The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
whose length equals the average road width and whose 
orientation is perpendicular to the secondary direction is moved 
along the secondary branch. The first place where the line 
intersects with the borders of the region on both sides of the 
skeleton branch is the place where the region is divided. After 
all branches have been examined, a second step follows in 
which the region is divided at places where the skeleton of the 
smoothed segment was disconnected due to the smoothing 
operation. Fig. 1 shows how one segment is partitioned. 
Figure 1. Partition of large segment. Original segment borders 
in yellow, skeleton in green, dividing lines in red. 
After the division of large segments, the evaluation in order to 
extract road parts follows. The criteria by which a segment is 
evaluated are: 
• intensity 
• NDVI (normalized difference vegetation index) 
• elongation, combined with convexity 
• width 
• width constancy 
Intensity and NDVI are radiometric criteria. The intensity 
should be higher than a threshold to exclude shadow regions, 
because shadow regions, for example of buildings, otherwise 
often have similar characteristics to road parts. The NDVI 
should be below a threshold in order to exclude areas with 
vegetation. 
The other criteria concern the shape of the region. A road part 
should be elongated, that means the ratio between the squared 
perimeter and the area should be high. If the area has a high 
convexity value (the ratio between the segment area and the 
area of the convex hull), lower values for elongation are 
permitted in order to include shorter road parts but to exclude 
regions with ragged borders, which also have a relatively high 
elongation according to the criterion used here. The width of a 
road region should be close to the average road width, and it 
should be relatively constant. For the calculation of the width, 
first a centre line is calculated for the region. This is done by 
first finding the two points on the boundary that are farthest 
away from each other. At these points the region boundary is 
split into two parts and for both parts a distance transform is 
calculated. The points where both distance transforms have the 
same values make up the centre line. The average width of the 
region is calculated from the distances of the centre line to the 
region borders. Twice the average distance from the borders 
gives the average width of the region which for a road part 
should not be too far from the average road width. The width 
constancy is defined by the standard deviation of the width 
divided by the mean value of the width. This value should be 
below a threshold. 
All regions are checked for these criteria, and those regions that 
fulfil all criteria are selected as road parts. The values for 
elongation, NDVI, width constancy and deviation from average 
road width are saved as evaluation results in order to give a 
quality measure of the road part. The results are mapped on an 
interval between 0 and 1 such that values that suggest higher 
probabilities towards road parts are close to one. Then the 
transformed values are multiplied to obtain a single quality 
measure. 
2.4 Assembling of road parts 
In many cases, one road part covers one complete road, from 
one junction to the next. But this is not always the case; 
sometimes one road is covered by several road parts with gaps 
between them. Therefore, in this step it is checked if the road 
parts have neighbours with which they can be connected. 
The search starts with the road part that has the best evaluation 
results from the step before. The intersection points between the 
centre line and the segment borders are used for calculating the 
criteria which determine if a road part in the neighbourhood is 
added to the examined road part. The criteria are: 
• distance between the segments, measured from the 
endpoints of the centre lines 
• direction difference between the road parts 
• continuation smoothness between the road parts 
The direction difference is measured by comparing the 
directions defined by the endpoints of the centre lines of both 
road parts. The continuation smoothness is determined by 
calculating the direction differences between the directions of 
the road parts to the direction of the connection between both 
road parts (Fig. 2). The smoothness is low if the differences 
between the directions of the road parts and the direction of the 
connecting line are high. The distance and the direction 
difference should be low and the continuation smoothness 
should be high for two road parts to be linked. The search for 
neighbouring road parts continues until no more road parts can 
be added. If no neighbouring road parts are found, the road part 
constitutes a road subgraph on its own. Then, the search 
continues with the next road part with the best evaluation result 
until all road parts have been examined. 
565