A novel approach to object recognition and localization in automation and handling engineering

Abstract

The industry is in need of reliable, computer aided object recognition and localization systems in automation and handling engineering. One possible application is bin picking, i.e. the task of grasping work pieces out of a storage container with a robot. Therefore, the parts do not have to be ordered or semi-ordered but can be totally unordered. 2D image processing techniques often can not perform such sophisticated tasks since the gray scale or color information provided is just not enough. An alternative is the examination of the other dimension. In this paper we discuss a novel approach to a 3D object recognizer that localizes objects by looking at the primitive features within the objects. The basic idea of the system is that the geometric primitives usually carry enough information to make possible proper object recognition and localization. The algorithms use 3D best-fitting combined with clever 2.5D preprocessing. The feasibility of the approach is demonstrated and tested by means of a prototypical bin picking system. The time taken to recognize and localize an object is < 0.5 sec., and the accuracy of the result is in the order of magnitude of the measurements inaccuracy, < 0.5 mm.