pixel gradient
bothness of
arnborough
scene.
roject boundary
o predict edge
omponents of a
1 and the edge
mponent.
s is the second
; map boundary
| with minimum
>thod generated
ure 3, and also
produced a match point file containing co-ordinates of these
match points in the map space and its corresponding points
in the image space. This match point file is the source for
the further processing to register image with the map.
_ 3 *
£ 3 “a «
© ” 1 ^ .1 % ? me
P Af At fe i
AT k pe a] . » 2
* ^ m vd 8 = À
! x a " [7 A
35 p. vss * . A
4 Ny 2. i re » m
Tto 974 DN e
- E sh gi 7 í
m d A
Wa - y * &
2 53 Mm 1
ecc ^ > > ^^
# 2 £3
$
Figure 3. Matched points of map-image pixels.
4. REGISTRATION OF IMAGE TO MAP
The matching of polygons produced 923 match points and
these points were used to find transformation parameters for
the registration of the image to the map. An affine
transformation gave a rmse of 2.0 pixel using match points.
This is not an acceptable result. It is thought that the clutter
as shown in Figure 2(c) may be the cause of large errors.
A post region segmentation processing is performed to
remove the clutter as is shown in Figure 4.
Figure 4. Building region segmentation without clutter.
The edges of buildings are extracted from the image. Then
the matching of polygons in the map and the image are
performed and which resulted in 393 match points. An affine
transformation on 393 match points resulted in a rmse of
1.8 pixel. Figure 5 shows a graph of the magnitude of the
absolute residual vector against frequency. A graph showing
stepwise a curve raised an interesting question. Is the
difference of height of buildings creating stepwise curve? It
is clear that the perspective geometry of the images will
cause displacement of the roof line with respect to the
building line. Buildings of differenent heights will therefore
cause different errors. This appears to be the effect shown in
Figure 5 in which building heights are clustered at certain
levels causing clustering in the errors. An analysis of the
errors and known building heights showed this hypothesis
to be valid. Further more the errors are similar for each
building edge. This can also be illustrated by using the
projection of a 3D object on to a 2D plane if the position of
the sensor is known as shown in Figure 6. In this figure
BcdB' is the area which will produce large residuals and AA
is the line with very low residuals.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
Least Squares Estimation
Normalised Histogram (as estimate of PDF)
0.5 Ch mire ul cran ue Lu eur eu M eae al
5 J number of pixels = 393
8 0.443
3 t iz minimum value = 0.820
S J] mean = 1.555
£ 0.34
= = maximum value = 3.930
g E
= 024 / m . RMS- 1781
T 1! | TT sd- 0.869
E 0:1 :
o Uo
z j >
FTE P P A | bM
magnitude of residual vector (pixels)
Figure 5. Residual vector vs. frequency of 393 match points.
O Sensor
d B A'
Figure 6. Model of perspective distortion.
A selection of 64 good points on undistorted edges gave a
rmse of 0.32 pixels after an affine transformation. This
demostrates that the method can generate sufficient
planimetric points for absolute orientation.
The application of bilinear resampling using the parameters
on the image resulted in the registration of the image to the
map as shown in Figure 7.
Figure 7. Farnborough subscene registered to the map.
5. DISCUSSION AND CONCLUSIONS
The system developed in this work is semi-automatic for the
registration of the image to the map. A few components of
the system are performed manually as mentioned below:
* converting buildings to solid objects in preparing the map
for matching with the image.
907
