A Curvature based Method to Extract Natural Landmarks for Mobile Robot Navigation

Landmark extraction is an essential task for robot navigation which not only requires an effective measure, but also the characterisation of landmarks to reduce the subsequent data association ambiguity. This paper describes a new method to detect natural landmarks from the adaptively estimated curvature function associated to 2D laser scans. This set of landmarks is composed of items associated to real and virtual features of the environment (corners, center of tree-like objects, line segments and edges). A novelty of the proposed system is that, for each landmark, characterisation provides not only the parameter vector, but also complete statistical information. Experimental results show the effectiveness of this method to deal with structured environments.