specific submodels which are adapted to contextual environment 
(open area, forest, suburb; or more specific: crossing in suburb), 
sensor, and scale (resolution). The submodels emphasize certain 
characteristics of the objects and therefore they can be regarded 
as specialized models. 
In the proposed road extraction scheme only resolution depen- 
dent submodels are employed. The development and integration 
of submodels for the contextual environment and with lower pri- 
ority for other sensors is considered an important task for future 
work. For the resolution dependent submodels there are a lot of 
partly interwoven problems: How many resolution levels are nec- 
essary for a reliable road extraction? Which resolutions provide 
additional evidence for the road recognition? Which characteris- 
tics of objects should be used at the chosen resolution? Essential 
theoretical clues to these questions can be found in the relation- 
ship between abstraction and scale-space events (Mayer, 1996). 
The general answer is that the resolutions depend on the inner and 
the outer scale of the object to be extracted. This means that the 
required resolutions can be expressed as a function of the size of 
smallest details of importance for the application and of the extent 
of the whole object. Since it is mostly impossible to see global 
characteristics of an object and every detail as well at the same 
resolution it is proposed in this paper to use more than only one 
resolution level to get a reliable road extraction. 
In the approach described below road extraction is based on 
the extraction of parallel edges which border homogeneous areas 
from an image with a ground resolution of about 25cm and on 
the extraction of lines in a version of reduced resolution of the 
original image. By fusing the results of the two resolution levels 
most of the errors in the individual results are eliminated. 
3.2 Road Detection at Low Resolution 
The notion “low resolution” cannot be fixed to a certain scale. 
In this paper “low resolution” means that roads are only a few 
pixels wide and appear as light or dark lines. Therefore, the 
resolution considered as low depends on the width of the roads 
in the imagery. If the road width varies widely in an image, e.g., 
between 4m (path) and 30 m (motorway), more than one “low 
resolution” level, e.g., one for paths and normal roads and one for 
motorways, would be needed. 
  
Figure 4: Image at low resolution 
Figure 4 shows a version of reduced, i.e., low, resolution of 
the original image. The ground resolution is 2m. Light lines 
are extracted with an approach based on differential geometrical 
properties of the image function. Points which have a vanishing 
56 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
gradient and a high curvature in the direction perpendicular to the 
line are considered as line points and linked into contours. For 
more details see (Steger, 1996). The contours are approximated 
by polygons. The extracted polygons are hypotheses for road axes 
(cf. Fig, 5). 
  
Figure 5: Hypotheses for road axes at low resolution 
3.3 Road Detection at High Resolution 
At high resolution roads are modeled as bright homogeneous areas 
bordered by parallel edges. Edges are extracted from the orig- 
inal image and approximated by polygons. These polygons are 
grouped into relations of pairs of parallels, and the area enclosed 
by the parallels is examined (Steger et al., 1995). The area to be 
investigated is indicated in Figure 4. 
3.3.1 Edge Extraction and Polygonal Approximation The 
edge extraction is performed using a modified Deriche edge op- 
erator (cf. section 2.2.1). After a thinning operation by a non- 
maximum-suppression algorithm one pixel wide edges are ob- 
tained. The computation of contours from these edges and a 
polygonal approximation is done as in section 3.2. 
3.3.2 Perceptual Grouping of Parallel Edges In the next step 
relations of parallel polygons are computed. Polygon segments 
are included in the parallel-relation if several criteria are fulfilled. 
First, the segments have to be approximately parallel. Since 
roadsides are never perfectly parallel, two roadsides are labeled 
as parallel if the angle between the line segments is below a certain 
threshold. Because longer line segments determine the direction 
more accurately the threshold becomes the smaller the longer the 
involved segments are. Second, parallel segments have to overlap. 
Third, since roads have a certain width, the distance between the 
parallels has to be smaller than a certain threshold. Results of this 
intermediate step are shown in Figure 6. 
3.3.3 Selection of Parallels Bordering Homogeneous Regions 
Up to now only geometrical properties have been employed for 
road extraction at the high resolution level. This step makes use 
of the radiometric characteristics of roads. It is assumed that 
the intensity of roads is relatively constant in the direction of the 
road, whereas it can vary considerably across the road due to 
road markings and tire tracks. To check this, the homogeneity 
of the rectangle enclosed by a pair of parallels is determined by 
examining slices which are parallel to the centerline. The slices 
are 1 pixel apart and the intensity within each slice is computed by 
bilinear interpolation. If the mean in each slice is within a certain 
   
  
range and th 
the region i 
are acceptec 
3.3.4 Ext« 
some parts ( 
geometrical 
no parallel | 
problematic 
angles, i.e., 
section 3.3. 
regions. Re 
neighboring 
homogeneit 
for roadside 
  
Figure 7: H 
which bord: 
3.4 Fusio 
Level 
Both resolu 
extraction. 
of the road: 
onthe road 
as at high r