zes,
een
up
the
s or
] in
nal
oad
ely
the
eof
eas
ing
ten
ugh
10St
] in
lual
mal
rra.
har
and
95,
ral
For
ork.
ing
sed
ion
hen
ads
Annett Faber
with super regional function which cross the regions in an arbitrary orientation can be identified. This allows to perform
a classification of the roads with respect to their role in the city, thus discriminating local and regional of global road
connections. Such a situation can be clearly seen in the city map of New York (cf. Fig. 5(a)). Our task may be divided
into two parts:
(1) modeling the generic structure of the expected road network and its appearance using the sensor model
(2) analyzing the image for establishing a description of the road structure.
Maps
Image data | | Analysis Digital map
Modelling
Figure 1: Concept of the task
The model is used for controlling the analysis, whereas the results of the analysis are used to improve the assumed models
for increasing performance in subsequent analysis steps (Fig. 1). Initialization of the model may be done by hand or using
existing maps. This is the reason why we will test our procedure on existing city maps.
3 MODELING
Let us describe the model, first in general, then for our special segmentation task. Following our previous discussion
(Fôrstner, 1993, Fôrstner, 1994), our model consists of three parts: the object model, the sensor model and the image
model. We will describe this in the next subsections.
3.1 Object model
The object model describes the object, which we intend to extract from the image data. In our case the object city would
have three elements: the road network, the morphology of the built up areas and the distribution of the vegetation. The
attributes of the elements and their relations could be topological, geometrical, physical or biological. Our special model
describes dominant orthogonal road structures. The ideal map S consists of a partitioning
n
> = US
i=1
Each area §; contains two kinds of roads r;;(¢i;):
e roads, which belong to the network, their orientation angle is
T Rs.
¢ij = ¢i + kj k; € {0,1,2,3}, di € [0, 5); j = D: (1)
Thus, to each area $; is assigned an orientation ¢; which is the smallest orientation angle for one of its roads,
belonging to the orthogonal road network. The orientation of the roads differs by the angle of 7.
e roads, which are not part of the orthogonal network. Their orientation angle ¢;; is random. We assume uniform
distribution
Gi; - U(0, 27) (2)
The regions S;,7 = 1, 2, 3, ... are assumed to have piecewise smooth boundaries, i. e. have a maximum curvature except
at corners which need to have a minimum distance. This guarantees observability of the boundary, as too high curvature
parts or too dense corners can not be recovered due to the limited density of the roads within the region. This implicitly
excludes too small regions.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000. 275