Griffermittlung für anthropomorphe Endeffektoren mit Hilfe von geometrischem Vorwissen

摘要

Recent technological advancements increase the commercial availability of autonomous mobile systems, which can operate in a household environment. At present, they are still lacking the ability to recognize and grasp a priori unknown objects, which restricts their field of application. The reason for this deficiency is the inability of these systems to adequately determine the state of the environment on basis of sensor data. Recent approaches for image based object recognition use a priori knowledge in order to identify objects. This knowledge is usually based on specific image features like texture, although geometric data is required to determine a suitable grasp. Even if the exact geometry of the objects is known, the available methods for determining anthropomorphic grasps require time-consuming calculations, thus restricting real time application. In order to grasp a priori unknown objects, this thesis presents a new object representation, which describes the object shape using a topology of geometrical solids. Contrary to the exact quantitative nature of recent feature representations, this approach provides a qualitative description of the object shape without depending on specific color data. Thus, geometrically similar objects with varying dimensions can be encoded with a single object representation. In addition, grasping strategies are linked to the solids, thus removing the necessity to calculate the end effector configuration during the grasp process. The proposed object representation does not contain any specific image features, which would allow for a direct comparison between a priori knowledge and examined object. Instead, the geometrical bodies have to be identified in the image data. Their scale and pose can vary and some bodies can be partially or totally occluded by others. The recognition of an object is based on a novel concept, which identifies the topology of the geometrical bodies using the object silhouette as image feature. Point distribution models encode all possible geometric variations for the silhouettes of a solid. During object recognition, the object silhouette is separated into contour parts, which are matched to the known point distribution models using active shape models. A recursive algorithm separates the silhouette until all contour segments were successfully matched. In order to cover all solids of the topology, the system performs this analysis on image data acquired from three different viewpoints. The collected shape information is finally used to deduce the object type, its pose and a suitable grasp strategy. The approach was evaluated using the DLR-Hand. A hardware device of this anthropomorphic end effector was not available. Therefore, the evaluation took place in a virtual environment using 3D models of the end effector and household objects. This approach simplified the problem of image segmentation. Of 17 different objects, which were lying in known position and unknown orientation on a flat surface, 16 were identified correctly. Due to delays in the control of the end effector pose, 18 out of 133 different grasps did not create a stable fixation. The results show that objects can be recognized using a qualitative description of their shape as a priori knowledge. In contrast to classical feature based approaches, this kind of representation allows the storage and usage of predefined grasp strategies without performing a time-consuming online calculation of the optimal grasp.