Page 259 - Innovations in Intelligent Machines

252 J. Gaspar et al.
The search for the features can be approached as a general pattern match-
ing problem using e.g. a generalised Hough transform as in [93, 26]. We
approach the problem by coordinating the two navigation modalities. To find
the features, the uncertainty of the location of the robot is controlled by using
more detailed topological maps and by increasing the searching regions of the
features otherwise bounded according to the maximum speed of the robot.
During system initialisation, the robot will normally begin at a known
docking place and the undocking visual path following task may be immedi-
ately elicited. Of course, the robot may have to start at an unknown (within
the topological map) position, i.e. a drop-in-scene case. Should this occur,
then self-localisation is performed using the topological localisation module.
The combination of omnidirectional images and the Topological and Visual
Path Following navigation strategies are illustrated by the complete experi-
ments described in this section. We believe that the complementary nature of
these approaches and the use of omnidirectional imaging geometries result in
a very powerful solution to build efficient and robust navigation systems.

Topological Localisation Results

We perform two experiments to test the three topological localisation
methods, presented above. In the first experiment we test that the images
after compression by the various methods are still sufficiently different to
yield correct localisation results, and in the second experiment we test the
robustness of the methods against illumination changes.
The experiments are based on three sequences of images: one database
sequence describing the environment and two run-time sequences acquired
along a fraction of the represented environment. One of the run time sequences
was acquired at a time of the day different to the database set, resulting
therefore in very different lighting conditions.
Experiment 1: the run time sequence, as compared to the database, is
acquired under similar illumination conditions, the length of the traversed
path is about 50% of the original and the images are acquired at a different
sampling frequency (distance between consecutive images). Figure 12 shows
that the three methods give similar localisation results, as desired. The small
differences among the methods are due to the distinct image database (appear-
ance set) construction techniques. The figure shows that in this experiment
the three methods, despite compressing information, preserve enough detail
to distinguish each image relative to all the others.
Experiment 2: Fig. 13 shows topological localisation performed by each
of the methods for two sequences taken in the same path but at different
times of the day, resulting in very different lighting conditions. As expected,
the PCA-based method, i.e. the one using brightness values directly, fails to
obtain correct locations in areas of large non-uniform illumination change (i.e.
the last part of the test). The other two methods, which are based on edges,
obtain better results.