In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part 3A - Saint-Mandé, France. Septeniber 1-3, 2010 
282 
Extrinsic parameters (camera-centered) 
Figure 1: overview of the truck 
(and more specifically the GPS antenna position) as the center of 
the truck coordinates system, every data have to be replaced in 
this specific frame, and will be finally expressed in the world co 
ordinates system. We now present briefly the calibration methods 
used for replacing all devices in this coordinates system. 
Figure 3: calibration grid positions 
In order to compute Lidar positions and orientations in the GPS 
coordinates system, we decided to determine their positions with 
respect to the camera, then to transfer these positions using the 
camera extrinsic calibration results. 
Different approaches for lidar calibration have been developed. 
Antone and Friedman implemented a method where only lidar 
range data are required, but which is based on the design of a spe 
cific calibration object (Antone and Friedman, 2007). They claim 
that registration to any camera can be further processed by apply 
ing a pattern on this object. (Mahlisch et al.. 2006) developed a 
calibration method of a multi-beam lidar sensor with respect to a 
camera which has to be sensitive to the spectral emission band. 
Alignment is then performed when viewing a wall from differ 
ent orientations, through a reprojection distance minimization. 
Huang presented an algorithm for multi-plane lidar calibration 
using geometric constraints on the calibration grid plane (Huang 
and Barth, 2008). We chose Zhang and Pless approach (Zhang 
and Pless, 2004), a two step algorithm based on a geometric con 
straint relative to the normal of the calibration grid plane. This 
method uses a linear determination of the pose parameters, fur 
ther refined by a non linear optimization (generally performed 
through a Levenberg-Marquardt algorithm). 
In our experiments, we use about 15 images, where the calibra 
tion grid is seen in both lidar and camera views. The poses of the 
calibration grid with respect to the camera are previously deter 
mined in the 2.2.2 section and we select manually the grid area in 
the lidar scan (Cf. figure 4). Zhang and Pless two-pass algorithm 
is then performed with the collected data for the three front lidar 
range sensors. 
(a) Lidar and camera 
2.2 Camera Calibration 
(b) “Front” and “Sky” lidar sensors 
cones of view 
2.3 Lidar Calibration 
J 
♦ 
1 
—4— 
♦ w 
’a) Calibration grid is 
selected in the scan 
manually 
(b) corresponding image 
We present in this section a two step camera calibration ; first 
the camera is roughly oriented so as to be aligned with the truck 
main axis. Then it is finely calibrated, using a dedicated Matlab 
toolbox. 
2.2.1 Camera Rough Alignment We designed a calibration 
site presenting many parallel longitudinal and transversal lines 
and marks for positioning the vehicle wheels. A primary camera 
orientation is processed in order to align it with the truck main 
axis : it consists in a dynamic process allowing a rough setting of 
the pitch, the roll and the yaw. Pitch is set in a way that the vehi 
cle’s hood is not seen. Roll is set to zero from the transversal lines 
mean orientation. Yaw is set so that the longitudinal lines vanish 
ing point has an u-coordinate equals to the principal point - i.e. 
the projection of the camera center in the image - u-coordinate 
(Cf. figure 2). 
Figure 2: configuration tool output image, instructions are dis 
played at the top left of the image 
2.2.2 Fine Calibration We mainly focus here on extrinsic ca 
libration, i.e. the camera position and orientation with respect 
to the truck. We used Jean-Yves Bouguet's Matlab camera ca 
libration toolbox ’, which relies on Zhang calibration algorithm 
(Zhang, 2000) and returns calibration grid positions in the camera 
coordinates system (Cf. figure 3). 
Intrinsic parameters, though quite important for any image pro 
cessing algorithm, are not critical in the road description process. 
Indeed, this first stage goal is to define camera position and ori 
entation in the GPS antenna coordinates system, which is done 
using our calibration site. 
1 http ://www.vision.caltech.edu/bouguetj/calib_doc/ 
index.html 
Figure 4: lidar data and corresponding image used for calibration 
As an intermediate result of this stage, we can reproject lidar im 
pacts on the grid, as can be seen in figure 5. 
The final step in the calibration scheme consists in replacing all 
calibrated lidar in the GPS antenna coordinates system, in order 
to have all sensors in the same reference. As these sensors are 
2D lidar, the vehicle displacement provides a full scan of the sur 
rounding 3D world. Using RT Maps acquisition platform, all data