be carried out 
^ the chosen 
re marked in 
00 250 
nd Steger 
e in feature 
ormance. To 
ion with the 
ble, the limit 
the output of 
operator was 
ion on to any 
| parameters. 
perators with 
grey value 
ighter line. A 
sponses even 
ient contrast. 
0 250 
vith varying 
to prevailing 
e operator's 
0..3 and. 5. 
range. With 
performance 
to noise iS 
nd Deriche 
g noise. 
operators is 
h noise. The 
er the noise 
the trustable 
smaller the 
Istanbul 2004 
    
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
  
Roadway 
Cont. Stripe 
gv1 
exti 
p1 
| NC 
part-of T o. part-of 
left boundary part-of part-of central 
[ Readway |  |EdgeLine| _ [Lane Marking _ | Double | 
Margin Left | left | = left Z [Whiteline 
| Cont. Stripe | — Cont.Line « . Periodic Line. « Cont. Stripe] 
| O-> t Oo > o > re 
gv3 = gv2 = gv2 = gv2 
ext3 S ext2 2 ext2 e 2*ext2 | 
| 
p1 pj. i] pd pte! 
| | | 
i 
left-of [4*d1-d3] | | 
| 
| 
| 
| 
| 
No = T 
right-of [d1] 
Legend: 
| Object Part Name. 
| Object type 
{ Gray Value gv 
| Extent ext 
| Periodicity p 
Es | spatial relation | | 
[Distance d] | 
= optional —. : part-of 
part-of part-of | part-of | right boundary 
(Lane Marking, iRoad Work| ^ [Edgeline| | | Roadway 
. Righ | | Marking | @ | Right | § Margin Right 
| Periodic Line | ContLine | — | ContlLine + | Cont Stripe 
{ d ro | o 
gv2 | gv4 | = | gv2 = gv3 
ext2 | ext4 | 9j ext? 2 ext3 
| p1 LÍ 
Pl At dal 
|; right-of [d1] 
Feature Extraction 
Figure 6. Concept Net for Dual Carriageway at Largest Scale, Generated for Images with Ground Pixel Sizes of 3.3 - 7 cm/pel 
5. EXAMPLE FOR SCALE ADAPTATION 
For an exemplarily chosen application of a dual carriageway we 
created a semantic net following the developed constraints as 
described in section 3. Fig.7 shows our test image, a cut-out 
from an aerial image with a ground pixel size of 3.3 cm. The 
goal of this section is to show that the proposed kind of 
semantic net is suitable to follow the scale space events in 
digital images, and is therefore suitable to be used in an 
automatic approach. 
In the concept net, as presented in Fig.6, the roadway is 
modelled as a continuous stripe with certain ranges for grey 
values and extent, i.e. width. The roadway itself is composed of 
various parts, the road markings and roadway margins. While 
the road markings are of the object type periodic or continuous 
line, the object type of the margins is a continuous stripe. 
Attributes for grey value, extent and periodicity are assigned to 
the object parts of the roadway as well. The declaration of the 
spatial relations between these object parts is essential for the 
scale adaptation process, as previously described in section 3. 
Here, the distance dl represents the width of a single lane, d2 
corresponds to the distance between the outmost edge line and 
the roadway margins and d3 locates the optional road work 
marking from the outmost edge line. All nodes of the net are 
connected to the appropriate feature extraction operators, but 
only the operators connected to the bottom nodes are used. In 
addition, the boundary object parts are labelled to facilitate the 
search for adjacent objects. With this information groups of 
objects can be formed, which have to be analysed in 
conjunction regarding scale space behaviour. 
The following semantic nets represent some instance nets of the 
adaptations to smaller scales of this particular roadway scene. 
The adaptations are done manually based on a visual inspection 
of larger ground pixel sizes, i.e. smaller scales of the image as 
seen in Fig.7. The adapted nets correspond to a selection of 
smaller scales. At these scales at least one event in scale space 
necessitates the adaptation of the previous net, which is 
appropriate for a larger scale. 
   
  
Figure 7. Aerial Image, Dual Carriageway 
As scale is decreased, first the object type of the central object 
part, the double white line, changes from continuous stripe to 
continuous line, cf. Fig.8. Secondly, the road work marking 
merge with the neighbouring right lane marking due to the 
small separating distance between them. Since road work 
marking is an optional part, the term of the lane marking is 
maintained for the name of the resulting object part. The 
attributes of this modified object part, however, change. The 
resulting object type is continuous line. 
With further decreasing of scale the edge line markings merge 
with the roadway margins, both left and right side. Fig.8 depicts 
the semantic net adapted to this scale. The nodes of the edge 
line markings are combined with the nodes of the roadway 
margins, resulting in new values for the attributes, grey value 
and extent. Since these new object parts are now located at the 
border of the entire object dual carriageway, they have to be 
labelled as boundary objects. Even though the feature extraction 
operators are not included in Fig.8, the connections to the nodes 
still exist and the operators are called for the extraction of the 
bottom object parts in the image. 
When the scale gets so small that single lines are not detectable 
anymore, the lane markings vanish. Distances between the 
remaining object parts, double white line and roadway margins 
have to be modified, cf. Fig.9. The stripes of the roadway 
margins shrink to 2 pixels in width and thus, the object type 
changes to a continuous line.