32 
CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
Figure 8. Road parts extracted in subset 1 with DSM ( yellow). 
Figure 9. Road parts extracted in subset 2 with DSM (yellow). 
Figure 10. Road parts extracted in subset 3 with DSM (yellow). 
Compared to the visual impression of the extracted roads, the 
completeness and correctness values are relatively low. The 
computed correctness suffers from leakage at the borders of the 
road parts and from the fact that pavements, which are often 
extracted as roads, are not included in the reference data. The 
computed completeness would also be increased by 
constructing road parts corresponding to the gap edges that 
were accepted in subgraph evaluation. 
Figure 11. Road subgraphs (with DSM), subset 1. Different 
colours represent different road subgraphs. 
Figure 12. Road subgraph evaluation (with DSM), subset 1. 
4. CONCLUSIONS 
In this paper, an approach for the extraction of roads in 
suburban areas was presented, with the focus on a comparison 
between the extraction results achieved for image data alone 
and the results achieved for using a DSM as an additional 
information source. Our results show that the approach is 
suitable for the extraction of roads in suburban areas. The 
majority of roads can be detected even without a DSM, though 
there is a relatively high number of false positives, mostly 
buildings. Using a DSM improves both the completeness and 
the correctness of the results, primarily because buildings can 
now be clearly separated from roads. The correctness is 
improved because buildings are not extracted as false positives. 
The completeness is improved because incorporating the DSM 
into the grouping process provides a better grouping result from 
which more road parts can be extracted. Without a DSM, there 
are more subgraphs containing several branches, so that the 
importance of the subgraph evaluation is higher. The potential 
to find the real course of the road based on an optimisation of 
the interrelations between the road parts is shown in Figures 6 
and 7. Subgraph evaluation can thus compensate for the lack of 
height information in the road part extraction stage. However, 
the improvement caused by using the height information in the 
grouping phase cannot be compensated. Road parts that remain 
undetected due to a poor performance of grouping based on 
image data alone cannot be detected at a later stage. Using a 
DSM thus certainly improves the quality of the results. This can 
be seen in particular for subset 3 (Fig. 5 vs. Fig. 10). 
The road extraction process can still be improved in several 
ways. The parameters used for grouping, road part extraction 
and road subgraph generation could be learned from training 
samples, which probably would improve stability in different 
settings. The road extraction can also be improved by 
incorporating context objects such as trees, buildings and 
vehicles. It is planned to incorporate context objects into the 
evaluation of the gaps within the subgraphs, combined with the