Environment perception is a basic problem in the design of autonomous mobile cognitive systems, i.e., of a mobile robot. A crucial part of the perception is to learn, detect, localize and recognize objects, which has to be done with limited resources. The performance of such a robot highly depends on the accuracy and reliability of its percepts and on the computational effort of the involved interpretation process. Precise localization of objects is the all-dominant step in any navigation or manipulation task.
This paper proposes a new method for the learning, fast detection and localization of instances of 3D object classes. The approach uses 3D laser range and reflectance data acquired by an autonomous mobile robot to perceive the 3D objects. The 3D range and reflectance data are transformed into images by off-screen rendering. Based on the ideas of Viola and Jones , we built a cascade of classifiers, i.e., a linear decision tree. The classifiers are composed of classification and regression trees (CARTs) and model the objects with their view dependencies. Each CART makes its decisions based on feature classifiers and learned return values. The features are edge, line, center surround, or rotated features. Lienhart et. al and Viola and Jones have implemented a method for computing effectivly these features using an intermediate representation, namely, integral image [12,25]. For learning object classes, a boosting technique, particularly, Ada Boost, is used . After detection, the object is localized using a matching technique. Hereby the pose is determined with six degrees of freedom, i.e., with respect to the , , and positions and the roll, yaw and pitch angles. Finally the quality of the object localization is evaluated by fast subsampling of the scanned 3D data. The resulting approach for object detection is reliable and real-time capable and combines recent results in computer vision with the emerging technology of 3D laser scanners. Fig. gives an overview of the implemented system.