OBJECTS LAYOUT GRAPH FOR 3D COMPLEX SCENES 1 A. Adán, 2 P. Merchán, 2 S. Salamanca, 1 A.Vázquez, 1 M. Adán, 3 C. Cerrada 1 Escuela Superior de Informática. UCLM. Spain. Antonio.Adan@uclm.es 2 Escuela de Ingenierías Industriales. UEX. Spain. pmerchan@unex.es 3 Escuela T.S.I. Industriales. UNED. Spain. ccerrada@ieec.uned.es ABSTRACT This paper shows how to extract information about the parts and their layout in a complex scene when a single range image is available. In the worst case, the complexity of the scene includes: no shape-restrictions, shades, occlusion, cluttering, contact, surfaces viewed in oblique angles and without textures. This work is a prerequisite before carrying out further robot interaction actions in the scene. The process is based on a novel 3D range data segmentation technique that avoids most restrictions imposed on other techniques. Making use of the 3D segmented parts, the method carries out an objects silhouette classification which allows us to perform a layout graph of the objects in the scene. A brief description of this method and experimental results are presented throughout the paper. 1. MOTIVATION Suppose that we have a single view of a complex scene and a robot has to manipulate the objects that are in it. An interaction (grasping, pushing, touching, etc) of a robot in such a scene is highly difficult unless knowledge about the parts and their layout in the scene is provided in advance. In this paper we deal with the problem of parts segmentation and their relative position in complex scenes through an occlusion study. This becomes a serious problem depending on the complexity degree of the scene. For instance, in the case of no-textured images, any processing technique on the intensity image is clearly inefficient for extracting the objects that compose the scene. That is why we have applied solutions through range image segmentation techniques instead of image processing ones. Figure 1 presents the real environment with the components that we are using in our work: the prototype of scene, the immobile range sensor and the robot. Range image segmentation strategies can be categorized in edge-based approaches and region-based approaches. Reference [1] offers evaluation and experimental comparison among them. In edge-based approaches, the points located on the edges are first identified, followed by edge linking, contours and surfaces definition. In this field, a wide variety of algorithms, where edges or contours are segmented, can be found ([2], [3], [4]). In region-based approaches a number of seed regions are first chosen. These seed regions grow by adding neighbour points based on some compatibility threshold. Some methods based on region growing are proposed in [5], [6] and [7]. Fig.1.Robot interaction model in a complex scene On the other hand, the relative location of the segment in the scene is a problem that can be solved by techniques based on silhouettes. In [8] Super et al. use a part-based shape retrieval method as a hypothesizer for the system. So, they avoid the cost of comparing every object model in the database. Serratosa et al. [9] define the function- described graphs (FDG) that are applied for 3D matching and human faces recognition. An aspect-graph approach that measures the similarity between two views by a 2D shape metric is presented in [10]. In [11] Sebastian et al. proposes a recognition framework which is based on matching shock graphs of 2D shape outlines. All works cited above involve some kind of restriction: object limited poses [10], several views of the scene [9], without occlusion [10, 11], only for simplex scenes [8,9]. Our segmentation technique is related to the region-based approaches but is different from most segmentation strategies because a distributive segmentation notion is introduced in the solution. This makes the method robust and insensitive to restrictions imposed on other Robot Range Finder System Scene: contact, shades, occlusion, cluttering 0-7803-9134-9/05/$20.00 ©2005 IEEE III-433