Landmark Navigation for Unmanned Ground Vehicles

We propose a novel robot navigation paradigm based on high-level visual landmarks, including structural landmarks and object-level landmarks. The system will use our novel multilayered-feature graph (MFG) to combine multiple, heterogeneous features (from points to planes to vanishing points) robustly into a single, hierarchical framework for scene structure. It will combine this with salient scene element segmentation to generate object-level landmarks with rich geometry. It will use a behavior-based control model to navigate using with respect the high-level landmarks. A predictive display will enable tele-operation with supervisory control that will allow high-level commands, such as''follow this wall'and''go around that object'.