Hiroshi Masaharu
it is not necessary to consider horizontal displacement between
polygon data and height data in region segmentation because the
source data is the same. On the other hand, it is necessary to
consider the displacement when using 2-D map data because
there are positioning errors in laser scanner system. Second,
region segmentation method has no such problem of changes
caused by time difference between the polygon data and height
data.
We have tried another method to derive building boundaries
using edge detection technique (Hasegawa et al., 1999). In this
method, it is rather difficult to obtain closed polylines to
compose polygons. On the contrary, region segmentation
method gives closed polylines in very natural way.
The region segmentation method, however, has its intrinsic
problems that in some condition even large buildings cannot be
distinguished from road surface. This is related to the selection
of the threshold of region segmentation. Another problem is
that the extracted objects are not always buildings or houses but,
for example, trees are included. The separation of buildings and
trees is a problem to be solved. These problems are discussed
in the following sections.
To summarize, except for the above mentioned problems, the
region segmentation of laser scanner data is a good method to
generate 3-D scenery model in the sense that the method can
generate it through automatic processing and therefore makes the
most of the merit of prompt data acquisition of laser scanner.
4.2 Appropriate Threshold Value for Region Segmentation
We thought at first the threshold would be around 1.5 m since the
height difference between road surface and buildings seemed to
exceed this value in any case. But many buildings were not
extracted as different regions with this threshold. Figure 7
shows the extracted polygons with the threshold of 1 m and 0.5 m
overlaid on the DSM image of the laser scanner data. It can be
seen that many buildings are not extracted with the threshold of 1
m. It is seen that a large building (This is a supermarket
actually.) at the lower end of the figure is not extracted even with
the threshold of 0.5 m.
We consider the reason as follows. Individual regions derived
from region segmentation have height difference with
neighboring regions larger than the threshold at the entire
boundary points. This means that two regions are merged into
one region if there is a path where the height differences of
adjacent pixels are under the threshold and thus connecting the
regions. This is demonstrated at the above mentioned large
building at the lower end of Figure 7. There is a slope leading
to the parking on the roof of the building. This path connected
the region of the roof of the building with the ground surface.
We found from field examination that failure to extract buildings
happens where buildings have special shapes having slopes or
staircases outside such as mentioned before, or where there are
trees, such as garden trees, next to houses.
We estimate that appropriate threshold is about 0.5 m for the data
with grid interval of 0.5 m from our experience. But some
buildings are not extracted with this threshold. Reducing the
Figure 7. Extracted Polygons Overlaid on the
Laser Scanner DSM
(Threshold: Dark blue=1m, Yellow=0.5m)
Elevati on (m 5 :
61 - 62 D
me: |
: 66 a
rm
uu
E |
L|
L—1
82382888
g
0.07 0 0.07 0. 14 km
Figure 8. Ground Surface DEM by Laser Scanner
560 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
threshc
to deve
edge de
43 O
From t
Then i
neighb
as Figu
This is
all the
and are
buildin
44 DL
The re
be bad
We ha
sensor
trees a
8 FL
AS Wr
from 1
extrac
We ha
dimen
We ha
combi
to dev
horizo
To de
scann
6 C(
A me
region
using
scann
and of
ACK
We w
D mo
REF]
Ackei
Photo
Doug