Hierarchical localization using compact hybrid mapping for large-scale unstructured environments

Hierarchical localization frameworks provide efficient means for localizing a robot topologically and geometrically. The metric localization performance is enhanced since the searchable space is confined to a previously identified topological place. This adapts well to large-scale environments. In this paper, a two-level hierarchical localization using a compact hybrid map is presented. The map preserves information set at two different abstractions and resolutions, and possesses both geometric and non-geometric properties. The coarse-resolution information enables global topological localization through place matching. The higher-resolution enables local metric localization through triangulation. The presented hierarchical localization is totally independent on the robot´s motion model. For effectiveness in both mapping and localization, the map is constructed based on information-theoretic evaluation that selects only highly qualitative information. The approach is demonstrated using a vision sensor and the scale invariant feature transform.