International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part Bl. Istanbul 2004
Figure 5. The residual error vectors of 8 GCPs using a
projective transformation between the co-ordinates manually
measured in image and on a vector map, respectively.
19 Check points | RMSE| Mean | Max | Min [Extent
Easting 32 -].1 48 |-6.5 | 11.3
Northing DN 1.9 63 1-25] 3.5
Error Vector Length | 4.5 (unit: m; 1m-1.4 pixel)
Table 2. The accuracies of the 14 check points using a
projective transformation between the co-ordinates manually
measured in image and on a vector map, respectively.
2.7588 “
125728 2700 275 2333 12724. 12136 2708, 274. 2742. D744. 2748
E (m) x 10
Figure 6. The residual error vectors of check points.
4. AUTOMATIC IMAGE-AND-MAP REGISTRATION
The automatic procedures of the image-and-map registration
start from an initial location of an arbitrarily selected reference
points in the test image and the corresponding node on the
vector map. A reference point of good approximation is not
essential for the proposed algorithm for image-and-map
registration, but it does reduce the time required to carry out
image-and-map registration. The Quickbird image header data
provide useful information of map projection (UTM) and co-
ordinates of four corners, giving relatively close approximation
with deviations of translation on the order of less than 100m.
In addition, rotational errors are not significant in this case,
since the provided information of map projection of image co-
ordinates gives enough knowledge for datum transformation
between the local datum of the vector map and that of the
rectified satellite imagery. Since the proposed model of the
line feature 1s sensitive to noise in the matching procedure,
a huge amount of candidate locations can be produced by
the image-and-map registration. Thus, the critical
procedure in the image-and-map registration is to find the
best match of the line features of both types of data using
the proposed geometric-structure-matching algorithm. The
primary results as shown in Table 3 suggests that a large
portion of the population of candidate locations can be
eliminated up to 98% without using any other criteria for
the image-and-map registration. In Table 3, the searching
range is defined as the extent or the number of pixels in
respect to the reference point.
All of the polygons are registered according to the same
geometric structure as mentioned before, however, each
polygon needs to be registered individually in the first place.
Further refining processes for the remaining candidate
locations are required in order to pick up the best
estimation of registration, which is done by an analysis of
the density numbers for the linear features. Since the linear
features always convey similar radiometric characteristics,
such as homogeneous density numbers along a specific
linear feature or boundary line, it is proposed that the
variances of density numbers of all pixels along a specific
linear feature, such as a wall feature, have a minimal
difference.
Taam Candidate locations of| Geometric-
Searching
rater each polygon structure-
(in pixel) A B C D matched
candidates
20 S02 453 509 557 36
40 1826 2031 1382 1563 178
100 14188 13913 11192 14166 1096
200 57572 57651 47637 55710 4692
Table 3. The number of candidate locations of the primary
image-and-map registration using the geometric-structure-
matching technique.
Figure 7. The result derived using the geometric-structure-
matching technique for automatic registration of a
Quickbird image and a cadastral map.
Internatio
(©Q
Figure 8.
image-anc
the bound
color line.
The prim
matching
Quickbird
The four
of the sat
with the
boundary
automatic
of the ima
The algo
validated
cadastral
parcels ar
studied c
match im
error mod
been prop
algorithm
high level
for optimi
experimer
the order :
algorithm
predicted
requireme
maps, prc
satellite ii
automatic
of the in
applicatio
ground c
updating «
The resul
project N
