Object recognition and pose estimation using appearance-based features and relational indexing

This dissertation addresses the problem of recognition and pose estimation for industrial parts having planar, cylindrical and threaded surfaces. The vision system developed falls into the model-based category, where all the objects to be recognized are known a priori and are stored in a database. The recognition paradigm used is hypothesize-and-test, where model hypotheses are generated and then tested during a verification step. The models are defined in terms of their appearance-based features, which are the features of the 3D objects that are reliably detected in 2D images, and in terms of the relationships between these features. This relational description of models is used in a new matching paradigm called relational indexing that generates hypotheses of the models present in the image of an unknown scene. The hypotheses generated are tested by a verification procedure that makes use of a new linear method for computing object pose by simultaneously utilizing sets of point correspondences as well as ellipse-circle correspondences. A modified Hausdorff distance measure is utilized to rule out incorrect model hypotheses and to confirm correct ones. The system was tested on a large set of real images of single and multiple objects, with some objects occluding others and the results obtained are promising.