Pose Registration Model Improvement: Crease Detection Diego Viejo Miguel Cazorla Dpto. Ciencia de la Computación Dpto. Ciencia de la Computación e Inteligencia Artificial e Inteligencia Artificial Universidad de Alicante Universidad de Alicante 03080 Alicante 03080 Alicante dviejo@dccia.ua.es miguel@dccia.ua.es Resumen Several works deal with 3D data in SLAM pro- blem. Data come from a 3D laser sweeping unit or a stereo camera, both providing a hu- ge amount of data. In this paper, we detail an efficient method to find out creases from 3D raw data. This information can be used toget- her with planar patches extracted from 3D raw data in order to build a complete 3D model of the scene. Some promising results are shown for both outdoor and indoor environments. 1. Introduction One of the central research themes in mobi- le robotics is the determination of the move- ment performed by the robot using its sensors information. This methods are called pose re- gistration and can be used for automatic map building and SLAM [14] [9] [10]. The problem can be seen as finding out the correspondences between two consecutive sensor readings and then computing the transformation, which is the robot movement, from the corresponden- ces. In real 3 dimensional scenes the problem increases because the huge amount of infor- mation [17]. Modeling methods can be used to reduce scene complexity [15] [16] [20]. Our main goal is to perform pose regis- tration in semi-structured environments, i.e., man-made indoor and outdoor environments. We use a sweeping unit with a 2D laser Sick or a Digiclops stereo camera as a main sen- sor, mounted on a mobile robot. Sweeping la- ser provides 3D data with low error and higher range compared to stereo systems [11]. Howe- ver, our main aim is to deal with outliers, i.e., environments with people or not modeled ob- jects. This task is hard to manage because classics algorithms, like ICP [18] and its va- riants [19], are very sensitive to outliers. Furt- hermore, we will not use odometry informa- tion. In [12] we propose a method for pose re- gistration using 3D range data. We obtain a planar patches model from the raw data and perform over the reduced model a modified ICP-based method to achieve registration bet- ween two consecutive poses. This approach is reasonably successful when a high scope sensor is used. Nevertheless, for scenes captured with a stereo system, an improved model is needed. In the present work we propose to improve this model representation by adding crease infor- mation to the model. Surface creases have numerous applications in geometric modeling [3][4][5], image proces- sing [6][7], other fields [8], and also for the re- solving the SLAM problem [13]. To find out creases on a surface, two actions have to be performed. First, a local estimation of the cur- vature is computed along the hole surface. Se- cond, an analysis of this curvature is carried out to find the maxima and minima points on the surface. Also, a previous smoothing pro- cess may improve results [1][2]. In this paper we propose a new method for extracting crea- ses by using the algorithm described in [11] to also label points in a 3D scene as posible crea- ses. Then a posterior region growing algorithm