The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Voi. XXXVII. Part B3b. Beijing 2008 
566 
3. RESULTS 
The approach was tested on CIR images of a suburban scene in 
Grangemouth, Scotland. The images with a resolution of 0.1 m 
cover areas of approximately 250 m x 200 m. 
For the segmentation the image has to be divided into subsets 
for computational reasons. The size of the subsets is 
approximately 200 x 200 pixels. Each subset is divided into 20 
segments. Fig. 3 shows a segmentation example. 
Figure 3. Segmentation with normalized cuts. 
Figure 4. Grouping result. 
The next step is the evaluation of the segments to extract road 
parts after large segments (over 200 m 2 ) have been split. In 
order to speed up the computation, only large segments are 
considered for splitting. Small segments can be ignored because 
the splitting is done only at junctions with branches longer than 
10 m. The segments that were extracted as road parts are shown 
in Fig. 5. 
In the next step, the segments are grouped. The parameters used 
for the grouping are summarized in Table 1, along with the 
parameters for the following steps. Fig. 4 shows the grouping 
result. 
Grouping 
max. mean edge strength 
50 
max. standard deviation of colour 
40 
max. colour histogram difference 
0.4 
min. length ratio of shared border (if direction 
difference > 60°) 
0.2 
Road part extraction 
min. intensity 
40 
min. elongation 
70 
min. elongation (convexity > 0,75) 
40 
max. NDVI 
0 
width 
3m - 16m 
width constancy (max. ratio of standard 
deviation to mean value) 
0.6 
Road part assembling 
max. direction difference 
30° 
max. distance 
40m 
max. parallel shift (direction difference 
30° 
between road parts and their connecting line) 
Table 1. Parameters for grouping, road part extraction 
and assembling. 
Figure 5. Road part extraction result. 
On most roads in the image, road parts have been extracted, 
except for the road on the right. That road was not extracted 
because the width constancy criterion was not met due to areas 
at the sides of the road that were merged with the road areas, 
and could not be cut off by the splitting algorithm. This is the 
main reason for missed road parts. Some parts are falsely 
extracted as road parts; these are mainly roofs of buildings. 
Most of them are small and isolated and probably could be