004
(x
() —
the
luce
rm.
the
e.
dn
that
ints
ga
ion,
n at
ling
ans
ctly
ing
be
n to
ace.
VENIS
Is to
and
an
im.
1ave
ce 7
ond,
the
nber
vhen
t the
International Archives of the Photogrammetry, Remote Sensin g and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
value at the found global maximum in Hough Domain became
smaller than the given threshold 7.
In the example shown in Figure 5, the process was terminated
after four iteration steps, which means that the building is
described by four lines (Figure 5, last row).
I N
MN \
^ \
) X
| \ >
vi frac” N
N
f N
L \
) N
N — \
s \
X
vri 5
e
A
— et
-— s Ti
e
d E d
C E
] oT
J
^ \
Lf
| x SE
x X
A P T *
x 3 N x]
\ À \ ^
mr. TT
|
Figure 5. Iterations of Hough Transform (left column:
vectorized building; centre column: Hough Domain: right
column: lines of back-transformed maxima)
Summing it up, the user has to tell the system either how many
edges have to be found for each building (e.g. if four is chosen,
the buildings will have a quadrangle form), or the minimum
number of points lying on an edge. In the latter case the
building will be represented in more detail (Figure 6), but one
has to keep in mind that the additional detail might come from
vectorization errors. Furthermore, the user also has to tell the
program how big the tolerance value 7 should be. This value
depends on the quality of the vector model and hence the
quality and resolution of the image.
The intersection of the obtained lines, as mentioned before, can
lead to wrong points that do not lie on the building boundary.
To avoid this problem the program searches only for
intersections that lie close to the vector data.
- Each back-transformed edge is marked with a unique
attribute (Figure 6c).
— All lines (and intersections) are dilated by the
tolerance value z (Figure 6d).
— Starting at a point on the vectorized building (Figure
6b) and moving along the edges in counter-clockwise
direction each pixel position is checked for the
attribute of the closest dilated line.
— As soon as we have a change in line-attributes we
search for the closest intersection of back-transformed
edges and store the position as corner point (Figure
6e)!
— The procedure stops as soon as the start-position is
reached.
Each extracted building will be assigned one height value
(horizontal roof). So, all corner points of a building will have
the same elevation, which is taken from the DSM. The building
is completed by assuming vertical walls that are intersected
with the DTM.
Figure 6. Finding the building corners (a: Hough Domain with
6 determined maxima; b: vector representation; c: back-
transformed lines; d: dilated back-transformed lines with vector
data superimposed; e: final building represented by 6 points)
5. CHANGE DETECTION AND UPDATING
The old situation is compared to the newly calculated geometric
building properties and if a certain threshold in the difference is
exceeded the ‘old’ database is updated.
The users may select one of the two proposed types of change
detection dependent on their requirements:
- The comparison of two vector DCMs, namely the old
one and the newly generated one. This updating
procedure is capable of finding new buildings that do
not exist in the old data set.
— The second approach is just to check the old vector
DCM for changes. This process is much faster since
the whole step of building recognition can be
neglected, but it is not in position to detect new
buildings.
When checking for changes one always has to bear in mind
how good the quality of the extracted buildings is. This quality
factor is highly correlated to the image resolution and DSM
accuracy.
6. CASE STUDIES
The case studies comprise two spaceborne IKONOS stereo
images and three airborne three-line sensor data.
The IKONOS (PrecisionPlus) along-track stereo imageries were
taken over Athens (Greece) and Melbourne (Australia) with a
ground sampling distance of one metre. The data-preparation
(orientation, DSM extraction and Orthophoto production) was
done with the ERDAS Imagine software version 8.6. Only
subsets of these scenes were used for algorithm testing.
The ADS40 imagery of Nimes (France) and HRSC-AX data set
of Bern (Switzerland) were acquired from the DLR (German
Space Agency), where also the data-preparation was carried out
(see Scholten and Gwinner, 2003). The Nimes images were
captured at a flying height of 800 metres and have a ground
resolution of 8 cm, whereas the Bern data was acquired at an
1037