On recognizing and tracking 3D curved objects from 2D images
Jin-Long James Chen · Michigan State University Libraries · 1996
This thesis addresses the problem of recognizing arbitrary curved 3D objects from single 2D intensity images. We propose a model-based solution within the alignment paradigm involving three major schemes--modeling, matching, and indexing. The modeling scheme consists of constructing model aspects for predicting the object contour seen from any viewpoint. The indexing scheme generates from image features hypotheses giving candidate model aspects and poses. A hypothesis grouping and ordering method is used to order model hypotheses based on prior knowledge of pre-stored models and the visual evidence of the observed objects such that the most likely model hypotheses are tested first. The matching scheme aligns candidate model edgemaps to the observed object edgemap and the results of alignment are used to support/refute model hypotheses. Due to the unavailability of salient features in objects with sculptured surfaces, matching is carried out by the Newton's method with Levenberg-Marquardt minimization. A hierarchical verification strategy is incorporated in the matching scheme to quickly eliminate false model hypotheses from further consideration. Once a correct model is localized, a recognition success is declared and the whole verification procedure is terminated. If all candidate model hypotheses are refuted, a recognition failure is reported. When combined into an integrated system, these three schemes make progress toward improving accuracy and efficiency by pruning false model hypotheses and minimizing unnecessary verification tests. A prototype implementation has been tested in experiments conducted on a database containing 658 model aspects to evaluate the performance of recognition on 20 arbitrary curved objects, either non-occluded or partially occluded, seen from several viewpoints. Bench tests and simulations show that a large database of many kinds of objects including polyhedra and sculptured objects can be handled accurately and efficiently. We have also applied this model-based alignment paradigm to tracking a single moving object in a scene from an image sequence. Experimental results indicate the viability of this tracking method. From these results, we conclude that the proposed recognition-by-alignment paradigm is a viable approach to object recognition and tracking.