Driving in unknown areas: From UAV images to map for autonomous vehicles

Along with the rapid development of autonomous vehicles and driving assistance systems, suitable maps have been intensively studied in recent years. Besides the improvement of conventional maps, new types such as high-density (HD) maps, have been introduced to provide comprehensive detailed information particularly for autonomous driving. In the areas which are not or not well covered by these maps, however, the autonomous vehicles are basically on their own. I.e., besides GNSS signals, they have to rely solely on the onboard sensors with local measurements. In this paper, we present an alternative pipeline for map generation from UAV imagery. A map particularly suitable for autonomous driving is derived from the 3D scene reconstructed from high-resolution images. In addition to basic geometry and semantic features, the map contains abstract vertical landmarks for fast and accurate positioning and path planning, elevation as well as slope information for driving, and trafficability information for different types of vehicles in off-road areas. The potential of the proposed approach is demonstrated based on the results of an experiment.