3.4 Lane Detection 
With this procedure we refer to the approximation of 
the detected lines by analytical functions. These curves 
define lanes of the highway. Once the vectorization 
procedure has been completed a large number of polylines 
exists in the images. À few of them correspond to the 
edges of the road, most represent edges of other objects, 
such as cars, traffic signs, light poles, and trees. To detect 
the lanes of a highway, which are defined by the center and 
the edge lines, we have to eliminate all polylines which do 
not correspond to the real road edges. We implemented a 
number of rules which form a knowledge base that allows 
the computer to decide on the most likely candidates for the 
road edges. The major rules are given below: 
(a) We only want to keep straight lines as road edges. 
Therefore, we approximate each segment by a 
straight line (1). 
x=k*y+b (1) 
If the standard deviation between the straight line 
and the approximation is smaller than a threshold, 
this edge line is kept as a candidate. 
(b) As the edge and center lines of the road are 
parallel in object space they must appear as 
intersecting straight lines in the image (a central 
perspective projection). The intersection point is 
also called vanishing point. It is the location 
where parallel object lines intersect in infinity. 
The vanishing point is close to the center of the 
image; as the cameras of our stereo-vision system 
are tilted downward by about 89 the vanishing 
point appears in the upper half of the image. In 
order to keep a line as a candidate it has to satisfy 
condition (2), which says that it must pass through 
an area around the approximate vanishing point. 
pixel/3 < x = k*y + b < pixels * 2/3 
for y = scan/3 (2) 
(c) Next, we collect line segments of similar slope 
and offset of the origin (k and b) into groups. 
Additionally, we include information about the 
lengths of these lines. Segments collected into 
one group are defined by similar parameters of k 
and b. Each group of lines is defined by only one 
straight line, which is the best approximation of 
all segments contained in this group. As the edges 
and center lines of the road are much longer than 
all the other lines, we can simply distinquish by 
keeping only significant groups as edge 
candidates. A criterion for the selection is the 
number of individual points which is represented 
by a specific straight line. 
(d) Finally, we have to approximate the edges and 
center-lines of the road in each of the selected 
groupsby analytical function. A straight line is 
defined through the pixels of each group. This 
straight line really represents the edge of the road 
in a certain distance in front of the van. Of 
course, if the road is bending these 
approximations will no longer be true. However, 
in a more sophisticated procedure the straight 
lines can be approximated or replaced by 
analytical curves. 
3.5 Intersection of Linear Features 
The procedure discussed in the previous chapter results 
in a number of straight lines in each of the stereo-images. 
As we are mainly interested in the object coordinates of 
these feastures they must be projected into object space. 
For this purpose we have to find two corresponding points 
for each line in the two images. 
For any point selected in the left image we can find the 
conjugate point in the right image by intersecting the 
epipolar line with the road edge. If we do this for at least 
two points we can easily compute the object coordinates by 
the collinearity equations. These two points are connected 
by a straight line which represents the road edge in local 
coordinates relative to the GPS-Van. By applying 
information about the position of the van and its orientation 
the global coordinates of these edges can be computed. 
The procedure described here can be performed for 
each stereo-pair of a sequence. The parallel lines which are 
created for each image-pair can be plotted in the same 
world coordinate system and be displayed as a road map 
(figure 3). 
  
121