2. HARDWARE COMPONENTS 
Absolute positioning of the GPS-Van is achieved by 
two surveying type GPS-recievers (Trimble 4000ST). One 
serves as a base-station at a known location, the other one 
is a rover station mounted on the van. When satellite 
signals are blocked a dead-reckoning system takes over. It 
consists of a directional and a vertical gyro, and a wheel 
counter. The directional gyro measures horizontal angular 
changes (directions), the vertical gyro determines two 
angles (pitch and roll) that measure the van's tilts relative to 
the vertical. Additionally, a magnetic proximity sensor 
counts wheel-rotations at the disk brakes of the two front 
wheels. Together these sensors generate the absolute 
positions of the GPS-Van and its orientation (attitude) at 
any time. 
For relative positioning the stereo-vision system was 
installed. It consists of two fully digital CCD cameras 
(Cohu 4110) with a resolution of 732 x 484 pixels. They 
are mounted on a rack on top of the vehicle. We assume tht 
they are rigidly attached to the van and do not change their 
attitudes during operations. The two cameras directly 
interface to a real-time imaging system (Trapix Plus from 
Recognition Concepts Inc. (RCI)), where the images are 
temporarily stored in a frame buffer. They can also be 
processed on-line using a digital signal processor, or they 
can be sent to the Data Store real-time disk, which has a 
data transfer rate of 4 MBytes per second and holds 2 
GBytes of digital data. It is interfaced to an Exabyte digital 
tape drive through a SCSI connector. 
Finally, a touchscreen is used to control operations of 
the data-collection procedure and to key in a number of 
pre-defined attributes as the GPS-Van passes an object of 
interest. A color-video camera is applied for photo-logging 
of the road environment; the video scenes are also related 
to the GPS-positions. All sensors of the mobile mapping 
system are controlled by a PC. The system configuration is 
shown in figure 2. 
Stereo-Vision System (RCI TrapixPlus) 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
BC p IDEN O38 MByiep {D/A ) 
& P. 
à KDPI RAM converte ; 
= VIShd —7 Hard 
* Adapte E disk 
3 Pixel KRTP Da bod 
© Processor | sm ag d 
8 ere Adapte ze Dy'E 
= FLKRTP =. ape 
5b igital Pott [ 
Bc zi Interface [TT S Mn 
KDPI 
msn 
Interfa 
  
  
  
Figure 2: Hardware components of the GPS-Van. 
3. EXTRACTION OF ROAD EDGES AND CENTER- 
LINES 
One of the most important features we want to extract 
from highway image-pairs are edges and centerlines of 
roads. Two different approaches were developed: the first 
one relies on the definition of the edge or centerline by a 
certain pattern, e.g. by a bright line. Well defined patterns 
can be found fully automatically by the computer, once 
detected, they are approximated by analytical functions. 
The second approach permits the user to define one point of 
an edge on the screen. A line following algorithm traces 
this edge from the bottom to the top of the image until the 
edge disappears. In both cases we work in both images of 
the stereo-pair, which means that the edges can be directly 
related to a global object coordinate system. 
120 
3.1 Automatic Detection of Bright Line Markers 
There are many edge detection algorithms available. 
Most of them require considerable computer-power such as 
the LoG operator. For our special application we prefer to 
use a very simple technique, which should work in a 
horizontal scan line or in a narrow horizontal window of 
the digital image. In most cases we know where the center- 
line or the edge of the highway appears in the image-pair. 
Therefore, we can limit the search area and accelerate the 
procedure. 
Our algorithm is based on the assumption that the edges 
of the road are marked by solid bright lines and the center- 
lines or lane dividers are marked by bright dashed lines. 
We perform the following procedure to directly detect these 
bright lines in the highway images: 
* We calculate the average and standard deviation of 
the gray values (ga, Og) of each scan-line, 
* A median filter is applied to the scan-line to 
eliminate noise. Typically, a window of five pixels 
is used, the median gray value is assigned to the 
center pixel if the difference between the median 
value and the average is larger than 1.5 times the 
standard deviation the value is replaced by 1 
(white). Otherwise, the pixel is set to O (black). 
This means that we detected a pixel corresponding 
to a bright line image. This procedure is performed 
for every pixel of a scan-line. 
* If a number of bright pixels appear side by side, 
their center point is used to represent the edge. We 
also make sure that there are no discontinued line, 
in the images. This method detects only very bright 
linear features, which run from top to bottom. The 
result is a binary image. The extracted edges are 
represented as a raster and need to be vectorized for 
further 3-dimensional representation. 
3.2 Thinning of Binary Images 
Before the detected features can be vectorized their 
width must be reduced to one pixel. This means that the 
center pixels of the detected features have to be used to 
represent the lines. The principle of this procedure is to 
keep the central point at its correct location while 
iteratively replacing the boundary of the white (1) areas by 
0 values, which corresponds to shrinking the width of the 
line. 
3.3 Line Following 
In the next step we convert from raster data (digital 
images) to vector data (lines). This procedure starts at the 
corner of the image and looks for the first white pixel 
(value 1), which represents a pixel of the line. The 
algorithm tries to follow the line continuously replacing 
each processed pixel by 0, until the end of the line is 
reached. Then we procede to the next pixel and apply the 
same procedure. This method is chain coding. It enables 
us to define a line through the digital image without 
specifically writing any coordinates, but by expressing 
dependencies between consecutive pixels. 
After following a line its length is com 
| puted and 
checked. All lines which are shorter than a threshold (in 
our case shorter than 10 pixels) are eliminated. If point i as 
well as (i— 1) and (i + 1), appear to be on the same linear 
edge, point i is not registered. Thus we can reduce the 
amount of data. 
3.4 
the 
dei 
pre 
exi 
ed; 
suc 
the 
the 
no 
nu 
the 
TOZ