Albert Baumgartner
3 ROAD EXTRACTION MODULES
In this section we describe the main characteristics, the strategy, the advantages, and the deficiencies of our two basic
approaches for road extraction. We will combine these two approaches and use them as complementary modules in a
new road extraction scheme. In the remainder we will refer to them as “module I” and “module II”. Both modules use
local and global information for road extraction. However, module I mainly focuses on local criteria, whereas module II
exploits especially global criteria, i.e., the network characteristics of roads.
3.1 Module I: Local Grouping
On the local level, we use lines and edges as image features to construct road segments. According to the road model,
we use apart from the original image also a version of the image with a reduced resolution. The lines extracted in the
reduced-resolution image (about 2 m) are used to select edges extracted from the original resolution that are candidates
for road sides. In order to be selected as road sides edges must fulfill several criteria: The distance between pairs of edges
must be within a certain range. The edges have to be almost parallel, i.e., there is an overlap and the direction difference
between the edges is small. The area enclosed by a pair of parallel edges should be quite homogeneous in the direction of
the road. In addition, for each pair of road side candidates a corresponding line has to exist in the reduced resolution.
From these road sides, initial hypotheses for road segments are constructed (Fig. 3). The road segments consist of quadri-
laterals which are generated from parallel road side candidates. Quadrilaterals sharing points with neighboring quadrilat-
erals are connected. The geometry of the road segments is represented by the points of their medial axes, attributed by the
road width. These road segments are the semantic objects which are used as input for the extraction of the other parts of
the road network.
(a) (b)
Figure 3: (a) Initial hypotheses for road segments (b) Detail
The fusion of lines from low resolution and edges from high resolution has proven to be very useful in order to get more
reliable results. For easy scenes these steps are often sufficient to come up with correct hypotheses for road segments
which can easily be linked into longer segments. This advantage of the combination of line and edge extraction is also
confirmed by the results of (Trinder and Wang, 1998) who use a quite similar approach to fuse low and high resolution
imagery for road extraction.
However, the limits of this initial detection of hypotheses for road segments become clear when the approach is applied
to urban or suburban areas. Inside the village the number of correct initial hypotheses for road segments decreases
tremendously, in scenes with many buildings most of the hypotheses are displaced due to shadows and occlusions, or even
completely wrong (cf. Fig. 3).
Putting the correct hypotheses together and eliminating false ones is the task to be solved during the next steps. In
module I geometric properties of neighbored road segments are used to establish hypotheses for connections between
these segments. The connection hypotheses are verified by analyzing the gap between the segments based on radiomet-
ric and geometric criteria (e.g., thresholds on mean gray value, difference in width and direction). Applying so-called
60 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B3. Amsterdam 2000.
“ribbc
nectio
neigh
steps
Simul
are re
due tc
itatior
pothe:
the ro
remo
After
ficatic
metrk
ified.
there
netwc
extrac
only |
matio
netwc
fore, i
of loc
direct
elimi
paran
this n
3.2
Modi
even
econc
conne
good
conn.
This
tracti
the p:
Base
then
erties
The €
whic]
of ve
the s
“gap
tions
linea
fuzzy
deriv
Once
“Imp
tiona
other
reser
inste
pairs