Geometrical representations: special features

During the postdoctoral stay of Jose Castellanos (August 1998-January 1999), LAAS studied some problems related to special feature representation, using only data acquired by a 2D laser range finder, mounted on a mobile robot. The research concerns:

• the representation of uncertain geometrical features and the formalism used to register and fuse data acquired from successive positions,
• and the structure of the final model, taking into account geometrical ( e.g. a same wall perceived from its two sides) and topological ( e.g. uncorrelated features acquired in different rooms) constraints.

The objective was to integrate the Symmetries and Perturbations Model (SP-model) proposed by J.Tardos from the Zaragossa University, with the landmark-based modeling method proposed by M.Devy at LAAS-CNRS:

• the SP-model uniformly represents all features which could be extracted from the sensory data (points, segments, curves, surface patches, ...). A reference frame is linked to all features: the intrinsic symmetries which remains due to some degrees of freedom (for example, a translation along the straight line which supports a segment or on a surface which support a facet, a rotation around a point, ...), are taken into account by a bounding matrix which masks the unuseful parameters of the reference frame.
• the landmark-based method tries to group the features extracted from the images, so that the final model is a set of decoupled stochastic maps: a map for each landmark (in a structured environment, a specific object) and a map for each topological area (for a building, a room, a corridor, ...).

An off-line experiment has been done; the results are promising and the work is continuing at LAAS, with (1) an on line experiment concerning the learning of the 2D map of a building, (2) an extension in order to deal with the geometrical and topological constraints, and (3) a work on perception strategies using Bayesian Belief Networks, in order to make the modeling task more autonomous.

On the following figures, we show a landmark-based model built from one viewpoint, and the integration of thirty local models into a global one.

J.A.Castellanos, M.Devy and J.D.Tardos, "Towards a topological representation of indoor environments : a landmark-based approach", Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS'99), Kyongju (Korea), November 1999.