DetailseiteDetails

Bayesian Estimation of Object Poses from Multiple Viewpoints

(3D Bild-Analyse und Synthese)
Externe Arbeit
Betreuer: Anas Al-Nuaimi
Real world tabletop scenes are mostly partially known, which means that the scene is not fully unknown but also that 3D models are not available for all objects in the scene. Regarding robotic tasks such as grasping or manipulating objects on a table, at least the objects that should be grasped need to be a priori known. Other objects that may be occluded or are not in the field of view, remain typically unrecognized by autonomous system. Thus, additional actions are required, such as resolving the occlusions by multiple view points or acquiring a complete model with a certain accuracy and extending the recognition database. In robotic perception however, the recognition and localization of objects for tasks like grasping and manipulation is usually kept separated from environment exploration and modeling for path planning, self localization, and object model extraction. Though, for tackling the analysis of partially known scenes in an autonomous way, recognition and exploration parts have to cooperate as a single scene exploration system. Thereby, exploration can provide useful views from the global model for multi-view recognition, and vice versa, recognition can refine the global model with object information. Further, the detection of non-recognized clusters during recognition has to trigger autonomous object modeling and database update. This work will focus on merging of object recognition results from multiple views, relying on probabilistic methods operating in the space of the special orthogonal group SO(3, R ) in combination with the space of translations, based on [1]. The use of multiple pose estimation and object classification methods will be tested, merging their results in order to improve the overall detection rates. Additionally, the integration between view planning and pose estimation will be explored. Different strategies for sensor placement will be evaluated, and the detection of model deficiencies solved, in order to trigger re-modeling them. This will enable the building of an expandable object model database. If time permits, autonomous detection method selection and probabilistic object identity resolution will be explored, based on systems like [2,3]. [1] crl.sharkdolphin.com/common/srt_distance/index.html [2] ai.stanford.edu/~koller/papers.cgi [3] ias.cs.tum.edu/_media/spezial/bib/blodow10humanoids.pdf "
Keywords: Computer Vision. Point Cloud Processing.