Authors:
Frank Dittrich
;
Stephan Irgenfried
and
Heinz Woern
Affiliation:
Institute for Anthropomatics and Robotics (IAR) and Karlsruhe Institute of Technology (KIT), Germany
Keyword(s):
Virtual Environment, Scene Modeling, Real-world Modeling, Synthetic Data, Probabilistic Modeling, Probabilistic Graphical Models, Image Processing, Computer Vision.
Related
Ontology
Subjects/Areas/Topics:
Computer Vision, Visualization and Computer Graphics
;
Geometry and Modeling
;
Modeling and Algorithms
;
Modeling of Natural Scenes and Phenomena
;
Scene and Object Modeling
Abstract:
In this paper we present an approach for the automated creation of real-world scenes in an virtual environment. Here we focus on human-robot interaction and collaboration in the industrial domain, with corresponding virtual object classes and inter-class constellations. As the basis for the sample generation process, we probabilistically model essential discrete and continuous object parameters, by adapting a generic mixed joint density function to distinct scene classes, in order to capture the specific inter- and intra-class dependencies. To provide a convenient way to assert these object interactions, we use a Bayesian Network for the representation of the density function, where dependencies can directly be modeled by the network layout. For the conditioned and uncertain descriptions of object translations, we use hierarchical Gaussian Mixture Models as geometrical sampling primitives in the 3D space. In our paper, we show how the combination of a Bayesian Network with these samp
ling primitives can directly be used for the automated collision avoidance of objects, during the sampling process. For the illustration of the applicability and usefulness of our approach, we instantiate the generic and abstract concept using an example with reduced complexity.
(More)