Page - 89 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

model and scene point clouds. In general, keypoints are defined by detecting characteristic surface points. A simple and efficient alternative is to sample keypoints uniformly from the surface. The local geometry of each keypoint is described by the Signature of Histograms of Orientation (SHOT) descriptor [17], which delivers favorable results for the evaluated dataset. PCL provides a variety of different descriptor implementations. A comprehensive comparison can be found in [5]. Correspon- dences are generated by matching scene descriptors against a database of offline computed model descriptors. The next step clusters geometrically consistent correspondences into groups. Starting from a seed correspondence ci={pmi ,psi} (pmi andpsi denote corresponding key points of model and scene), geometrical consistency follows from the following relation |||pmi −pmj ||2−||psi−psj||2|<ε (1) where ε defines a distance threshold between the keypoints. A minimum of three correspondences is required to estimate a 6DOF pose. The absolute orientation step eliminates correspondences that are not consistent with a unique 6DOF pose. The utilized recognition pipeline provides an optional iterativeclosestpoint (ICP) refinementstep,whichcanbeappliedon the recognizedhypotheses. The numberof ICPiterationshasbeenset toa lowvalue. Runningmore than5ICPiterationsonthegiven datasetdoesnot result in significant recognition improvements. Thefinalhypothesisverificationstep determinesasetofnon-conflictingmodelhypothesis thatareinaccordancewiththescenepointcloud. Hypothesis that result from unexpected objects within the scene have to withstand the following quality measurement. An acceptance function evaluates the number of supported model points that are close to scene points, as well as the number of unsupported model points (visible model points thathavenocounterpart in thescene). Adetaileddescriptionof thehypothesisverificationalgorithm that hasbeenutilized in this paper is given in [12]. Figure2: Recognitionpipelineused in thispaper. 3.2. Model-FreePointCloudSegmentation Segmentation results from summarizing interesting and distinguishable image properties. Higher- level visual tasks like object recognition and pose estimation can benefit from such condensed image representations. A method that segments the signal of a RGBD-sensor, without explicit object model informationhasbeenpresentedin[1]. Homogeneousregions(segments)aregeneratedbyusingcolor information. In addition, the method exploits depth information in order to support the segmentation and tracking process. Figure 3 shows two example scenes that have been segmented by this method. The segmentation result depends on several factors like scene density, degree of occlusion, object geometry, light conditions, etc. Figure3: Point cloudsegmentation generatedbyacolor-based model-freemethod. 89