The next generation of service robots capable of such sophisticated services as errands, table/room settings, etc. should have in itself a database of a large number of object instances human use daily. However, it may be impractical, if not impossible, that human should construct such a database for robots and update it each time new objects are introduced. To reduce the level of human involvement to a minimum, we propose a method for a robot to self-model the objects that are referred to or pointed out by human. The approach starts with the generic description of object categories assumed available a-priori. The generic description represents a category of objects as a geometric as well as functional integration of parts the 3D shape of which can be depicted as generalized cones or cylinders representing geometrical primitives or Geons. Given the 3D point clouds from the scene, 3D edge detection segments out the target object referred by human by removing out the background and/or neighboring objects. At the same time, it decomposes the target object into its parts for modeling. Here, we show that a rotational symmetric part or object, a special case of generalized cylinder, can be modeled precisely with only partial 3D point cloud data available. The proposed approach based on the generic description of object categories allows the self-modeling to infer a modeling procedure, estimate a full model from partial data, and inherit functional descriptions associated with parts. Experimental results demonstrate the effectiveness of proposed approach.
展开▼
