3D-object recognition using model-based invariants

Scene analysis in robotics is divided into camera calibration, image scanning, determination of relevant information (a graph representation), object recognition and determination of positions for grasp planning. Often grasp planning is working well, when the conditions are optimal. But in many cases there are errors caused by bad illumination, e.g. this results in an incorrect image description. This article presents some new ideas to solve this problem. We especially concentrate on the problem of 3D-object recognition. For roughly choosing those models and pose parameters in a time-optimal way we use model-based invariants which make it possible to use a priori information. As a quality measure for the so-called "fine selection" we combine point to set distances and gradient information. We give a description of the theoretical background and we discuss problems in practical applications. We finish with experiments and applications.