In: Paparoditis N., Pierrot-Deseilligny M„ Mallet C., Tournaire O. (Eds), IAPRS, Vol. XXXVIII, Part 3A - Saint-Mandé, France, September 1-3, 2010 
et al., 2008). They detect pavement edges by analysing individual 
scan lines, assuming that the road surface is flat and lower than 
the pavement surface. Aufrére et al. have detected curbstones 
with the help of a dedicated vehicle (Aufrere et al., 2003). A 
laser scanner and a camera are located on the front bumper of the 
vehicle and oriented towards the pavement. For each scan line, 
the range histogram is thresholded to find the main class. This 
class represents the curbstone points. This laser-based detection 
is refined by a camera-based detection. The curbstone positions 
detected in the laser point cloud are projected on the camera im 
age. A region of interest (ROI) is defined. Within the ROI, a 
Canny edge detection is performed. The extracted new edge al 
lows to define a new ROI to repeat this operation and extend the 
detected pavement edge as much as possible. An occlusion detec 
tion module, based on laser points, increases robustness in case 
of occlusions. 
Jaakkola et al. have worked on road surface modelling from ter 
restrial laser point clouds (Jaakkola et al., 2008). The curbstones 
are detected studying gradients in a height image. Hernández and 
Marcotegui have also worked with a terrestrial mobile mapping 
vehicle and have detected curbstones using laser point clouds 
(Hernández and Marcotegui, 2009). The clouds are first subdi 
vided into building blocks, and artifacts (cars, pedestrians, etc.) 
are removed. Then the ground points are detected using a cloud 
segmentation. As some ground points are removed by the arti 
facts suppression module, a recovery step is performed. All the 
points belonging to the estimated ground plane are recovered. An 
elevation image is finally derived from the new cloud of ground 
points, then segmented into planar regions with a watershed al 
gorithm. The boundaries between planar regions finally give the 
pavement edges. 
Vosselman and Zou have worked on curbstone detection from 
aerial lidar (Vosselman and Zhou, 2009). Points located on or 
near the road are first selected, then a threshold is applied on their 
elevation to classify them as "high points" or "low points". Each 
high point is associated to the nearest low point and each low 
point is associated to the nearest high point. Pairs are selected 
only if they belong to both computed sets of pairs. For each se 
lected pair, the median point is computed and stored as a curb 
stone point. The curbstone points are linked into a single com 
ponent when they are near enough. Each component is finally 
vectorized using a line detection technique based on RANSAC. 
3 METHOD DESCRIPTION 
We propose a method to delineate the road surface associated to 
a given road axis. It relies on a combination of existing methods 
that were adapted to be robust to the road slope and to occlusions 
caused by parked cars. It also takes advantage of the existing road 
axes. The whole road surface is delineated and the road width is 
computed. The approach is composed of four stages, presented 
in this section: 
3.1 Ground point detection 
The curbstone detection method starts with the extraction of ground 
points from the laser point cloud. The right and the left clouds 
are processed independently. The elevation histogram of each 
vertical scan line is computed. Most ground points belong to the 
lowest significant peak of the elevation histogram. These points 
are therefore recorded as seed points then used within an iterative 
clustering algorithm applied to the whole cloud. Clustering tech 
niques are commonly used for ground detection from airborne 
lidar data Digital Elevation Models (Vosselman et al., 2004, To- 
vari and Pfeifer, 2008, Badea and Jacobsen, 2008, Chehata et al., 
2008). Our clustering algorithm works as follows. The ground 
seeds detected at the previous step are iteratively and chronolog 
ically stored into a small fixed size queue (typically 500 points), 
the acquisition order corresponding to the progression along the 
street. 
At each iteration, a mean square plane is computed over the stored 
points. All the points of the cloud belonging to this plane are 
marked as ground points. The last acquired ground points are 
used for updating the queue. Thus, the seed location moves along 
the street at each iteration and follows the ground curvature. A 
small queue size will be enable high ground curvature, at the risk 
of including low objects. 
This process is first performed chronologically (seeds are stored 
following the point acquisition times) then in the inverse direc 
tion. 
A drawback of this method lies in the fact that it keeps points 
belonging to the bottom of objects lying on the floor like car tires, 
pedestrian feet or façade bottoms. 
3.2 Curbstone detection 
The pavement edge detection is based on the analysis of the el 
evation gradient computed over the ground point cloud. Thus, it 
is guaranteed to detect only differences in elevation occurring on 
the ground level. An image of accumulation is first computed, 
whose pixel value is the minimal elevation of the corresponding 
ground points. In order to remove noise, the elevation image is 
smoothed with a Gaussian filter. Then gradients are computed 
with a Sobel filter and edges are determined by a hysteresis filter. 
Thresholds are chosen assuming that pavement height is about 10 
cm. 
To remove false detections due to object feet as tires, pedestrians 
or façades, the gradient is also computed over the elevation image 
of the whole point cloud. The pixels selected at the previous step 
are studied in the new gradient image. As illustrated in Figure 1, 
if the difference between the new gradient and the previous one is 
larger than a predetermined threshold, then the pixel is rejected, 
because the height variation is too important to correspond to a 
pavement edge. The threshold value is chosen equal to 20 cm. 
1. ground point detection; 
2. curbstone detection; 
3. pavement edge ordering and connection; 
ground gradient 
Figure 1 : Illustration of the method for removing false detections. 
4. road surface modelling and width computation. 
These stages are independently applied on the right and the left 
laser point clouds, using the same road axes. 
The resulting pixels give a coarse position of the curbstones. They 
are labelled as connected components representing sections of 
pavement edges. Each component is then processed individually. 
The accurate curbstone positions are computed with a method