RIA
0 detect
This ap-
ct roads
for road
tion. By
ch other
"nificant
; a lot of
insolved
ring the
criteria
focuses
satellite
solution
s able to
e. In the
ngths of
er et al.,
for road
4). E.g.,
magery:
straints.
- second
aluation
traction,
describe
ition are
ales and
1etwork,
nted. In
Albert Baumgartner
contrast to the image level, the representation at the geometry and material level is independent from sensor characteristics.
Depending on the scale, i.e., the image resolution, roads are either modeled as flat homogeneous regions, or as lines. The
solid dark lines establish the connections between the concepts at the different levels.
The use of different scale is motivated by the fact, that different characteristics of roads can be best detected at different
scales. At fine scale, i.e., in high resolution images, a better geometric accuracy can be achieved since the road sides can
be detected very precisely. Substructures on the road, e.g., markings, can give additional hints for road extraction. At
coarse scale, i.e., at resolutions where roads are only a few pixels wide, the network characteristics of roads are more
clearly visible, and small objects like single vehicles or trees do not influence the extraction as heavily as they do in fine
scale.
—— > Specialization relation
— — Part-of relation
= Concrete relation
----» General relation of objects
Junction | p------- connects... »-| Road-link
Small
Scale
Large
Scale
Road-segment
Pavement Real World
Solid Dashed
marking marking
er as iere Pon A iier eene tentent enne nnns
: | Mostly compact Long colored Short colored Elongated, flat Geometry
concrete or line line concrete or and Material
: | asphalt region asphalt area
wel nsdn bres fr —————————.
Bright Mostly straight | : |Mostly compact Long bright Short bright Elongated
blob bright line | bright region line line bright area Image
Figure 1: Road model
>
€ - à =
* Dow
&
E & : * p =
3 + x €
x ,
* € " "
. à Ed
> E
; #
No emi pe
Eb po. -
P E x >
i ; a
Se X i mU UNS X
X sc, # E -
Ss * wu
i2 x =; * T * :
e VES us BC i
7 3 i
Figure 2: (a) Image (b) Segmentation of open_rural context
In addition to the road model in Fig. 1, which comprises knowledge about geometric, radiometric, and topological prop-
erties of roads, our model contains also relations between roads and other objects, e.g., buildings, trees, and vehicles.
This type of knowledge is modeled by the context of the roads. The context model is split into /ocal and global context.
Whereas. the local context describes relations between individual objects of different types, the global context segments
the image into three regions, in which the appearance of roads in imagery is completely different: urban, forest, and
open rural context regions. A texture based segmentation of the open_rural context is shown in Fig. 2. The segmentation
of the different context regions provides a priori information about the typical problems which might occur during the
road extraction. Therefore, we can use the global context to guide the extraction and start at places where the extraction
is supposed to be easy and reliable.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 59