tanbul 2004 
lation. The 
erc are only 
neighboring 
a reference 
of the road 
of gravity of 
to separate 
ed from the 
ferent slope 
'en points is 
ace and the 
of a triangle 
ope and the 
most likely 
erline of the 
ies could be 
end to have 
‘a tent, and 
le envelope 
found (only 
ones. which 
ts for these 
e the higher 
on polygon 
of vehicles 
ec Figure 3 
(Sederberg 
sed data is 
es between 
lese can be 
in based on 
clusters for 
a principal 
stretches à 
information 
'ntation has 
sing image- 
  
eight, after 
'k is regular 
he minimal 
sed bilinear 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B2. Istanbul 2004 
  
Figure 7. Binary image 
[t is easier to use a binary image for classification, therefore, the 
original image has to be converted, using an appropriate 
threshold value. As described above, because of the elevation of 
the road, the same pixel value could represent both vehicle and 
road points. (It can be seen in Figure 5 and 6.) 
In Figure 7, the labeling method can be easily applied to get 
connected components, which based on the investigation of the 
neighboring pixel values (4 or 8 neighbors), and sort them into 
classes; each separate class gets a unique identifier. 
6. CONCLUSIONS 
This paper demonstrated a way to extract vehicles from a 
LiDAR point cloud to provide data for vehicle classification. 
The comparison results of the introduced three methods are 
shown in Table 2. 
  
  
  
  
  
  
  
  
Route 35. Dayton, | Toronto, Canada 
Ohio 
Thresholding 4/4 13/14 
Triangulation 4/4 14/14 
Image labeling 4/4 12/14 
  
Table 2. Segmentation methods comparison 
(Extracted vehicle/present vehicles) 
In Figure 8, the results obtained for one vehicle are shown. 
Because of the automatic Delaunay triangulation, which 
operates without any constraints, one point is missing, but it 
provides the most precise boundary of the vehicle out of the 3 
algorithms. The fastest method is the thresholding-based 
technique. 
Figure 8. Vehicle points with polygon boundary 
(Red - Triangulation, Green - Labeling. Blue - Thresholding) 
The results are promising. but further work is needed for refined 
segmentation. In addition, using denser point cloud is expected 
to result in better point selection. 
Acknowledgements 
The authors would like to thank to Woolpert LLC and Optech 
International for providing the LiDAR datasets. and to Charles 
K. Toth, Center for Mapping. The Ohio State University, for his 
extremely useful contribution during the research. 
References 
Barsi, A., Detrekéi, A. Lovas, T.. Tóvári, D., 2003. Data 
collection by airborne laserscanning (in Hungarian). Geodézia 
és Kartográfia, Vol. LV, No. 7, pp. 10-17 
Tovari, D., 2002. Analysis of airborne laserscanner data. 
Toth, C. K.. Barsi, A., Lovas, T., 2003. Vehicle recognition 
from LiDAR data. International Archives of Photogrammetry 
and Remote Sensing, Vol. XXXIV, Part 3/W13. pp. 162-166 
Lovas, T.. Barsi, A., Toth, C. K. 2004: Detecting Moving 
Targets in Laser Scanning, Proc. ASPRS Annual Conference, 
May 23-27, in press 
Pitas, I. (2000): Digital Image Algorithms and Applications, 
John Wiley & Sons, Inc. 
Sederberg, T. W. and Anderson, D. C..1985 .steiner Surface 
Patches," IEEE Computer Grapichs and Applications, May 
1985