Topological navigation and qualitative localization for indoor environment using multi-sensory perception

Abstract

This article describes a navigation system for a mobile robot which must execute motions in a building; the robot is equipped with a belt of ultrasonic sensors and with a camera. The environment is represented by a topological model based on a Generalized Voronoi Graph (GVG) and by a set of visual landmarks. Typically, the topological graph describes the free space in which the robot must navigate; a node is associated to an intersection between corridors, or to a crossing towards another topological area (an open space: rooms, hallways, …); an edge corresponds to a corridor or to a path in an open space. Landmarks correspond to static, rectangular and planar objects (e.g. doors, windows, posters, …) located on the walls. The landmarks are only located with respect to the topological graph: some of them are associated to nodes, other to edges. The paper is focused on the preliminary exploration task, i.e. the incremental construction of the topological model. The navigation task is based on this model: the robot self-localization is only expressed with respect to the graph.

Introduction

Navigation is a critical task for a mobile robot to allow it to move and act autonomously in its environment. Because internal sensors on the robot are not accurate enough or may give false measurements, a navigation system must be based on exteroceptive sensors like cameras, sonars or laser range finders.

As opposed to the classical methods based on explicit localization of the robot with respect to the environment, other methods [6], [9] make the robot localization relative only to discriminant features learned and successively perceived by the robot or relative to an area (environment modeling in topologically independent areas: corridors, room, …) [1]; the continuity of a path is guaranteed by a graph which expresses some relationships between landmarks; for example, landmark A is connected to landmark B only if B is visible from A, or if a sensor-based motion (wall following, visual servoing, for example) can be executed to go from A to B. This kind of approach could be more generally embedded in the family of qualitative or topological navigation methods. Note that these methods alone will never be sufficient to provide a truly reliable navigation system in a general indoor environment but have to be integrated in a more general, adaptive system as the ones described in [5].

This paper proposes such a topological navigation ability. A service robot must execute motions in an office environment, so the Generalized Voronoi Graph (GVG) representation proposed by Choset and coworkers [2], [7] could be well adapted to solve the navigation problem; in such a graph, nodes are associated to transitions between areas (corridor crossings, area entrances, doors, …); an edge typically corresponds to a path in a corridor or in an open space. In a corridor, the robot motion can be controlled using sonars to maintain the robot on the GVG; the robot localization is expressed according to the GVG (the robot is on this node or is moving on this edge).

Nevertheless, a self-localization problem may occur because this kind of environment is very ambiguous using only sonars whenever human presence or topological modification (an open door) may occur. If the robot is equipped with several sensors—in our experiment, monocular vision and sonars—it can take advantage of different topological representations (visual landmarks and GVG) to validate an hypothesis about a node recognition. Vision gives stable, reliable information from a large part of the environment, which may be helpful in comparison with ultrasonic sensors.

Kortenkamp and Weymouth [4] have already presented results combining these two sensors. Predetermined forms of gateways are searched with sonar and these distinct places are associated with some simple visual landmarks. The learning phase was not done autonomously, and the processing steps of sonar data made the algorithm usable in only orthogonal corridors. Our contribution is two-fold:

• we make the robot learn autonomously the environment model, without preliminary guided route traversals and with extended environment structural configurations, although considering only corridors-based environment;

• we use a visual landmark intrinsic representation independent from the viewpoint and as stable as possible with respect to illumination, scale changes and small occlusions.

Section 2 proposes an overview on the environment representation and the navigation system. 3 Learning the topological graph, 4 Learning landmarks present our strategy to build a hybrid topological map—landmark-based and GVG; in Section 5, experimental results are commented. Finally in Section 6, discussions about this work and some future researches are considered.