How to provide pattern representations for learning and recognizing differently located objects from arbitrary camera positions

Abstract

Our robot vision system is able to recognize differently located objects from arbitrary positions of a hand mounted camera. The type, position and orientation of the workpieces are provided with sufficient accuracy for gripping. Unknown objects are learnt, and may be recognized as soon as a name is assigned. Shift invariance is due to correct foveation by the moving camera. Parametric mappings controlled by distance and object orientation provide normalized representations to the associative network. These representations based on a special retina are sufficiently tolerant against minor deviations in size, position, and orientation, and allow the system to extract object orientation at high accuracy.