ond 
ure. 
and 
for 
iain 
(2) 
the 
des 
een 
  
tion 
ber- 
nds 
ted 
the 
2 of 
ads. 
IV) 
s of 
oad 
loys 
In: Stilla U, Rottensteiner F, Hinz S (Eds) CMRTO5. IAPRS, Vol. XXXVI, Part 3/W24 -— Vienna, Austria, August 29-30, 2005 
4.1 Extraction of Lines 
Line extraction is carried out using the differential geometric ap- 
proach (Steger, 1998). Bright or dark lines can be extracted with 
sub-pixel accuracy based on only a few parameters, e.g., the max- 
imum width. Lines are extracted in the image channels as well as 
after the classification in the road class image to generate as many 
road candidates as possible. The road class image is useful as an 
additional channel for line extraction, because of the well known 
behavior of roads in it and the resulting simple and robust adjust- 
ment of the extraction parameters: Roads in the road class image 
always appear as salient bright lines with dark surroundings. 
To allow reasonable speeds for driving, roads usually have a lim- 
ited curvature. Thus, lines often change their meaning from roads 
to non-roads at points where the curvature is high. To deal with 
this, the extracted lines are split at points with high curvature and 
only then act as preliminary hypotheses for road parts for the next 
step. 
4.2 Road Hypotheses Verification 
Apart from the homogeneity of the grey value along the road, 
the assessment of the road hypotheses is done in (Wiedemann et 
al., 1998) only based on geometric conditions. Spectral proper- 
ties are not used. To use as much knowledge as possible for the 
verification of road hypotheses, we developed a new method for 
assessing road hypotheses. Particularly, we calculate fuzzy val- 
ues (Zadeh, 1989) for the following parameters: 
e length ur 
e average width uw 
e road energy, i.e., average membership value ju of the road 
hypotheses to the road class 
Figure 3 shows the criteria for the assessment of road hypotheses 
and the linear membership functions used for the calculation of 
the corresponding fuzzy values. The individual fuzzy values are 
combined into one final weight jg for each line with the fuzzy 
AND operator using equation (1). 
pr = pr A pw A pg = MIN(UL, Uw, LE) (1) 
The road hypotheses are considered to be verified and are used 
for the following generation of a road network, if the final weight 
Kr is above a given threshold. 
4.3 Road Network Generation 
For the verification the individual spectral channels are regarded 
separately leading to a strong redundancy. To generate a consis- 
tent set of roads, the results from the different channels are fused 
eliminating redundant roads in the same way as in (Wiedemann, 
2002). Starting from the fused roads a network is generated in 
two steps: 
First, remaining small gaps are eliminated based on a weighted 
graph. In this graph missing connections are detected by finding 
31 
  
Figure 3: Criteria for the assessment of road hypotheses together 
with the membership functions for the evaluation of the corre- 
sponding fuzzy-values 
the best path between pairs of seed points using the Dijkstra algo- 
rithm and calculating weights for the possible connection. A gap 
is closed if the connection is part of an optimum path, without 
making use of image information (Wiedemann et al., 1998). 
In the second step, larger gaps are closed using the approach 
of Wiedemann (Wiedemann and Ebner, 2000) extended with a 
new module proposed by (Bacher and Mayer, 2003, Bacher and 
Mayer, 2004) for the verification of potential connections. Poten- 
tial gaps are detected using the property of road networks, that 
most points can be reached from all other points with minimum 
detour. To make use of this property, link hypotheses are gen- 
erated. The distance of adjacent points is calculated once along 
the network and once along the hypothetical optimal path, i.e., 
if nothing else is known the Euclidean distance. From these dis- 
tances a detour factor is computed using equation (2). 
t k dist 
detour factor = rman wor i ange (2) 
optimal distance 
The link hypotheses are checked with the new module, detailed 
below, starting with the hypotheses with the largest detour factor. 
If a link hypotheses is accepted, the new connection is inserted 
into the road network. Due to changes in the network, the gener- 
ation of link hypotheses has to be repeated. This is iterated until 
no more new link hypotheses can be generated. The result of this 
global grouping step is the final road network. 
The module for the verification of link hypotheses starts with the 
two end points of the potential connection. The goal is to find the 
optimum path between these points with respect to the geomet- 
ric and radiometric characteristics of roads and use it to verify or 
reject the link hypotheses. For the determination of the path a zi- 
plock snake (Neuenschwander et al., 1995) is optimized between 
the two end points employing the image information of the road 
class image (cf. Fig. 4). To verify or reject a link hypotheses, i.e., 
to decide if the hypotheses actually corresponds to a road, a grey 
value profile perpendicular to the snake direction is calculated for 
every snake point in the road class image. When evaluating the