Page 47 - Autonomous Mobile Robots

30 Autonomous Mobile Robots

The principle behind the localized correlation sensor fusion is: instead of
directly averaging L ij and S ij to get F ij , a search is performed to find the best
match within a small neighborhood. The averaging of the center pixel at a
matched point produces the final fusion map.
In case an obstacle map only takes three values: obstacle, traversable, and
undefined; the approach above can be simplified as


L
 ij S ij = undefined



S ij L ij = undefined


F ij = 1 L ij = 1, C so > T 1 , D < T 2 (1.11)


1 S ij = 1, C lo > T 1 , D < T 2




0 otherwise
where T 1 and T 2 are preset thresholds that depend on the size of the search
window. In our experiments a window of size 5 × 5 pixels was found to
work well. The choice of size is a compromise between noise problems with
small windows and excessive boundary points with large windows. C so and
C lo are obstacle pixel counts within the comparison window w c , for L ij and S ij ,
respectively, D is the minimum distance between L ij and S ij in :
 
w c /2 w c /2

D = min  |S i+m+p,j+n+q − L i+p,j+q |  (m, n) ∈ (1.12)
p=−w c /2 q=−w c /2
w c /2 w c /2

C so = S i+m+p,j+n+q (1.13)
p=−w c /2 q=−w c /2
The advantage of implementing correlation-based fusion method is two-
fold: it reduces false alarm rates and compensates for the inaccuracy from
laser and stereo calibration/synchronization. The experimental results of using
above mentioned approach for laser and stereo obstacle map fusion are shown
in Figure 1.9.

The geometry of 2D range and image data fusion. Integration of sensory
data offers much more than a projection onto an occupancy grid. There exist
multiple view constraints between image and range data analogous to those
between multiple images. These constraints help to verify and disambiguate
data from either source, so it is useful to examine the coordinate transformations
and the physical parameters that define them. This will also provide a robust
framework for selecting what data should be fused and in which geometric
representation.

© 2006 by Taylor & Francis Group, LLC

FRANKL: “dk6033_c001” — 2006/3/31 — 16:42 — page 30 — #30