Large Consistent Geometric Landmark Maps

Abstract

The autonomous operation of an intelligent service robot in practical applications requires that the robot builds up a map of the environment by itself, even for large environments like supermarkets.

This paper presents a solution to the problem of building large consistent maps consisting of geometric landmarks. This solution consists of three basic steps: - incremental extraction of geometric landmarks from range data - recognition of previously mapped parts of the environment and identification of landmarks originating from the same structure and finally - removing the inconsistencies by unifying those landmarks while retaining local relations between the other landmarks.

The recognition is based on comparing partial maps of geometric landmarks. This is done by enhancing an individual landmark with features derived from its environment. Care is taken that these features are invariant with respect to missing landmarks, rotation and translation of the map and varying landmark lengths. Based on this set of features, different landmarks originating from the same real world object can be identified.

For the purpose of correcting these inconsistencies the geometric relations between landmarks are modeled by links of variable length and variable angles between a link and the adjacent landmarks forming a flexible truss. Replacing two identified landmarks with their mean modifies length and angles of the related links, thus introducing energy into the truss. The overall energy in the truss is minimized by means of numerical optimization resulting in a consistent map.

Experience in the field with about 20 robots has shown that it is possible to build up maps of large environments robustly in real-time.