rent data sets, 
f the buildings: 
nd. The point 
Square meter. 
id simple roof 
e or combined 
igs in this data 
] possess large 
iccount for the 
uildings of this 
tiveness of the 
'g which is on 
$ test area is 
luding various 
it roof shapes. 
e meter. 
cm in position 
aw irregularly 
'mented point 
uilding, some 
s of adjacent 
s results from 
y of Dresden 
ains rasterised 
t serves as a 
a obtained by 
aces different 
lding model 
‘hich arise in 
. necessary to 
ing. For this 
e, which roof 
characteristics 
ces. Thus, the 
aight lines in 
hich should be 
le roofs, pent 
of these basic 
ess maximally 
h other. The 
onal to one of 
e building are 
the orientation 
nore complex 
f shapes, it is 
1g can possess 
re parallel or 
that the lower 
dge is parallel 
plane vertical 
sideration of a 
. Istanbul 2004 
certain roof projection. Opposite roof faces, which intersect in 
a ridge line, do not necessarily have the same inclination. 
3.2 Overview 
The basis for the developed algorithm is a segmented laser 
point cloud, which represents a potential building in each case. 
The individual point clouds are processed one after the other by 
means of the following algorithm: 
a) Read laser scanner points and reduce coordinates to 
barycentric coordinates 
) Elimination of alleged bottom points 
c) Determination of the azimuth of the ridge direction and 
rotation of the data points by the azimuth around the z-axis 
d) Projection of the laser points in the z-x and z-y-plane 
e) Search for lines in these projections and determination of 
the extension of the roof faces, which are represented by 
the lines 
f) Determination of the roof face outlines 
g) Blending the roof faces, which were obtained from the 
different projections 
h) Determination of the walls 
i) Determination of the ground plan polygons and 
visualisation of the building as VRML model 
3.3 Determination of the ridge direction 
As the method is based on the principle of line detection in 
projections of the point cloud orthogonal to the direction of the 
ridge of the roof, the first step of the modelling procedure is the 
determination of the main directions of the buildings. Potential 
ground points are eliminated by analysing a height-bin 
histogram. The minimum of laser points in the height layers 
within the range of the walls can be used as a criterion for the 
separation of potential roof points from ground points. 
With the remaining points that are classified as roof points, the 
search for the ridge direction of the building takes place. The 
main ridge direction is then given by the azimuth x (see 
Figure 3-1 a). 
The principle used is based on the investigation of the 
orientation of the points in individual height layers of the point 
cloud. Not only the upper height interval containing ridge 
points shows the orientation of the building, but also the lower 
height layers of the roof contain this information. The idea is to 
search for lines within the points of each height layer. Within 
the range of the roof the dominant direction of the detected 
lines corresponds to one of the two main directions of the 
building. In contrast, the distribution of points in height layers 
of vegetation has a random character. 
The most pronounced direction of the detected lines is the one 
that is accepted as the main roof direction. The point cloud is 
now rotated by the angle k around the z-axis, so that the main 
direction (the main roof ridge) of the building runs parallel to 
the y-axis (see Figure 3-1 b). 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
a) 
  
Figure 3-1: a) roof ridge direction; b) rotated point cloud 
3.4 Detection of roof faces in projections 
In the next step the points are projected onto vertical planes, 
defined by the detected azimuth direction. First the data is 
projected onto the z-x-plane. Points, which lie on a roof plane 
with a normal vector parallel to the projection plane, are 
displayed as a line in the projection plane. 
In the projection, lines that represent roof planes are intersected 
and their end points are determined. In dense datasets it may be 
necessary to thin out the points on the line to warrant a proper 
performance of the line detection procedure. On this basis, 
knowing the start and the end points of the lines, the inclination 
and the width of the roof areas represented by the lines are 
given (see Figure 3-2). 
  
  
Z-X-projection 
  
  
  
Figure 3-2: Inclination and width of detected roof faces