Robust algorithms for model-based object recognition and localization

TL;DR: The idea of tolerance is introduced which measures the robustness of a recognition and localization method when noise is allowed and a localization algorithm that achieves any desired tolerance in a relatively low order worst case asymptotic running time is presented.

Abstract: We consider the problem of model-based object recognition and localization in the presence of noise, spurious features, and occlusion. We address the case where the model is allowed to be transformed by elements in a given space of allowable transformations. Known algorithms for the problem either treat noise very accurately in an unacceptable worst case running time, or may have unreliable output when noise is allowed. We introduce the idea of tolerance which measures the robustness of a recognition and localization method when noise is allowed. We present a collection of algorithms for the problem, each achieving a different degree of tolerance. The main result is a localization algorithm that achieves any desired tolerance in a relatively low order worst case asymptotic running time. The time constant of the algorithm depends on the ratio of the noise bound over the given tolerance bound. The solution we provide is general enough to handle different cases of allowable transformations, such as planar affine transformations, and scaled rigid motions in arbitrary dimensions. Thesis Supervisor: Sanjoy k. Mitter Title: Professor of Electrical Engineering