated by the unknown correspondence between the scene and object features. We use a generalized maximum likelihood approach, which requires finding the likelihood of the best one-to-one correspondence between the scene and object features. We develop efficient combinatorial algorithms for finding the optimal correspondence.; An important issue in object recognition is that objects can appear with variable pose; that is, objects may be rotated, translated, scaled or deformed from their models. We extend our statistical framework for partially occluded object recognition in the presence of unknown 2D affine pose transformations. We treat the pose as a nuisance parameter in the generalized maximum likelihood approach. The recognition problem reduces to finding the optimal correspondence and pose for computing the likelihood. We develop a two-step algorithm for finding the optimal correspondence and pose. We extend the approach to additional object variations such as articulation and deformation.; Our proposed statistical models and algorithms can be used to unify and extend other algorithms for object recognition. In particular, we use the MRF occlusion model and algorithms to extend several statistical object recognition algorithms proposed in the recent literature.; We illustrate the advantages of our algorithms using both simulated data and experimental data from optical and synthetic aperture radar imagery.