estimation, 
vide a good 
essed. First, 
the surface 
face). Once 
ich as travel 
dance of an 
hicle blobs. 
ques, both 
as standard 
 Delaunay- 
1d provides 
effectively 
eight of the 
e used the 
d form of it 
a 
tasets, with 
1 July 2000 
. whilst the 
> flight data 
  
Cs 
  
  
Im“ 
  
"anada 
  
a passenger 
t direction. 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
In the opposite direction, this value is under 10 (Lovas et al. 
2004). 
Figure | shows a cross section of the LiDAR strip, which is a 
view about the cross section at the centerline. 
EFFET, 
rg 
he 
er 
Bg 
  
Figure 1. Cross section of the road with surroundings 
In this paper we present three, different methods for vehicle 
segmentation (Figure 2). 
UE OW AQ IQ AP QUERI EST ET 
| Initial LIDAR data | 
i 
i Data filtering | 
| Keep road points only | 
  
  
Image preparation | Surface warping 
Surfacing from rough 3D | Remove elevation, using | | 
data, to a regular grid | interpolated plain 
x s i em | 
| 4 | ; 
Thresholding | 
Select points above the | 
road surface | 
Delanuay triangulation 
Slope measurement 
Image-processing 
Labeling, based on 
| 
| 
neigborhoods | 
tet 2 AT TS 
Get LIDAR points = ua Me — 
Select original points, [ Classification | 
correspond to labeling Create classes for each vehicle | 
result 
Figure 2. Data processing flowchart 
2. DATA FILTERING 
In Figure | not only the road but also the surroundings 
(vegetation. ground work, landmarks, transmission line and 
vehicles) can be seen. First, we can easily detach the points not 
belonging to vehicles. If the position of the centerline of the 
road, and the number of lanes and their width are known, the 
usable swath can be obtained. 
If the centerline is not given. we can develop a semi-automatic 
algorithm that is based on the cross sections. Roads are usually 
located on embankments. We have to mark one cross section 
and the road direction. then using the calculated parameters 
from the sample section (height of the trapeze. angles and 
lengths, road slope from centerline), and the basic properties of 
the road (angle of slope - both for the long and cross direction, 
curve radius). Then the same data for the next cross-section 
should be calculated, close to the last one (e.g.. 10 meters). 
Combining this with the original dataset we can decide whether 
the calculation is right or not. If the calculations are correct, and 
the matching is good, the middle position and the parameters of 
the given cross-section can be recorded. If not. the same 
calculation with the same parameters in a different position 
should be performed (rotating by a small angle around the 
middle point of the last recorded cross section ). If in this 
position a properly matching cross section cannot be found. this 
has to be ignored, and a shorter distance from the last recorded 
one have to be used. 
3. THRESHOLDING 
In order to perform vehicle extraction, we have to separate all 
points above the average road height in a local environment. 
We cannot accomplish that without knowing the road level at 
every position of the vehicles or other objects (e. g.. vegetation). 
Using a zone with a little bit smaller width than the sampling 
density, we can ensure that only one point can fall inside. A 
polyline connecting these points and the centerline represent the 
road surface. This should not be very accurate because we use 
only the first pulse reflected from the tops: the lowest part of the 
vehicle is the engine hood, which is higher above the road than 
the distance between the points (Figure 3.). All points above 
that surface possibly belong to a vehicle. This new set is the 
basis for our further examinations. 
14096 
4.4036 
4.403 
     
  
$ 
x 10 
    
  
1 
2 
a3 
Bl 
P 
   
Fieure 3. Test data in 3D: rear and front view 
In case of sloping roads the same height could represent a road 
and also a vehicle. In order to identify vehicles more easily we 
have to compensate for the slope of the road. The centerline of 
the road is given or being calculated only horizontally. In 
Figure 4, the sampled point heights are shown along the 
centerline. The long section of the road is shown, where the 
sloping angles are different, but can be approximated with lines 
segments (marked in red). Decreasing all point height to the 
value of the regression line, at the point's horizontal position 
this goal can be achieved (Pitas 2000)