Page 432 - Introduction to AI Robotics

415
11.7 Localization
often be a disadvantage in the field as tweaking more than one can have
conflicting side effects. The basic increments, I + and I are often changed.
Less frequently, the size of the mask W and the individual weights W p;q are
changed.

11.7 Localization

Fig. 11.1 shows a metric map built up from sensor data using shaft encoders
to localize. As can be seen, the shaft encoders are so inaccurate that the hall-
ways never connect.
ICONIC Localization can either use raw sensor data directly (iconic)oruse fea-
FEATURE-BASED tures extracted from the sensor data (feature-based). For example, iconic lo-
LOCALIZATION calization would match current sensor readings to data fused with the previ-
ous sensor measurements in the occupancy grid. Feature-based localization
might first extract a corner from the sonar data or occupancy grid, then on
the next data acquisition, the robot would extract the corner and compute the
true change in position. Feature-based localization is conceptually similar to
the idea of distinctive places in topological navigation, in the sense that there
are features in the environment that can be seen from several viewpoints.
Current metric map-making methods rely heavily on iconic localization,
and many methods use some form of continuous localization and mapping.
Essentially the robot moves a short distance and matches what it sees to what
it has built up in its map. Map matching is made more complex by the uncer-
tainties in the occupancy grid itself: what the robot thought it was seeing at
1 may have been wrong and the observations at t n are better. These
time t n
methods can be extremely accurate, though are often computationally ex-
pensive.
There is rising interest in feature-based methods for topological map-mak-
ing because gateways are of interest for maps and can be readily perceived.
The primary issue in topological map-making is the possibility that the ro-
bot mistakes one gateway for another, for example, interprets an intersection
with a hallway as a door.
Shaffer et al. compared iconic and feature-based methods. 127 They con-
cluded that iconic methods were more accurate for localization than feature-
based methods with fewer data points. Also, they noted that iconic methods
impose fewer restrictions on the environment (such as having to know the
types of features that will be available). However, feature-based algorithms
were often faster because there was less data to match during the localization