Geometric Matching for Free-Form 3D Object Recognition Heinz Hügli, Christian Schütz, Dimitrios Semitekos University of Neuchâtel Institute for Microtechnology Rue Tivoli 28 CH-2003 Neuchâtel Switzerland hugli@imt.unine.ch KEYWORDS: free-form 3D object recognition, pose estimation, range imaging, closest point matching Abstract This paper investigates a new approach to the recognition of 3D objects of arbitrary shape. The proposed solution follows the principle of model- based recognition using geometric 3D models and geometric matching. It is an alternative to the classical segmentation and primitive extraction approach and provides a perspective to escape the difficulties found with it when dealing with free- form shapes. The heart of this new approach is geometric registration which is performed by a closest point matching algorithm. Reported investigations examine the practical effectiveness of this approach for views obtained from range imaging and address relevant aspects of associated computational costs. The paper proposes solutions allowing to keep track with these costs and presents results assessing the practical feasibility of this approach. Introduction Traditional approaches to 3D vision proceed according to the signal to symbol paradigm. A basic assumption behind it is the existence of significant tokens that can be extracted from the signal and which intrinsically characterize the objects. Unfortunately, the true existence of significant and universal tokens is still an open question and after years of investigations and partial successes with tokens like planar or curved algebraic patches, it appears that their generalization for complex shapes is difficult and that with it, it is hard to progress towards the recognition of objects of arbitrary shapes. To further investigate model-based 3D vision for arbitrary-shaped objects, we opted for a recognition principle that proceeds by geometric registration of 3D shapes and works directly on the 3D coordinates of the object surface as measured by a range finder. An important component of this approach lies in the fact that the method is independent from object geometry assumptions: the representation of objects by sets of 3D points confers the method high shape modeling versatility, a property that permits to describe arbitrary shapes [3]. Geometric registration The needed geometric registration implies to find a best fit between a reference and test set of 3D data. Recently, an iterative closest point algorithm (ICP) [1] [2] was proposed to solve this problem. The algorithm proceeds iteratively by changing the objects relative poses (position and orientation) until convergence towards a best fit is obtained. Because the full search for optimal registration is computationally costly, we examine in this paper theoretical and practical possibilities to lowering it by use of adequate and fast search methods. In a previous paper [5], we considered the case of free-form 2D shapes and a simple 3D object. These experiments have since then been extended [6] and the present paper presents results for complex 3D objects. Recognition configuration Our investigations refer to a recognition configuration used for classification and pose estimation of 3D industrial objects in automatic Published in Proceeding ACCV'95 (Asian Conference on Computer Vision) 1995 which should be used for any reference to this work 1