International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
  
The computation of best direction of the road in the last 
extracted point is carried out by the template which is a matrix 
of m rows and n columns. For matching the template with the 
road: 
— The shorter side of template m is chosen equally to the 
road width 
— The longer side n is taken several times greater than m. 
The points of the template, which is designed to rotate 360° 
around the last extracted point, are resampled from the original 
image by bilinear interpolation. The new extrapolated point and 
the best road direction are used by the edge-based extraction 
process to extract a new road center point. Whenever this 
process fails in finding a new point, the template is translated to 
the new extrapolated point and the computation of the best local 
direction and the piecewise linear extrapolation are repeated 
and so on (see more details and equations, Poz, 2001). 
This method’s extraction process is based on the edge-based 
model presented by Nevatia and Babu (Nevatia and Babu, 
1980) and a little modified by McKeown and Denlinger 
(McKeown and Denlinger, 1988). First, a 5x5 Sobel gradient 
operator is applied to the points along the line that is 
perpendicular to the best road direction at the extrapolated point 
. The Sobel values (direction and magnitude) for each point are 
used to compute a score based on the weighted sum of three 
component scores. Each of these components is a linear 
function varying in the interval [0; 1]. The component scores 
are the edge strength, the difference in orientation between the 
edge orientation and the road orientation, and the difference in 
orientation from each neighboring point (Poz, 2001). 
In order to compute the component scores, the Sobel is first 
applied to every sampled points of cross-section of the 
extrapolated point and the point sequences whose magnitudes 
exceed a pre-defined minimum magnitude are detected. The 
point sequences are potential edge region, which are two for 
cases close to the ideal situation. The following values are then 
stored for all detected sequences: 
— The point coordinates, 
— Sobel magnitude and direction 
— The difference in direction between neighboring points 
— The maximum magnitude. 
In order to allow sub-pixel precision, the selected distance 
between sampled points needs to be less than 1 pixel. The next 
step is to compute the component scores and the resulting score 
for each point of the detected sequences on both sides of the 
road (see more details and equations in Poz, 2001). Having 
computed the maximum scores on both sides of the road center, 
three cases can be identified: 
— Two edge points are found on both sides of the road 
center; 
— One edge point is found on one side of the road center 
— No edge point is found 
If first case occurs, it means that the road center is computed as 
the middle point between the two edge points. If second case 
occurs, the road center has to be computed based on the 
extracted edge point, the computed road width and the current 
cross-section orientation at the point. On the third case the 
520 
extrapolation process guesses ahead and another attempt is done 
to extract another point. 
3. APPLICATIONS 
These two approaches previously investigated are applied by 
the help of Borland C++ software. 500x500 pixels (each pixel is 
| m) size black and white image is used for the tests. In both 
methods the same image is used. 
Figure 1 shows the results of the road tracer by profile matching 
and Kalman filter. The traced road is shown in green (means 
good profile matching) and red (means bad profile matching). 
The red parts show that the road position is based on the time 
update only. It works quite good on most parts of the road. 
Especially at the region where the trees are and at the 
intersections where the profile matching fails, it works very 
efficiently and is able to correctly predict the road path. The 
experiment shows that this method has some problems at the 
big curvatures. 
  
  
Figure 2. Road extraction based on edge and correlation 
analyses 
Figure 2 shows the results of the road extraction based on edge 
and correlation analyses. The extracted road is shown in red. 
The visual interpretation of results reveals that the method 
works satisfactorily on most parts of the road. This method has 
some problems where the road width changes too much, 
on 
Sp 
Po 
on 
Ce