In: Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
GEOREFERENCED
' ■
USER RAW DATA
PREPROCESSING
3D CLOUD POINTS
INTO 2D ACCUMULATOR
EXTRACTION OF FEATURES
MODELING
BUILDING FOOTPRINT
Figure 4: Data flow diagram of the building footprint extraction
approach.
the dataset by thresholding these distances. Moreover, this tem
poral effect distinctly appear on the image of intensity of the laser
beam and looks like a succession of rows having the same inten
sity (see figure 3). Therefore detecting redundant frames can also
be based on differences between the intensity of return of two
consecutive rows. This step solves only the problem of data re
dundancy related to the sensor immobility.
3.1.2 Volume of interest The sensor characteristics are used
to select the 3D points belonging to the facade plane. The 3D
available points are principally positioned above the vehicle to
reduce the problem of occlusions. A horizontal band is defined
between two horizontal planes. The lower plane passes through
the sensor center. The upper plane is shifted by a certain dis
tance that is related to the height of the buildings under study. We
precise that the ground altitude could be simply deduced by mea
suring manually the laser sensor height. Finally, the volume of
interest is defined by the georeferenced trajectory of the vehicle
and the above horizontal band. The 3D points not included in this
volume will be removed from the dataset.
having a high score are kept. The process uses a global threshold
which is compared to the maximum score. By this technique, the
erratic points of the cloud are removed from the data. The cells
with high scores are principally facade points with high density.
Several techniques for the detection of outliers in laser point clouds
can be found in (Sotoodeh, 2006) and (Sotoodeh, 2007).
3.2
Building footprints extraction
3D cloud points
Projection 3D to 2D
T
Hough counting space
/r-nieans clustenng
e
Lines extraction
0
Figure 5: The main steps of our proposed approach for building
footprint extraction.
3.1.3 Exploiting the linearity of 3D data After the preced
ing filtering steps, the frames have undergone a horizontal crop
ping. The data structure represented by frames is now repre
sented by a sequence of 3D points. We exploit the fact that in
this representation facade points are locally aligned. We seek fa
cade points which are principally organized vertically. Thus, the
dataset in this sequence is parsed by triplets. The central point
of each triplet is kept in the dataset if the triplet is aligned, other
wise it will be removed from the dataset. Therefore, the coplanar
points of the building facade are kept. Besides, we observe that
the 3D points belonging to other linear structures are also kept.
3.1.4 Mapping the 3D point cloud onto a 2D accumulator
The goal of this step is twofold. Firstly, it aims at removing noisy
and outlier points. Secondly, it gives a very compact representa
tion of the filtered 3D points. Since we are interested in the verti
cal structures that generally represent the facades, we project the
3D cloud on a horizontal plane. More precisely, the 3D points are
projected into a 2D grid to create an accumulation space. Each
point of the cloud votes in one cell, giving a score. Only cells
The goal is to automatically extract the building footprint using
the 3D filtered points cloud contained in the compact 2D accumu
lator. The building footprint is a set of 2D segments that can be
detected in this 2D space. Recall that the vertical structure of the
facades is captured by the scores of the cells. Each cell contains,
if any, a set of 3D points P(x,y,z). Furthermore, an efficient
extraction can be obtained by working with the barycenters (2D
coordinates) of the cells together with their scores. Our approach
combines the use of the counting space of Hough Transform, the
fc-means clustering technique and the RANSAC method. We
briefly describe these three techniques and their properties ap
plied in our context. Figure 5 illustrates the main extraction steps.
The Standard Hough Transform (SHT) allows the extraction of
the 2D lines among 2D dataset points (Hough, 1962). In the
field of our application, this method is currently used to detect
the building boundaries in aerial images using the edge points.
This method is also used to extract buildings in LIDAR data (e.g.,
(Tarsha Kurdi et al., 2007) and (Karsli and Kahya, 2008)).