Indexing based object recognition using projective invariant

Abstract

This thesis explores the use of projective invariants for indexing-based object recognition method. Indexing approaches using projective invariants are very effective when perspective effects cannot be ignored. In some applications, the features of two-dimensional objects may appear as different features in an image due to perspective effects. The central issues that should be considered in designing an indexing-based object recognition system include model representation, the extraction of invariant features, feature matching, hypothesis formation, and object verification. This thesis presents an unified framework for indexing-based object recognition that can be applied to three-dimensional objects as well as two-dimensional objects. The framework consists of model representation based on projective invariants, analysis of error modeled by Gaussian noise, model database construction using hash table, and matching by indexing. For model representation based on projective invariants, this thesis develops a new shape descriptor for 2-D curved objects and a 3-D invariant relationship by single-view for 3-D objects. The descriptor is very stable and not sensitive to noise because it does not include derivatives. The 3-D invariant relationship is more general than the previously proposed invariants. To construct the model-base for indexing by efficient hash-table, error on the proposed invariants is modeled by Gaussian noise. It is used to determine what score (or voting) a correct or incorrect hypothesis is likely to have. To show the usefulness of the method, it applies to three vision problems: "the self-localization of a mobile robot" as the 2-D point problem, "2D curved object recognition" as the 2-D curve problem, and "3D object recognition from single image" as the 3-D point problem. The self-localization and obstacle detection for a mobile robot are very important in vision-based navigation. By applying the method, the self-localization problem is solved ...