The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3h. Beijing 2008 
(a) One-orientation building (b) Two-orientation building 
Figure 6: Building main orientations; The yellow points repre 
sents the building outline and the red and the blue lines are the 
lines found by Hough transform. 
Figure 7: 3D prismatic model overlaid on digital terrain model 
creates the LOD1 representation 
5 A NOVEL APPROACH FOR BUILDING 
RECONSTRUCTION BASED ON PROJECTION 
BASED ANALYSIS OF 3D POINTS - LOD2 
The concept of our projection based building reconstruction ap 
proach is as follows. Geodesic image reconstruction with a very 
small height difference (cf. Figure 2) captures ridge points and 
roof outlines very reliably. This allows to deduce the main ori 
entation of buildings or building parts and a corresponding buffer 
zone (cf. Figure 10(a)). Next, a cuboid region which covers the 
building or building part is extracted. The spatial direction is used 
to define a 3D to 2D projection of the cuboid region. All 3D laser 
points included in the cuboid are projected onto a 2D projection 
plane, which is one of the planes of the cuboid. The projection of 
all laser points of the 3D volume results in point accumulations 
in the 2D projection. The cumulation of points corresponds to 
the main building shape in terms of a profile which represents the 
roof and typically the vertical walls. In our approach only a lim 
ited number of roof models is taken into account which are flat, 
hipped and gabled roofs. Figure 10 shows an example of a gable 
roof for a part of a building. Robust line fitting approximates the 
profile by straight line segments from which a polygon with the 
roof and the vertical walls is derived. This automatically elimi 
nates any details of the shape the building or building part. By 
extruding the extracted 2D information to 3D along the normal to 
the projection plane a 3D model of the building or building part 
is determined. The 3D model of the whole building is obtained 
by intersecting the models of its parts. The result is considered 
as the LOD2 representation. Refinement to greater detail follows 
the same conceptual idea. Instead of working with all data of 
the cuboid in one projection plane, a sequence of section planes 
is used which accumulate the respective part of the points of the 
cuboid. 
The proposed approach for generating 3D building models con 
sists of the following steps: 
5.1 Extract ridge line and determine main orientation 
It begins with image reconstruction by geodesic morphology to 
extract the pixels of the highest part of the building segment. A 
small height offset value, e.g., 0.2m is chosen for this purpose. 
As outcome all pixels that belong to the local peaks and their 
neighborhood are detected as shown in figure 8(b). For flat roofs 
the detected pixels represent the complete roof region. The region 
segments obtained by labeling connected components are classi 
fied into flat roof and ridge points using Gaussian curvature and 
surface normal as features for the classification. The number of 
extracted points in this step depends on the selected offset value 
and the inclination of the roof face. Some other regional max 
ima are also detected in this step (cf. Figure 8(b)). Next, straight 
(a) Range image (b) Roof top pixels; Difference between 
original image and reconstructed image is 
represented by red points 
Figure 8: Determination of roof top 
line segments are extracted with RANSAC from the ridge points. 
The orientation of the ridge line segments are calculated and ver 
ified by the orientation of the boundary lines (section 4). Since 
in most cases the ridge lines are parallel or perpendicular to the 
building edges, the orientation of the ridge is compared with the 
main orientation of the building. If the deviation angle (£) be 
tween the ridge line and the main orientation is less than, e.g. 
±5°, the ridge line is rotated around its center of gravity with 
the value of £. The orientation for building parts with flat roofs 
is calculated based on the minimum bounding rectangle for the 
roof outline. Figure 9 shows the points classified as ridge points 
and the RANSAC lines superimposed on the original LIDAR im 
age. Ridge points shown in blue in this figure are outliers of the 
RANSAC process or lines which are not approved because not 
enough inliers are found. 
(a) Ridge points and ridge (b) Ridge points and ridge lines (detail) 
lines (overall view) 
Figure 9: Classified points as ridge points (blue points) and lines 
fitted by RANSAC (red lines) superimposed on LIDAR image 
5.2 Localization of the building parts 
For a rectangle parallel (or perpendicular) to the main orienta 
tion the points located inside it are extracted using the point-in