567 
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vo!. XXXVII. Part B3b. Beijing 2008 
easily discarded in a later step when the road network is 
constructed. 
Some roads are covered by several road parts with gaps 
between them. These are connected in the next step (Fig. 6). 
Figure 6. Assembled road parts. Connections between road 
parts are shown as lines in the same colour. Intensity image 
used for clarity of display. 
Road parts that lie on the same road have generally been found. 
In the case of the group of road parts displayed in yellow, at the 
top of the image, two road parts were connected to the same end 
of the first road part. This branching is permitted so that no road 
hypotheses are lost. In a later step both alternatives will be 
examined and the better one will be kept. The blue road part 
between the yellow ones was not added to the group because of 
the overlap with the left yellow one. 
The figures 7 and 8 show the results of two further subset 
images. Here the roads were typically covered as a whole by 
one road part so that the assembling step had no effect. 
Figure 7. Extraction result on second subset. 
In these examples the majority of the roads are covered by 
extracted road parts. There are few false positives, and those are 
mainly small and could be eliminated in a later step. In 
summary, also considering subsets that are not shown here, 60- 
70% of the roads are found. 
Figure 8. Extraction result on third subset. 
4. CONCLUSIONS AND OUTLOOK 
In this paper, an approach for the extraction of. roads in 
suburban areas was presented. The results show that this 
approach is applicable to suburban areas. The majority of roads 
could be extracted as road parts. The number of false positives 
is small, and as the results presented here are of an intermediate 
stage, we are confident that these false positives can be 
eliminated in a following step. 
Several parts of the algorithm still need to be improved. For 
example, the splitting step does not always succeed in dividing 
the regions in a meaningful way. If the border of the region is 
very irregular, the splitting can be incomplete. Also, loops, as in 
the not extracted road on the right hand in Fig. 7, are not 
handled properly by the current algorithm. 
The width constancy value can only be calculated meaningfully 
if the road is not curved too much. If it is, the two endpoints of 
the centre line cannot be derived from the points on the border 
that are farthest away from each other. In this case the centre 
line is calculated wrongly and a road segment would fail the 
width constancy test. Fortunately, curved roads are rare in 
suburban areas. But this problem can also occur if two roads are 
connected in one segment at a junction and the splitting step 
does not separate them because the skeleton just has a sharp 
bend there instead of a junction. 
The parameters are currently defined empirically which will 
probably lead to problems if the approach is applied to images 
of another area. The combination of the criteria for grouping, 
road part extraction and assembling is currently done in an all- 
or-nothing way; all criteria have to be fulfilled. A better method 
would include weighing the criteria against each other. 
The next steps of our work will include dealing with the above 
mentioned issues as well as completing the road network 
extraction. For completing the extraction, first the road 
subgraphs which contain several branches, like the yellow one 
in Fig. 6, need to be examined in order to find the best solution 
to solve the ambiguity. Then, the roads can be connected to a 
road network by searching for junction hypotheses at the end 
points of roads. False positives can be eliminated in this step 
because they would mainly be isolated.