3-D Vision Techniques for Autonomous Vehicles

A mobile robot needs an internal representation of its environment in order to accomplish its mission. Building such a representation involves transforming raw data from sensors into a meaningful geometric representation. In this paper, we introduce techniques for building terrain representations from range data for an outdoor mobile robot. We introduce three levels of representations that correspond to levels of planning: obstacle maps, terrain patches, and high resolution elevation maps. Since terrain representations from individual locations are not sufficient for many navigation tasks, we also introduce techniques for combining multiple maps. Combining maps may be achieved either by using features or the raw elevation data. Finally, we introduce algorithms for combining 3-D descriptions with descriptions from other sensors, such as color cameras. We examine the need for this type of sensor fusion when some semantic information has to be extracted from an observed scene and provide an example application of outdoor scene analysis. Many of the techniques presented in this paper have been tested in the field on three mobile robot systems developed at CMU.