Adaptive localization of an autonomous mobile robot in natural environments

We focus on the perceptual and decisional functions required for the self-localization of a mobile robot in a natural environment. Two main localization approaches are considered: based on the terrain numerical representations (by elevation map or B-splines), or using explicit object or landmark representations. They are not applicable in the same situations, nor do they yield the same accuracy. In order to take into account the diversity of the possible environment configurations and situations, the scene model is built with heterogeneous representations: on the one hand, a hierarchical numerical "terrain model", which involves B-spline surfaces as nonplanar primitives, well adapted to describe the geometry of uneven terrains, and, on the other hand, an "object model", which provides a more symbolic description suitable when the scene includes objects laying on a flat ground. The decisional functions which select the better localization strategy and combine them according to the localization task context are discussed.<>