In: Paparoditis N., Pierrot-Deseilligny M„ Mallet C., Tournaire O. (Eds). IAPRS. Vol. XXXVIII, Part 3A - Saint-Mandé, France, September 1-3, 2010 
adapted from (Vosselman and Zhou, 2009), described at the end 
of section 2.2. The ground points are classified as high or low, and 
pairs of corresponding high and low points are selected. Median 
points are stored as curbstone point. Figure 2 illustrates its prin 
ciple. This method was designed for aerial lidar data but works 
well on terrestrial data. Furthermore, the terrestrial data give di 
rect information about the altitude of the road associated to the 
pavement, thanks to the vehicle altitude. By subtracting the road 
altitude to the laser point altitude, it is possible to compute the 
relative pavement altitude. Thus, the determination of the thresh 
old delimiting high and low points is not influenced by the road 
slope. At the end, the curbstone points are stored as a set of con 
nected points. 
• the segment is close to the pavement edge (typically less 
than 10 meters); 
• the angle between the road segment and the main direction 
of the pavement edge is small (typically less than 40°). 
In the road model, one road can be represented by several seg 
ments, even if no crossroad occurs. Thus, a single pavement edge 
can correspond to several segments. At the previous step, an edge 
is associated to the nearest segment. It is now necessary to split 
pavement edges according to the road segments, and to assign 
the curbstone points to the right segments. Figure 3 illustrates the 
result expected at this step. 
(a) 
(b) 
Figure 2: Principle of the curbstone determination method. The 
real pavement edge is in green (straight line), enlighten pixels of 
the connected component are in white, high points are in light 
orange and low points are in dark blue. Selected pairs and their 
median points (detected curbstone points) are in black. 
3.3 Pavement edge ordering and connection 
As the curbstone points can be irregularly distributed within a 
connected component, they need to be ordered and connected 
within each component. Each component is processed indepen 
dently. The distances between all curbstone points within a con 
nected component are computed and the two farthest points are 
stored as pavement edge extremities. A path between these ex 
tremities is then found by iteratively selecting the nearest point 
that has not already been stored as a path point. 
As laser points are acquired with a mobile vehicle, the relative 
acquisition times of the points give the ordered pavement edge 
sequence along the road. Then different pavement edges are con 
nected if they are near enough. The distance is chosen to connect 
pavement edges separated by the shadow of a car. The direction 
of each component is given by the chronological order. 
3.4 Road surface delineation 
The aim of the road surface modelling stage is first to associate 
each pavement edge to the available road axes. The road axes 
are derived from aerial imagery, simultaneously to 3D building 
models. The road axes are registered to the laser point cloud via 
the registration of the 3D building models, using the method de 
scribed in (Denis and Baillard, 2010). 
The existing road model is composed of a series of 3D polylines 
describing road axes. Each polyline is made of "road segments" 
connecting two successive points. Each segment describes a road 
portion. A road intersection always implies a node in the model, 
and consequently a segment end point. However, a road portion 
without intersection can be described by several segments. The 
3D road model brings more information than the vehicle track 
recorded during data acquisition. The street intersection positions 
are not provided by the vehicle track unless all the streets are 
covered. 
Each detected pavement edge is associated to the nearest road 
segment, if the following criteria are respected: 
Figure 3: Scheme illustrating the road segment / pavement edge 
association, (a) before edge split and merge, (b) after edge split 
and merge. 1. initial segment edge, 2. extended edge, 3. edge 
from neighbouring segment. 
For this purpose, the points of each pavement edge are projected 
onto the corresponding road segment axis. Only points project 
ing on the road segment are kept. Points projecting out of the 
segment are reassigned to another road segment, according to the 
following rules: 
• the segment is connected to the current one; 
• the angle between the road segment and the main direction 
of the pavement edge is small; 
• if several segments respect the two previous conditions, the 
nearest is selected. 
If not any road segment can be chosen, the edge points are lost. If 
several pavement edges are associated to the same road segment, 
they are extended and merged. As no road intersection can occur 
within a road segment, the pavement edges are assumed to be 
continuous along a road segment. 
Finally, when a road segment is associated to an incomplete pave 
ment edge, this latter is extended from its last point assuming a 
constant road width. 
At the end of the process, the pavement edges derived from the 
right and the left laser point clouds are associated to the corre 
sponding road segments of the existing model. Then, the road 
surface associated to each road segment can easily be delineated. 
3.5 Average road width 
The final purpose of this work is to provide a complex road model 
with geometrical information. In particular, the average width is 
an information that must be computed and recorded for all the 
road segments. 
The road surface delineation allows to compute the average width 
of each road segment. The right and the left sides of the segments 
are processed independently.