Annett Faber
Determination of an estimated dominant orienta-
tion The middle of the smallest interval which con-
tains p % of all of the grey values now describes the
estimated dominant orientation (Meer et al., 1990). It
: ; : : H(a)
is a robust estimate allowing at least 50 % outliers,
when assuming the other data to be uniformly dis- a,
tributed. For p = 50% we will get the least median |
square estimation. We actually use the p = 25%, al-
lowing 25% outliers. The robust estimation is done :
for all positions r,c of the window and so we ob- I
tain for all these positions a vector representation ^
Oo
p%
f = (coso(r,c), sino(r, c))!. We represent the i
fourfold angle « in order to cope with discontinuities
at 27.
This way we get an image which contains nearly Figure 4: Histogram of the orientation angles a. The
homogeneous regions describing the estimation of robust estimation will be obtained as the middle of the
orientation into the regions of interest. Within the shortest interval containing p% of the values. For p — 0.5
searched regions the estimated orientation will be it leads to the least-median-square-estimation.
quite good. At the borders there might be a bias if
the density of the roads in the neighboring regions is
different.
Partitioning To find homogeneous regions corresponding to the dominant orientation of the area in the road network we
use our feature extraction program FEX(Fuchs, 1998). The homogeneity measure, essentially the gradient magnitude of
the planar orientation, is used to decide whether a pixel will be interpreted as belonging to the region or not. The value in
the homogeneous regions represents the mean direction in this regions. We extract the homogeneous regions at this two
channel image only. The exoskelett we use as region boundaries.
5 RESULTS
Now we will describe results of the presented segmentation algorithm for two examples. The first example uses a piece
of a printed map of New York (Fig. 5(a)), the second a piece of an original MOMS-02 (Fig. 5(a)) image of Ballarat
(Australia), taken during the D2-mission.
e City map of New York: Fig. 6(a) shows the result of line extraction, Fig. 6(c) the homogeneous regions as result of
segmentation and Fig. 6(c) the original image overlapped with the border lines of segments.
We can observe, that the main segments are clearly visible and correctly found by the algorithms. The boundaries are
somewhat irregular. The boundaries of the small region above the center of the map is found quite well. The other
regions are contaminated by border effects of the segmentation procedures due to the large window size of appr. 15
% of the complete map.
e MOMS-02 image of Ballarat (Australia): Fig. 7(a) shows the result of line extraction, Fig. 7(c) the homogeneous
regions as result of segmentation and Fig. 7(c) the original image overlapped with the border lines of segments.
In this image exist only two larger regions with different orientations: the main region, extending over the whole
image nearly, and a smaller one at the left image margin. These two regions were clearly segmented. Additionally
some small regions inside the main region were missed segmented, because there are any errors in the detected lines.
The result for the MOMS-02 image is not so clear as the result of city map of New York. The main reason is the
irregularity of the density of the road network.
278 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.