In this paper we present an approach to dynamic model-based feature prediction and feature planning for our laser scanner-based navigation system. The system enables autonomous precise maneuvering of vehicles relative to target objects based on 2D laser range data. An adaptive tracking algorithm uses predicted features to localize the target. Jump edges, line segments, line intersections and free space areas are considered. Feature prediction is based on an attributed polygonal 3D object model representing object surfaces, masking volumes and free space volumes. The model is intersected by a virtual scan plane to dynamically determine all detectable features for the current sensor pose. In order to further increase range, robustness and precision of navigation, feature planning is performed prior to or dynamically during the vehicle's approach towards the target. A rating function allows to compare feature sets with regard to visibility and quality. This enables active localization, i.e. controlling and optimizing the pose of an actuated sensor, which is especially useful for long range approaches and compensation of ground unevenness. The system has been evaluated and in part it has already been tested successfully with a Mercedes-Benz truck on an outdoor test yard under varying environmental conditions.
展开▼
