Page 261 - Introduction to Autonomous Mobile Robots

Chapter 5
246
that the robot will always be able to localize successfully. This work also led to a real-world
demonstration of landmark-based localization. Standard sheets of paper were placed on the
ceiling of the Robotics Laboratory at Stanford University, each with a unique checkerboard
pattern. A Nomadics 200 mobile robot was fitted with a monochrome CCD camera aimed
vertically up at the ceiling. By recognizing the paper landmarks, which were placed approx-
imately 2 m apart, the robot was able to localize to within several centimeters, then move,
using dead reckoning, to another landmark zone.
The primary disadvantage of landmark-based navigation is that in general it requires sig-
nificant environmental modification. Landmarks are local, and therefore a large number are
usually required to cover a large factory area or research laboratory. For example, the
Robotics Laboratory at Stanford made use of approximately thirty discrete landmarks, all
affixed individually to the ceiling.

5.7.2 Globally unique localization
The landmark-based navigation approach makes a strong general assumption: when the
landmark is in the robot’s field of view, localization is essentially perfect. One way to reach
the Holy Grail of mobile robotic localization is to effectively enable such an assumption to
be valid no matter where the robot is located. It would be revolutionary if a look at the
robot’s sensors immediately identified its particular location, uniquely and repeatedly.
Such a strategy for localization is surely aggressive, but the question of whether it can
be done is primarily a question of sensor technology and sensing software. Clearly, such a
localization system would need to use a sensor that collects a very large amount of infor-
mation. Since vision does indeed collect far more information than previous sensors, it has
been used as the sensor of choice in research toward globally unique localization.
Figure 4.49 depicts the image taken by a catadioptric camera system. If humans were
able to look at an individual such picture and identify the robot’s location in a well-known
environment, then one could argue that the information for globally unique localization
does exist within the picture; it must simply be teased out.
One such approach has been attempted by several researchers and involves constructing
one or more image histograms to represent the information content of an image stably (see
e.g., figure 4.50 and section 4.3.2.2). A robot using such an image-histogramming system
has been shown to uniquely identify individual rooms in an office building as well as indi-
vidual sidewalks in an outdoor environment. However, such a system is highly sensitive to
external illumination and provides only a level of localization resolution equal to the visual
footprint of the camera optics.
The angular histogram depicted in figure 4.39 of the previous chapter is another example
in which the robot’s sensor values are transformed into an identifier of location. However,
due to the limited information content of sonar ranging strikes, it is likely that two places