Page 253 - Innovations in Intelligent Machines

246 J. Gaspar et al.
The appropriate choice of the sensor and environmental representations,
taking into account the task at hand, results in an efficient methodology that
hardwires some tasks requiring precise navigation.

3.2 Topological Representations

A topological map is used to describe the robot’s global environment and
obtain its qualitative position when travelling long distances. It is represented
as a graph: nodes in the graph correspond to landmarks, i.e. distinctive places
such as corners. Links connect nodes and correspond to environmental struc-
tures that can be used to control the pose of the robot. In order to effectively
use this graph the robot must be able to travel along a corridor, recognize
the ends of a corridor, make turns, identify and count door frames. These
behaviours are implemented through an appearance based system and a visual
servoing strategy.
An appearance based system [62] is one in which a run–time image is com-
pared to a database set for matching purposes. For example, in our corridor
scene, the appearance based system provides qualitative estimates of robot
position and recognizes distinctive places such as corner or door entrances.
Therefore, the topological map is simply a collection of inter-connected
images. To go from one particular locale to another, we do not have to think
in precise metric terms. For example, to move the robot from one corner to
the opposite one we may indicate to the robot to follow one corridor until the
first corner and then to follow the next corridor until the next corner, thus
reaching the desired destination, or to complete more complex missions such
as “go to the third office on the left-hand side of the second corridor”.
To control the robot’s trajectory along a corridor, we detect the corridor
guidelines and generate adequate control signals to keep the robot on the
desired trajectory. This processing is performed on bird’s eye views of the
ground plane, computed in real-time.
When compared to geometric approaches, topological maps offer a parsi-
monious representation of the environment, are highly computationally effi-
cient [85], scale easily and can explicitly represent uncertainties in the real
world [7].

Image Eigenspaces as Topological Maps

In general, sizeable learning sets are required to map the environment and so
matching using traditional techniques, such as correlation, would incur a very
high computational cost. If one considers the images as points in space, it
follows that they shall be scattered throughout this space, only if they differ
significantly from one other. However, many real-world environments (offices,
highways etc.) exhibit homogeneity of structure, leading to a large amount of
redundant information within the image set. Consequently, the images are not