Page 271 - Innovations in Intelligent Machines

264 J. Gaspar et al.
In terms of perception, omnidirectional vision has the additional advantage of
providing output views (images) with simple geometries. Our sensors output
Panoramic and Bird’s Eye views that are images as obtained by cylindrical
retinas or pin-hole cameras imaging the ground plane. Panoramic and Bird’s
Eye views are useful for navigation, namely for servoing tasks, as they make
localisation a simple 2D rigid transformation estimation problem. Successful
completion of the door crossing experiment, for example, relied on the tracking
of features surrounding the sensor. Such experiments are not possible with
limited field of view (conventional) cameras. Even cameras equipped with pan-
and-tilt mounting would be unable to perform the many separate landmark
trackings shown in our experiments.
Designing navigation modalities for the task at hand is easier and more
effective when compared to designing a single complex navigation mode [8].
Therefore, in this work, emphasis was placed on building appropriate repre-
sentations rather than always relying upon highly accurate information about
the environment. The decision to use this representation was partly inspired
by the way in which humans and animals model spatial knowledge. Our com-
bined navigation modalities, Visual Path Following and Topological Naviga-
tion, constituted an effective approach to tasks containing both short paths
to follow with high precision and long paths to follow qualitatively.
Interactive Scene Reconstruction was shown to be an effective method
of obtaining 3D scene models, as compared to conventional reconstruction
methods. For example, the model of the corridor corner, in Sect.4, was built
from a single image. This constitutes a very difficult task for automatic recon-
struction due to the low texture. These 3D models formed the basis for the
human-robot interface. Unlike many other works, a unique feature of this rep-
resentation was that the user could specify a given destination, at a certain
orientation, simply by rotating the 3D model.
When considering the system as a whole, (i) our approach to visual percep-
tion was found to be useful and convenient because it provided world-structure
information for navigation, tailored to the task at hand, (ii) the navigation
modalities fulfilled the purpose of semi-autonomous navigation by providing
autonomy while naturally combining with the human-robot interface, (iii) the
human-robot interface provided intuitive way to set high level tasks, by com-
bining limited user input with the simple output of the sensor (images).
In the future, omnidirectional vision will certainly have many develop-
ments. Many current catadioptric setups assume a rigid mounting of the
mirror on the camera. Pan-tilt-zoom cameras have been demonstrated to
be convenient for surveillance tasks, because of providing a large (virtual)
field-of-view while having good resolution when zooming at regions of interest
[75]. Adding convex mirrors will allow enlarging the field-of-view and achiev-
ing faster pan-tilt motions, obtaining the so termed active omnidirectional
camera. Networking cameras poses new calibration challenges resulting from
the mixture of various camera types, the overlapping (or not) of the fields-
of-view, the different requirements of calibration quality (many times can be