MODEL-BASED VEHICLE DETECTION FROM LIDAR DATA 
T. Lovas *, C. K. Toth”, A. Barsi® 
* Department of Photogrammetry and Geoinformatics, Budapest University of Technology and Economics, 
H-1111 Budapest, Hungary - tlovas@mail.bme.hu, barsi@eik.bme.hu 
" Center for Mapping, The Ohio State University, 
1216 Kinnear Road, Columbus, OH 43212-1154 USA - toth@cfm.ohio-state.edu 
Commission Il , WG Il/2 
KEY WORDS: LIDAR, Modeling, Automation, Data fusion, Terrestrial laser scanning 
ABSTRACT: 
Airborne laser scanning has established itself as a dominant technology providing high quality surface data for a variety of 
applications. LIDAR has substantially widened the use of mapping, for instance, in the late nineties, telecommunication industry 
required large volumes of high-density DSM data. In addition, research has recently shown promising results in extracting features 
such as man-made objects, for example buildings, from the point cloud. More recently, research started to explore the feasibility of 
using Lidar data for transportation applications, including infrastructure, emergency and environmental mapping along corridors. 
Initial investigation on assessing the performance of extracting vehicles from LIDAR data and then categorizing them has proved 
that valuable traffic flow information can be obtained. The vehicle classification was mainly based on simple four-parameter 
description of the vertical profile of the vehicles. This paper is a continuation of that research effort by introducing an improved 
model of the vehicle profile description. A model library is formed based on the ground-based laser scanning data and an analytical 
approximation of the vehicle profile will replace the previous four-parameter description. The anticipated benefits are twofold: 1) a 
better extraction and a more robust coarse classification of the vehicles are expected, and 2) it is very likely that subclasses of 
vehicles can be introduced such as small cars, full-size cars, light trucks, SUVs and so on. This paper describes a newly developed 
model of vehicle profile description, the classification method, implementation, and algorithmic aspects. Extensive tests have been 
carried out to validate the method and assess its performance. 
1. INTRODUCTION 2. FIRST APPROACH: PCA 
In our initial approach we derived basic geometric features of 
the vehicles, such as length, width and height values. First, the 
vehicles have to be segmented from the data set, which can be 
accomplished applying well known techniques, such as 
thresholding, or edge detection. However, it’s not a 
straightforward task, since various effects may corrupt the 
resulted vehicle point cloud in the LiDAR data set. 
In order to reduce the dimensionality of the feature space, 
principal component transformation had been carried out. The 
Due to decreasing sensor prices and improved navigational 
accuracy, laser scanning technology (also known as LiDAR) 
has become a dominant technology in topographic mapping. 
LiDAR rapidly produces high density, accurate spatial data 
with minimal need for post processing (Flood 1999). Due to its 
data acquisition. characteristics, LiDAR is mainly used for 
digital surface/elevation model (DSM/DEM) generation, but its 
application rapidly widens for instance in urban planning, 
agriculture or transportation. 
  
Transportation as a major momentum of modem economy 
creates even more complex tasks for engineers. Its 
environmental impact, accidents, and wasted time spent in 
congestion mean serious economical damage for motorized 
countries (Zhao 1997). Old city structures do not allow major 
changes in the road network; the problem should be solved in 
the area of operation and controlling. Nowadays, intelligent 
transportation systems (ITS) continuously gain ground in 
transportation management. ITS. as a sophisticated information 
system, however, requires accurate, high-density spatial data, 
which demand cannot be satisfied. with conventional data 
acquisition technologies. 
Remote sensing, including LiDAR can contribute to this task. 
Thé research presented in this paper has been conducted as part 
of the NCRST-F (National Consortium for Remote Sensing in 
Transportation - Flows) project. We will discuss some aspects 
of the potential use of laser scanning technology in 
transportation, especially, focusing on different methods of 
vehicle classification. 
input matrix contained the above mentioned parameters of all 
the involved vehicles. For the training, we used a data set 
acquired in Ohio, and for the tests, LiDAR data sets obtained 
from Michigan and Ontario. Further details of the Principal 
Component Analysis can be found in Toth et al. 2003a. As a 
result, the desired vehicle categories (e.g., passenger car, multi- 
purpose vehicles, and trucks) can be nicely separated and 
easily distinguished (Lovas 2004a). Figure 1 shows the vehicle 
categories in the 2D classification space. 
The result of this clustering served as a basis for automated 
pattern recognition methods. All three developed techniques 
(rule-based method, minimum distance method (Duda et al. 
2001), and neural network-based classifier, (Rojas 1993) are 
able to categorize the vehicles (72 training vehicles + 30 test 
vehicles) with accuracy better than 80 % (Toth et al. 2003b). 
Internation 
Hi died 
2.1 PCA 
Applying 
PCA-base 
intensity ‘ 
parameter 
the one w 
how the v 
  
Figur 
Supposed 
reflection 
very steep 
car theor 
intensity \ 
We storec 
vehicles 1 
previously 
and four | 
next appr 
columns « 
int4), see