ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision", Graz, 2002 
  
All found parallel edges are considered as Possible Road Sides 
that are Parallel (PRSP). They compose a weighted graph. The 
nodes of the graph are PRSPs, the arcs of the graph are the 
relations between PRSPs. Note that in occlusion areas, the arcs 
also represent the missing parts of a road between a pair of 
PRSPs. The width of two PRSPs should be similar. If there is 
no gap between two PRSPs, i.e. one PRSP shares points with 
another, and the linking angles between them in 3D space 
comply with VEC25, they are connected directly. In case of an 
existing gap, we first check the connecting angles between 
PRSPs and the gap. If the angles comply with the VEC25, the 
gap area is further evaluated using additional information, and 
we compute the possibility of the gap belonging to road. This is 
called reasoning process in our work. If the gap is not too long, 
and 
e within the gap is a road region, or 
e within the gap is a shadow or shadow mixed with road 
region, or 
e the gap is caused by tree occlusion (determined from the 
image classification results and nDSM data), or 
e within the gap is terrain as determined by the DSM, or 
e road marks are extracted within the gap 
Then, we consider the gap as possibly belonging to a road. 
Suppose N is the total number of pixels in the gap, and Nr, Ns 
are numbers of pixels of road and shadow respectively. Ng is 
the number of pixels of ground objects. The possibility of the 
gap belonging to a road is computed as P, — w, * w, , where w, 
and w, are measures using height and image information 
respectively. They are given as 
Neg r E 
a a (b 
The road is then found by searching the graph using a best-first 
method that maximizes the length while minimizing the 
curvature difference between the extracted road and VEC25 in a 
merit function. The function is defined as 
(j * lgap t 1j) * wc (2) 
i and j are adjacent PRSPs, /; and /; are their lengths, /,,, is the 
gap length between i and j, w, is a measure that is inversely 
related to the curvature difference between the curve formed by 
i, j and the corresponding curve on VEC25. The function 
defined in (2) gives high values to long curves that have similar 
curvature to VEC25. 
For main roads, on which the system knows that road marks are 
present, the system also extracts roads using detected road 
marks and zebra crossings. The road marks are linked using a 
similar method as described in the previous paragraph. This 
procedure increases the effectiveness and reliability of our 
system. In complex areas, such as in city centers, the roadsides 
are usually occluded very much, and sometimes it is impossible 
to identify them. However, the road centerlines are successfully 
extracted by the system using road marks. In rural and suburban 
areas, the extracted road using road marks is used by the system 
to verify the extraction results using 3D parallel edges. 
5. RESULTS 
The described system is implemented as a standalone software 
package with a graphic user interface running on SGI platforms. 
The system reads color stereo imagery, old road databases and 
other input data, and outputs the extracted roads in 3D Arc/Info 
Shapefile format that is readily imported in existing GIS 
software. The system has been tested using different datasets in 
various landscapes. Reports of the test results and the system 
performance can be found in Zhang et al. (2001a) and Zhang et 
al. (2001b). A recent benchmark test has been conducted 
independently by our project partner Swiss Federal Office of 
Topography with several stereopairs. We will show in this 
section some results of the benchmark test. Besides, we also 
tested our system on quite different image data provided by the 
National Geographic Institute (NGI), Belgium. The test results 
and performance evaluation will be given. 
The benchmark test images were over Thun, Switzerland. The 
area fluctuates from 550m to 2200m. Almost all types of roads 
in Switzerland can be found in this area. The image data have 
the same specifications as described in Section 1. During the 
test, our system was only applied to extract roads in rural areas, 
while roads in urban and forest areas were not processed. 
Figure 3 presents a portion of 3D road extraction and road 
network generation. The landscape of Figure 3 includes open 
rural and forest areas distributed with some houses. All the 
roads in this area are correctly extracted by our system. The 
details of automatic 3D road extraction and junction generation 
are presented in Figure 4, where the outdated roads from 
VEC25 are shown in white lines, and the extracted roads in 
black lines. Figure 4a is an example of road extraction and 
junction generation in rural areas. Figure 4b shows a highway 
with 4 lanes, the system extracted the lane border lines and lane 
centerlines through road mark extraction. 
Figure 5 presents some results from the Belgium site. The image 
data and the old road databases are provided by the National 
Geographic Institute, Belgium. The test area is generally flat. 
The image scale is around 20,000, and the camera focal length 
around 150mm. The image is black and white, and scanned 
using a PS1 scanner with 15 microns. As can be seen in Figure 
6, the image does not have good quality, and many road-like 
lines are observed in the fields in the test site. A DTM in the 
test area is provided with 40m interval, the RMS error is around 
10m. The RMS error in old road databases is around 9m, with 
maximum one around 25m. Some road attributes, e.g. road 
width, can be derived from the old databases. However, these 
parameters are too strict, they have to be relaxed for the test. We 
did not change anything in our system except in the image 
classification procedure. With black and white image, we only 
cluster a single band data to try to find road regions. Figure 5 
shows a portion of the test results. The roads in rural area are 
correctly and reliably extracted by our system. In Figure 6, the 
details of road extraction and junction generation for this 
dataset is presented. 
In order to evaluate the extraction results, we developed a 
method in previous paper to compare the extracted roads with 
the reference data. The method was applied to the datasets of 
Switzerland and Belgium. The reference data were measured by 
L+T and NGI at analytical plotters. The quality measures aim at 
assessing exhaustiveness as well as geometrical accuracy. To 
evaluate exhaustiveness, completeness and correctness are used. 
Completeness measures the amount of the reference data that is 
covered by the extracted roads, while correctness is the amount 
of correctly extracted roads covered by the reference data. The 
geometrical quality is assessed by mean and RMS of the 
distances between the extracted road and the reference data. The 
comparison results for Figure 3 and Figure 5 are listed in Table 
2 
A - 403