Fast 3D – Localization of Flat Natural Objects from Crowded Volumes by Binocular Robot Vision

In this paper we derive a model and a method suitable for robot vision to determine the 3D–positions of single flat, not entirely convex objects from crowded scenes holding many fuzzy similar shaped objects. Preferably our approach aims to objects from natural scenes. We propose to use a fast feature based stereo algorithm to match feature positions we compute from the distance transform and the morphological skeleton of the binarized stereo input. The background motivation of our method is given by the design-task of a vision guided robot system which will have to disassemble potted holiday cactus plants into separate cactus segments. On such objects with various natural mutational changes in shape, fast 3D–localization methods designed for industrial vision will fail because of lacking a sharp constrained geometric model for object recognition and/or the high amount of occlusions occurring in natural scenes.