In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds), IAPRS, Vol. XXXVIII. Part 3A - Saint-Mandé, France. September 1-3, 2010 
302 
centre lines and average widths. The centre line of a road string 
is composed from the centre lines of the individual road parts 
and the connecting lines between them, and then approximated 
by a polygon approximation. The polygon approximation starts 
with a straight line between the end points of the centre line and 
iteratively adds vertices if the distance of the approximation to 
the original is too high. The average width is calculated from 
the average widths of the individual road parts. A road region is 
calculated from the centre line and the average width. The new 
centre lines are used for the connection of the road network. 
In the first step of the network generation, pairs of parallel roads 
that lie close together are searched for. From such a parallel 
pair, only the road that has the better quality measure is kept. 
Additionally, roads that are overlapped by other roads for a 
substantial part of their road region (defined by their centre 
lines and average widths as described above) are also deleted. In 
this way, some false extractions can be eliminated before the 
actual network linking. After that, junction connections are 
searched for among the remaining roads. At the end of each 
road a search region is defined as a semicircle whose centre is 
the end point of the road centre line and which points in the 
direction of an extension of the road. The radius of the search 
circle depends on the quality measure of the road: a road with a 
good quality measure has a large search radius. If another road 
is found inside the search region, a junction connection is 
created. Depending on whether both roads are collinear or not, 
the junction connection is created in different ways. If the roads 
are collinear, i.e. have a small direction difference, they are 
connected if the end point of the second road lies inside the 
search region of the first, and the junction connection is the 
connection of the two end points. If the roads are not collinear, 
the junction connection is constructed from the extension of the 
first road, and, if necessary, from the extension of the second 
road, i.e. the junction connection is either the connection 
between one road and the intersection point on the other road or 
it is the short polygon connecting the two end points via the 
intersection point. Additionally, intersections between two 
roads are searched for. Junction connections are verified before 
they are accepted: if several competing junction connections 
(e.g. two parallel but not collinear roads) exist at the end of a 
road, only one is kept. The junction connections are evaluated 
according to their length: shorter junction connections are 
considered more reliable. 
After the creation and verification of junction connections, the 
road network consists of one or more connected components. A 
connected component consists of at least one road and possibly 
junction connections. Connected components are checked for 
significance: the total length of all roads must be more than the 
total length of all junction connections, and the total length of 
the connected component (roads and junction hypotheses) must 
exceed a minimum. An exception for the last condition is made 
if at least two open ends of the connected component lie near 
the image border; then it is possible that the connected 
component belongs to a road network beyond the image border. 
3. EXPERIMENTS 
The approach was tested on two different data sets. The first 
data set consists of an orthophoto generated from a scanned CIR 
aerial image of a suburban area in Grangemouth (Scotland) with 
a resolution of 10 cm, and a DSM obtained from image 
matching with manual post processing. The second data set 
consists of orthophotos generated from digital CIR aerial 
images with a resolution of 8 cm and a DSM from LIDAR data 
from the DGPF (German Association for Photogrammetry and 
Remote Sensing) test site at Vaihingen (Germany) (Cramer and 
Haala, 2009). In our tests, we used six image subsets from the 
Grangemouth scene and three subsets from the Vaihingen 
scenes, each of them depicting suburban scenes. The roads 
extracted by our method were compared to a reference to assess 
the completeness and the correctness of the extraction results. 
3.1 Results 
Figures 1, 2 and 3 display the results from the subgraph 
generation for three of the subsets used for evaluation. Figures 1 
and 2 show subsets from Grangemouth, whereas Figure 3 is 
taken from the Vaihingen data set. The subgraphs consist of the 
extracted road parts and the connecting lines found during the 
subgraph generation and evaluation. The results show the 
subgraphs after the evaluation, which means the road subgraphs 
consist of only one road string each. The subgraphs are depicted 
in different colours; roads that belong to the same subgraph 
have the same colour. 
Figure 1. Accepted road subgraphs, subset 1 (Grangemouth). 
Figure 2. Accepted road subgraphs, subset 2 (Grangemouth). 
Large parts of the road network could be extracted as road parts. 
Areas where the extraction fails typically lie at the image border 
(most notably in subset 2), at sharp turns or where the 
appearance of the roads is disturbed by trees and shadows. False 
extractions are rare, thanks to the DSM; most of them are 
driveways or parking lots. After the subgraph generation, most 
road parts that lie on the same road are connected. In subset 2, 
two road parts were first connected across two buildings, but 
the connection was eliminated after the context object 
evaluation (white dashed line in Figure 2). One connection in 
subset 2 was missed (white dotted line in Figure 2).