Page 53 - Autonomous Mobile Robots

36 Autonomous Mobile Robots

features used for reference pass out of the FOV: in recursive mode, there is no
guarantee at initiation that features will persist. Errors (drift) are introduced
when the reference features are changed and the consequence is that a robot
will have difficulty in returning home or knowing that it is revisiting a location.
Chiuso has a scheme to reduce this problem but drift is still inevitable. On the
other hand, SLAM has to rely on sparse data because it needs to maintain a full
covariance matrix which will soon become computationally expensive if the
number of data points is not restricted. It can be difficult to associate outdoor
data when it is sparse.
The two techniques offer different benefits and a possible complementary
role. SLAM is able to maintain a sparse map on a large scale for navigation but
locally does not help much with terrain classification. SFM is useful for building
a dense model of the immediate surroundings, useful for obstacle avoidance,
path planning, and situation awareness. The availability of a 3D model (with
texture and color) created by SFM will be beneficial for validation of the sens-
ory data used in a SLAM framework: for example, associating an object type
with range data; providing color (hue) as an additional state; and so on.

1.5 CONCLUSION
We have presented the essentials of a practical VGS and provided details on
its sensors and capabilities such as road following, obstacle detection, and
sensor fusion. Worldwide, there have been many impressive demonstrations of
visual guidance and certain technologies are so mature that they are available
commercially.
This chapter started with a road map for UGVs and we have shown that the
research community is still struggling to achieve A-to-B mobility in tasks within
large-scale environments. This is because navigating through open terrain is a
highly complex problem with many unknowns. Information from the immediate
surroundings is required to determine traversable surfaces among the many
potential hazards. Vision has a role in the creation of terrain maps but we have
shown that practically this is still difficult due to the physical limitations of
available sensor technology. We anticipate technological advances that will
enable the acquisition of high-resolution 3D data at fast frame rates.
Acquiring large amounts of data is not a complete solution. We argue that
we do not make proper use of the information already available in 2D images,
and that there is potential for exploiting algorithms such as SFM and vision-
based SLAM. Another problem is finding alternative ways of representing the
environment that are more natural for navigation; or how to extract knowledge
from images and use this (state) information within algorithms.
We have made efforts to highlight problems and limitations. The task is
complex and practical understanding is essential. The only way to make real

© 2006 by Taylor & Francis Group, LLC

FRANKL: “dk6033_c001” — 2006/3/31 — 16:42 — page 36 — #36