Model-based visual recognition of 3-D objects using pseudo-random grid encoding.

Abstract

This thesis presents a solution to the following problem: Given a set of 3-D objects having their visible surfaces marked with symbols representing the terms of a pseudo-random array (PRA), given the mapping between the planar faces of the polyhedral models of these objects, and given a monocular (2-D) image partly showing an object of this set, determine object identity and its POSE (Position, Orientation and Scale Estimation) parameter relative to the 3-D frame of the video-camera. From the symbols which could be recognized in the image, a broken portion of the PRA is reconstructed and then inspected until a complete pseudo-random window (PRW) is found. Such a window has the uniqueness property which allows the full identification of the window's absolute coordinates (i,j) within the PRA. By consulting the "object model/ PRA" mapping database it is possible to identify the object and its specific face containing the recovered window. Once the object face is identified, its POSE parameters are calculated from the perspective transformations relating the 2-D positions of the recognized symbols in the image frame to their correspondent PRA grid nodes defined in the 3-D object model frame (a classical "camera calibration" problem.)