In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009 
highest vote for each cluster. We propose to employ a more accu 
rate method such as the RANSAC method to model the building 
footprint. One advantage of our approach is the following. A 
curved facade will be approximated by a single line. In addition 
to this, much information deduced by the clustering step allow to 
automatically adjust the parameters of the RANSAC algorithm 
and to thus improve the precision of the detected lines. 
The RANSAC method is commonly used to detect lines among 
edge points (Bretar and Roux, 2005) and (Sester and Neidhart, 
2008). We use a classic method (Fischler and Bolles, 1981). For 
each detected cluster, we use the following process. We use the 
original space of parameter (x,y). Two different points belong 
ing to the cluster are randomly selected, characterizing a line. A 
neighborhood is defined along this line by a minimal distance be 
tween a point and the line. This process is iteratively repeated 
until the number of inlier points in the neighborhood is maxi 
mized. In our approach, the number of facade points is roughly 
equal to the number of points for each cluster. Furthermore, the 
minimal distance associated with the RANSAC technique can be 
determined from the dispersion of the cluster. When this step 
is carried out, the best fit lines of 2D points are extracted us 
ing the Least Squares Adjustment (LSA technique). A set of 2D 
segments giving the building footprint is then obtained from the 
detected 2D lines. 
4 PRELIMINARY RESULTS 
The acquired 3D data correspond to the facades of buildings in 
the 12th district of the city of Paris. In this study, we use a high 
precision 2D laser sensor LMS-Q120i made by RIEGL company 
1 . The laser sensor is positioned on the roof of the vehicle. Its 
beam plane is perpendicular to the vehicle trajectory. The system 
allows us to carry out 10000 measurements per second and the 
beam vertically sweeps with an opening of 80 degrees (-20 to 60 
degrees with respect to the horizontal). The angular precision of 
the beam is equal to 0.01 degrees. More specifically, the precision 
of laser-based measurements is approximately 3 cm at 150 m. In 
this study, the angular resolution was configured to 201 points by 
frame (see figure 7). The ground based laser range transmits laser 
pulses with simple echo. 
Figure 7: Acquisition of the 3D point cloud using the 2D laser 
sensor. The frame shows a selected band without occlusions. 
The raw measurements provided by the laser sensor are points 
that are parameterized by distance and angle. The reflected in 
tensity of the laser is between 0 and 1. The coordinates of the 
3D points are expressed in the laser sensor coordinate system and 
also in a common coordinate system, namely the ground refer 
ence (absolute) Northern, Eastern and Altitude in Lambert 93. 
The precision of a 3D point is not easy to evaluate because it 
depends on the laser precision and on the referencing system pre 
cision. 1 
1 Link to RIEGL company: http://www.riegl.com 
Figure 8: Two difficult facades for the classical Hough Trans 
form: i) a curved facade, and ii) a facade with a low density of 
3D points. The detected lines correspond to the local maxima of 
the Hough space accumulation using the filtered cloud. 
The experiments are carried out on two building facades having 
different architecture and different density of acquired 3D points. 
One can thus assess the robustness and the efficiency of the pro 
posed approach. 
Figure 9: Extracting the building footprint lines using the pro 
posed approach. 
Figures 8 and 9 show the extraction of the building footprint lines 
using the classical Hough Transform and our proposed approach, 
respectively. The 3D point cloud is presented in the upper part of 
the figure. The projection onto the 2D accumulator is presented 
in the lower part of the figure. The studied building illustrates two 
difficult cases for a classical Hough Transform. Indeed, the left 
facade does not suffer from occlusions but it is slightly curved. 
On the other hand, the right facade which is a planar structure is 
partially occluded, that is, the density of its 3D points in the 2D 
accumulator space is much lower than that of the left facade. 
Cluster characterizing one building facade 
Extraction based on the score maximum in the cluster 
Extraction based on a reestimation by the RANSAC method 
Figure 10: Comparative schema illustrating the precision of lines 
detection step on one simulated facade.