, .. . '~v 
BttAaMi 
urn*. 
The images !(HSI')- G//, C5 and G/+//. were initially filtered 
with a median filter to reduce the salt/pepper noise, which comes 
from bad screening of the cables. Then, the images were 
represented on scales. The scale space representation does not 
only minimise noise after JPEG compression but also enables to 
enhance dominant information along the scales. For eacli(HSI)" 
Gj-f, Gs and G/+/-[ image the edge detection was performed on 
the coarsest scale level. Then, the sum of gradient images was 
calculated (Eq 2). 
V/ C (*,y) = 
dG 
dG H dG s dG 
——+ — + 2-+ 
dx dy dx 
| dG UH t dG ltH 
dy dx dy 
(2) 
In the next step the VIq (x, y) image was compared with 
threshold image (Gj) to remove all edges that do not belong to 
the road (Fig 2). 
/! 
Figure 2. Detected road edges. 
The edge linking was the last step in the boundary detection. 
Similar to models presented in section 2.1 the Hough 
transformation (Duda and Hart, 1972), (Gajdamowicz 1998) was 
applied (Fig 3). 
7T 
Figure 3. Detected 2D lines. 
To measure the road width one has to find the distance between 
two 3D lines. In this approach the road width is calculated in the 
following way: 
1. Extraction of two points, the cross-section 
between sloped lines belonging to road 
boundaries and horizontal line passing through the 
centre of the ROI (Fig 4). 
Figure 4. Intersection points on the road boundaries. 
2. Matching of intersection points with help of 
Cross-correlation in colour space and epipolar 
geometry constrains (Gajdamowicz 1998). 
3. Calculation of 3D co-ordinates of the intersection 
points and the distance between them. 
The measured road width is visualised in the original image (Fig 
5). 
£}le £tfit ^aw Qeteci ytfndcr# jfteip 
I « 1 » I a ! № l-u-l 
■HP-- 
Ghbd coorcinaloa 
X V i 
|325in 78 44555! 06 [218X7 
|3S6VM4~ 4430231.91 [21Sli 
I Length afvooto 1 pi 
[1379 
Figure 5. Results of automatic road width measurement. 
2.3 Test 
In order to test the model and the algorithms the georeferenced 
images were used. The images were acquired with the MMS On- 
Sight. The images were collected under various conditions e.g., 
different colour of the road, different light condition and different 
road environment. 
The cameras were situated on the roof of a van2 m above the 
ground. The camera system was forward looking. The images 
were acquired with a spacing of 5 m, 8 m, 10 m and 15 m 
depending on the complexity of the environment. Consequently, 
in urban areas the image spacing was 5 m and in open terrain the 
spacing was 15 m. The images were compressed with the JPEG 
compression algorithm with a factor 1: 30. 
The algorithms were tested on a limited data set representative 
for specific cases. The first test (Table 1. Test 1) contained 
images of bitumen, dark road with painted lines indicated 
centreline and road boundaries. The images were acquired during 
sunny weather with a light from behind the cameras. 
The second test (Table 1. Test 2) included images taken in an 
area with buildings on the left side and with the open area on the 
right side of the road. Additionally, some big trees were on the 
left side of the road. The road had no white painting indicating 
the road boundaries. The light was from the left side and 
perpendicular to the road axis. 
The images in the third test (Table l.Test 3) were acquired in a 
built-up area with trees on the left and right sides of the road. The 
road was dark (asphalt) and had no paintings indicating road 
boundaries. The side of the road was made from gravel and it 
was slowly changing into grass. The light was from the right side, 
perpendicular to the direction of photography. 
7B-2-3 
pigfl 
§1 
m 
mmw. 
m