Page 270 - Introduction to Autonomous Mobile Robots

255
Mobile Robot Localization
features and avoids a great deal of irrelevant detail. When the robot arrives at a topological
node that could be the same as a previously visited and mapped node (e.g., similar distin-
guishing features), then the robot postulates that it has indeed returned to the same node.
To check this hypothesis, the robot explicitly plans and moves to adjacent nodes to see if
its perceptual readings are consistent with the cycle hypothesis.
With the recent popularity of metric maps, such as fixed decomposition grid represen-
tations, the cycle detection strategy is not as straightforward. Two important features are
found in most autonomous mapping systems that claim to solve the cycle detection prob-
lem. First, as with many recent systems, these mobile robots tend to accumulate recent per-
ceptual history to create small-scale local submaps [51, 74, 157]. Each submap is treated as
a single sensor during the robot’s position update. The advantage of this approach is two-
fold. Because odometry is relatively accurate over small distances, the relative registration
of features and raw sensor strikes in a local submap will be quite accurate. In addition to
this, the robot will have created a virtual sensor system with a significantly larger horizon
than its actual sensor system’s range. In a sense, this strategy at the very least defers the
problem of very large cyclic environments by increasing the map scale that can be handled
well by the robot.
The second recent technique for dealing with cycle environments is in fact a return to
the topological representation. Some recent automatic mapping systems will attempt to
identify cycles by associating a topology with the set of metric submaps, explicitly identi-
fying the loops first at the topological level. In the case of [51], for example, the topological
level loop is identified by a human who pushes a button at a known landmark position. In
the case of [74], the topological level loop is determined by performing correspondence
tests between submaps, postulating that two submaps represent the same place in the envi-
ronment when the correspondence is good.
One could certainly imagine other augmentations based on known topological methods.
For example, the globally unique localization methods described in section 5.7 could be
used to identify topological correctness. It is notable that the automatic mapping research
of the present has, in many ways, returned to the basic topological correctness question that
was at the heart of some of the earliest automatic mapping research in mobile robotics more
than a decade ago. Of course, unlike that early work, today’s automatic mapping results
boast correct cycle detection combined with high-fidelity geometric maps of the environ-
ment.

5.8.2.2 Dynamic environments
A second challenge extends not just to existing autonomous mapping solutions but to the
basic formulation of the stochastic map approach. All of these strategies tend to assume that
the environment is either unchanging or changes in ways that are virtually insignificant.
Such assumptions are certainly valid with respect to some environments, such as, for exam-
ple, the computer science department of a university at 3 AM. However, in a great many