Learning Operators for View Independent Object Recognition
Josef Pauli · 1996
In the context of vision based robotics our work focusses on the recognition of target objects for object grasping. The objects are arbitrarily shaped, and the viewing position and orientation of the camera is arbitrary as well. Due to various imponderables it is hard to geometrical model all relevant 3D object shapes and all effects of perspective projection. Therefore 3D model based approaches for object recognition are unfavorable in our robot application. Rather we use typical 2D appearance patterns of the target object which will be learned in a training phase. To acquire training patterns the turning angle of the object and the focal length of the camera lens must be changed systematically. A multidimensional Gaussian is defined for each typical pattern and used as basis function (GBF) for computing similarities to certain image patches. By appropriate linear combination of the GBFs we get a smart operator for recognizing the target object (regardless of object turn or distance)....