Locating geometric primitives by pruning the parameter space

This paper examines the detection of geometric primitives using 2D edge pixels or 3D range points. The geometric primitives are described by parametric equations in the image space. An explicit error model allows the primitives to be extracted robustly, while a hierarchical search of the parameter space with conservative pruning allows the primitives to be located e$ciently. The result is an e$cient search strategy that is robust to distractors, missing data and noise, and that does not require an initial estimate of the positions of the geometric primitives. We apply these techniques to circle detection, for locating craters on planetary bodies and analyzing engineering drawings, and to cylinder detection, for "nding unexploded ordnance in test ranges. 2001 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.