The paper is organized as follows: Section 2 gives an overview on 
the proposed algorithm for generating 3D models from LIDAR 
point clouds in three LOD. DTM generation from LIDAR data is 
explained in section 3. Section 4 explains building outlines are 
detected and approximated to produce second LOD. In section 
5, the idea of modeling based on projecting 3D data into a 2D 
plane and generating the LOD2 model is explained and finally, in 
section 6, the achieved quality of the reconstructing buildings is 
discussed. 
2 PROPOSED ALGORITHM FOR 3D BUILDING 
MODEL GENERATION IN THREE LEVELS OF 
DETAIL 
Figure 1 presents the proposed work flow for automatic genera 
tion of building models. The process begins with separating non 
ground from the ground regions by hierarchical filtering using 
geodesic reconstruction. A DTM is produced by interpolating 
the gaps obtained by the filtering process. The result represents 
the first LOD, i.e. LODO. The approach continuous with extract 
ing building regions from the ALS range data. A segmentation 
and classification algorithm groups and classifies the laser range 
pixels into building, vegetation and other classes. Next, the build 
ing outlines are detected and approximated to reduce the number 
of boundary pixels to some significant nodes. After estimating 
an average height for the building, a prismatic building model is 
generated to form the second LOD, i.e. LODI. Projection based 
analysis of the LIDAR data is proposed for 3D building recon 
struction to form the LOD2. The algorithm uses geodesic mor 
phology for line detection and a 2D model driven technique for 
building reconstruction. 
3 AUTOMATIC DTM GENERATION - LODO 
A hierarchical approach for filtering of the non-ground regions in 
LIDAR data and generating digital terrain models has presented 
in (Arefi and Hahn, 2005, Arefi et al., 2007b). Image recon 
struction based on geodesic dilation is the core of this algorithm 
which is proposed by (Vincent, 1993). The image reconstruc 
tion is achieved by applying geodesic dilations until stability is 
reached (Jahne et al., 1999). The idea of image reconstruction is 
shown in figure 2. A profile of some non-ground objects located 
on an uneven surface is shown in figure 2(a). Laser points (red 
dots) are overlaid to the profile. The only input to generate im 
age reconstruction is the height difference h shown in figure 2(b). 
The result of geodesic image reconstruction is displayed in figure 
2(c). The reconstructed image is subtracted from the original im- 
(a) Profile representing the ground (black color) and the lo 
cation of the laser points (red dots) 
Mask 
(b) mask and marker; marker = mask - h except pixels 
at the boundary of the image when marker —mask 
LIDAR DSM 
Hierarchical filtering 
of non-ground regions 
DTM generation (LODO) 
Building extraction 
building outline 
Determine main orientation 
of building part 
Localize building parts 
T 
Project 3D into 2D plane, 
(d) nDSM of (c) 
Figure 2: Geodesic image reconstruction by selecting a marker 
image by subtracting h as offset value from the mask image. 
age to shape the normalized DSM 2(d). In this initial nDSM or 
nDSMO small hills may incorrectly be included. To avoid this 
problem the initial nDSM of non-ground regions is evaluated by 
a feature criterion that highlights jumps. The surface normal or 
the local range variation (LRV) defined by the difference between 
dilated and eroded LIDAR image in 3 x 3 local windows can be 
used as features. Thresholding and connected components analy 
sis leads to potential non-ground regions. The boundary pixels of 
these regions are evaluated by these features and the regions with 
height jumps are classified as non-ground. A sequence of marker 
images, provided by different offsets h, are used hierarchically to 
detect high as well as low non-ground objects. After separating 
the ground and non-ground pixels in the LIDAR data the gaps are 
filled using interpolation to get the DTM (LODO).