Page 262 - Introduction to Autonomous Mobile Robots
P. 262
247
Mobile Robot Localization
in the robot’s environment may have angular histograms that are too similar to be differen-
tiated successfully.
One way of attempting to gather sufficient sonar information for global localization is
to allow the robot time to gather a large amount of sonar data into a local evidence grid (i.e.,
occupancy grid) first, then match the local evidence grid with a global metric map of the
environment. In [129] the researchers demonstrate such a system as able to localize on the
fly even as significant changes are made to the environment, degrading the fidelity of the
map. Most interesting is that the local evidence grid represents information well enough
that it can be used to correct and update the map over time, thereby leading to a localization
system that provides corrective feedback to the environmental representation directly. This
is similar in spirit to the idea of taking rejected observed features in the Kalman filter local-
ization algorithm and using them to create new features in the map.
A most promising, new method for globally unique localization is called mosaic-based
localization [83]. This fascinating approach takes advantage of an environmental feature
that is rarely used by mobile robots: fine-grained floor texture. This method succeeds pri-
marily because of the recent ubiquity of very fast processors, very fast cameras, and very
large storage media.
The robot is fitted with a high-quality high-speed CCD camera pointed toward the floor,
ideally situated between the robot’s wheels, and illuminated by a specialized light pattern
off the camera axis to enhance floor texture. The robot begins by collecting images of the
entire floor in the robot’s workspace using this camera. Of course, the memory require-
ments are significant, requiring a 10 GB drive in order to store the complete image library
of a 300 x 300 area.
Once the complete image mosaic is stored, the robot can travel any trajectory on the
floor while tracking its own position without difficulty. Localization is performed by
simply recording one image, performing action update, then performing perception update
by matching the image to the mosaic database using simple techniques based on image
database matching. The resulting performance has been impressive: such a robot has been
shown to localize repeatedly with 1 mm precision while moving at 25 km/hr.
The key advantage of globally unique localization is that, when these systems function
correctly, they greatly simplify robot navigation. The robot can move to any point and will
always be assured of localizing by collecting a sensor scan.
But the main disadvantage of globally unique localization is that it is likely that this
method will never offer a complete solution to the localization problem. There will always
be cases where local sensory information is truly ambiguous and, therefore, globally unique
localization using only current sensor information is unlikely to succeed. Humans often
have excellent local positioning systems, particularly in nonrepeating and well-known
environments such as their homes. However, there are a number of environments in which
such immediate localization is challenging even for humans: consider hedge mazes and