Page 371 - Autonomous Mobile Robots

Map Building and SLAM Algorithms 361

9.4.3 Matching and Fusion after Map Joining
The map resulting from map joining may contain features that, coming from
different local maps, correspond to the same environment feature. To eliminate
such duplications and obtain a more precise map we need a data association
algorithm to determine correspondences, and a feature fusion mechanism to
update the global map. For determining correspondences we use the JCBB
algorithm described in Section 9.3.2.
Feature fusion is performed by a modified version of the EKF update
equations, which consider a nonlinear measurement equation:

z ij = h ij (x) = 0 (9.25)

with null noise covariance matrix, which constraints the relative location
between the duplicates F i and F j of an environment feature. Once the matching
constraints have been applied, the corresponding matching features become
fully correlated, with the same estimation and covariance. Thus, one of them
can be eliminated.
The whole process of local map joining, matching, and fusion can be seen
in the example of Figure 9.9.

9.4.4 Closing a Large Loop
To compare map joining with full EKF–SLAM we have performed a map
building experiment, using a robotized wheelchair equipped with a SICK laser
scanner. The vehicle was hand-driven along a loop of about 250 m in a pop-
ulated indoor/outdoor environment in the Ada Byron building of our campus.
The laser scans were segmented to obtain lines using the RANSAC technique.
The global map obtained using the classical EKF–SLAM algorithm is shown
in Figure 9.10a. At this point, the vehicle was very close to the initial starting
position, closing the loop. The figure shows that the vehicle estimated location
has some 10 m error and the corresponding 95% uncertainty ellipses are ridicu-
louslysmall, givinganinconsistentestimation. Duetothesesmallellipsoids, the
JCBB data association algorithm was unable to properly detect the loop closure.
This corroborates the results obtained with simulations in Reference 9: in large
environments the map obtained by EKF–SLAM quickly becomes inconsistent,
due to linearization errors.
The same dataset was processed to obtain independent local maps at fixed
intervals of about 10 m. The local maps were joined and fused obtaining the
global map shown in Figure 9.10b. In this case the loop was correctly detec-
ted by JCBB and the map obtained seems to be consistent. Furthermore, the
computational time was about 50 times shorter that the standard EKF approach.

© 2006 by Taylor & Francis Group, LLC

FRANKL: “dk6033_c009” — 2006/3/31 — 16:43 — page 361 — #31