In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3. 2010 
293 
TOWARDS ROAD MODELLING FROM TERRESTRIAL LASER POINTS 
Eloïse Denis, Renaud Burck and Caroline Baillard 
Si rade1 
3, allée Adolphe Bobierre 
CS 14343. 35043 Rennes, France 
edenis@siradel.com, renaud.burck@ensg.eu, cbaillard@siradel.com 
http://www.siradel.com 
Commission III, WG III/2 and WG II1/4 
KEY WORDS: Urban, Terrestrial, Laser scanning, Point Cloud, Modelling, City. Feature, Measurement 
ABSTRACT: 
This paper presents a complete pipeline for extracting and modeling urban roads as 3D surfaces, using laser points acquired from a 
terrestrial mobile mapping system (MMS) and existing 3D road axes derived from aerial imagery. First ground points are extracted from 
the laser cloud with an adapted surface growing method. Then the pavement edges are detected by analysing the elevation gradient, 
then ordered and connected. After registration to the laser data, the existing road axes are associated to the detected pavement edge 
components. Finally, the road surface and the road width are estimated for each road segment. Our method is illustrated on test data 
sets describing narrow historical streets. It shows good results, despite the significant slope of the street and the numerous parked cars 
occluding the curbstones. The road surface is delineated all along the street, and the estimated width is very close to ground truth. 
1 INTRODUCTION 
City modelling is an expanding subject of study, stimulated by 
the advent of virtual worlds such as the ones provided by Google 
Earth or Visual Earth. Modelling an urban scene consists in de 
scribing not only buildings (location, textures, roofs, façade de 
tails. etc.) and vegetation, but also ground objects such as roads, 
parking lots, pavement, etc. Thus road modelling has become in 
creasingly complex and generates more and more interest within 
the scientific community. 
This paper presents a new approach for modelling urban roads 
as 3D surfaces, using vehicle-based laser points and existing road 
axes. More precisely, the aim of this study is to locate and connect 
pavement edges, to delineate the road surface and to compute the 
road width. This information will be used as an input to a more 
complex road modelling system. 
The laser point clouds used for this work have been acquired 
with a mobile mapping vehicle equipped with two lateral laser 
scanners. Each one is mounted on a side of the vehicle and pro 
duces vertical scan lines, describing any object located between 
the ground near the vehicle and the top of the lateral buildings. 
For each data set, right and left laser point clouds are provided. 
The mobile mapping vehicle is equipped with a very precise geo- 
referencing system composed of a Global Positionning System 
(GPS), an Inertial Measurement Unit (IMU) and an odometer, 
that provides precisely geo-referenced laser point clouds. More 
over. as a data provider, Siradel has a huge amount of 3D geo 
graphical data bases at an accuracy better than 1 meter, derived 
from aerial imagery. These data include a geometric description 
of the road axes, which will be used as input data of our road 
modelling process. 
Section 2 presents some previous work and section 3 describes 
the principles of our method. Our algorithm was tested on ex 
perimental data sets and the results are presented in section 4. 
Section 5 finally concludes this paper and presents some future 
evolutions. 
2 RELATED WORK 
2.1 Ground point detection from terrestrial laser points 
Rao et al. have presented two methods to classify ground and 
non-ground points in a vehicle-based laser point cloud (Rao et 
al., 2006). They are both based on point projection onto a hori 
zontal plane. The first method consists in generating an accumu 
lation image of the point cloud. Each cell stores the data related to 
the projected points. The cells containing few points, with small 
height deviation, are considered as ground cells, and the corre 
sponding points are classified as ground points. This method can 
unfortunately also detect points belonging to roof slopes or trees. 
However, as these points float in mid-air, they can easily be de 
tected and removed. The second method takes advantage of the 
local spatial characteristics of the point cloud. Within the projec 
tion plane, points are grouped into large cells and the algorithm 
iteratively finds regions with low height variations that are spa 
tially correlated. With a modified minimum filter, laser points are 
classified into ground and non-ground points. This method ac 
curately extracts points belonging to the ground but also points 
located at the foot of objects (bottom of tires, trees or walls). 
Goulette et al. have developed a real-time algorithm to segment 
vehicle-based laser points into four classes: ground, façades, trees 
and other (Goulette et al., 2007. Abuhadrous, 2005). Their method 
is based on prior knowledge on the different classes: the ground 
is a horizontal plane with a high density of points, the façades 
are vertical planes, and the trees are free shapes whose projec 
tions onto the ground have a specific width. The histograms of 
each scan line are studied to get a first segmentation of the points. 
Then fuzzy logic helps to refine the segmentation. 
2.2 Pavement edge detection and vectorization 
Most work about pavement detection is either based on edge de 
tection in an image of point accumulation or based on histogram 
analysis. 
Shi et al. have worked on automatic extraction of road boundaries 
and road marks from terrestrial images and laser range data (Shi