Robust Extraction of Vertices in Range Images by Constraining the Hough Transform Dimitrios Katsoulas Institute for Pattern Recognition and Image Processing, University of Freiburg Georges-Koehler-Allee 52, D-79110 Freiburg, Germany dkats@informatik.uni-freiburg.de Abstract. We describe a technique for extracting vertices from range images of cluttered box-like objects. Edge detection is performed and an edge map is acquired. Extraction of vertices is carried out using the edge map and comprises two steps: Linear boundary detection in 3D and boundary grouping. In order to recover the four parameters of a 3D linearsegment,wedecomposetheproblemintwo2D subproblems, each recovering two line parameters. These subproblems are solved by means of the Hough Transform, constrained in this way so that accurate and efficient propagation of the edge points localization error is achieved. Pairs of orthogonal boundaries are grouped to form a vertex. The or- thogonality of a boundary pair is determined by a simple statistical test. Our strategy comprises many advantages, the most important of which robustness, computational efficiency and accuracy, the combination of which is not to be found in existing approaches. 1 Introduction Automatic unloading and sorting of piled objects is of great importance to the industry, because it undertakes a task that is very monotonous, strenuous and sometimes quite dangerous for humans. Objects which are often encountered in industrial sites and distribution centers are mainly rigid boxes as in Fig. 4 (a) or deformable box-like objects (sacks) full of material as in Fig. 5 (a). It is advantageous to employ range imagery for dealing with the problem mainly due to relative insensitivity on lighting conditions and object texture. It is since years known in the computer community that a three-dimensional visible vertex provides the strongest constraints for accurately determining the position of convex, three-dimensional objects and thus are very good approximations of the location of the objects in space. Since the objects we are dealing with are either boxes or box-like objects, their vertices can still be used for generating accurate object location hypotheses. For this reason the robust and accurate detection of object vertices in range images is of extreme importance to this application. Although a variety of methods for detecting corners in intensity images have been reported, this is not the case for range images. The majority of the existing approaches (like [3],[1] and others) use region information to extract vertices. The disadvantage is that the objects need to expose more than one surfaces to F.J. Perales et al. (Eds.): IbPRIA 2003, LNCS 2652, pp. 360–370, 2003. c  Springer-Verlag Berlin Heidelberg 2003