Vision-based Global Localization of Large-Scale Indoor Environments with Hierarchical Map

Abstract

In this paper, we propose a method to make the mobile robots building a vision-based hierarchical map and then quickly localizing itself by this map. The top-level map is a topological map which consists of “Places”. Topological map is represented through the graph and suitable for the large-scale environments. In this map, the vertices comprise some visual information for efficiently and robustly identifying the “Places” of the robot’s environment and the edges denote the spatial rela-tion between these “Places”. The visual information con-sists of the characteristics of segments and SIFT features. The former indexes “Place” coarsely and the latter is with the capability of indentifying the position within the cur-rent Place finely. When localizing itself, we merely com-pute the similarity by voting corresponding SIFT features between current information and the vertices indexed with the similar characterization of segments. This locali-zation is efficient because of the low time complexity of characterization of segments and the few candidates of the sets of SIFT features. In the other aspect, the scale, orien-tation and illumination invariant of SIFT features make this localization robust.