The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008 
487 
There are some advantages on filtering of the laser images using 
geodesic image reconstruction: 
• The process is not sensitive to the size of the objects to be 
filtered. Spacious as well as small buildings can be filtered 
using this approach. 
• Contrary to standard morphological processing, for which 
proper structuring elements have to be defined this is not the 
case in this process. In geodesic dilation the marker image 
is dilated by an elementary isotropic structuring element. 
• Another benefit is that, the geodesic image reconstruction 
does not effect ground pixels. Therefore the normalized 
DSM can be simply segmented using a threshold value of 
zero. 
• The filtering approach based on geodesic dilation is rela 
tively fast. In many cases even in hilly regions the filtering 
can be implemented with a single marker image. A marker 
image which represents the minimum height value of the 
mask image except pixels at the boundary when marker = 
mask (Arefi et al., 2007b) can be used. 
(b) DTM generation result (LODO) plus contour lines superimposed 
on it 
Figure 3: Generation of digital terrain model by hierarchical fil 
tering of non-ground objects 
4 BUILDING OUTLINE DETECTION AND 
APPROXIMATION FOR GENERATING 3D 
PRISMATIC MODEL - LODI 
The normalized DSM shown in figure 2(d) contains buildings 
as well as vegetation pixels and other 3D objects might be also 
present in the data. Classification of the regions is carried out 
rule based using geometric and other region properties. Size of 
the regions, vegetation index based on first and last pulse range 
data and variance of the surface normals have been employed in 
rule based classification to separates building and vegetation re 
gions. To model the second level of detail the extracted build 
ing outline is simplified to a polygon which includes only few 
significant points such as corners. For this purpose two meth 
ods are employed: fitting a rectilinear polygon by iterative fitting 
of minimum bounding rectangles (MBR) and straight line fitting 
and merging based on RANSAC (Arefi et ah, 2007a). The first 
method is simple and relatively fast to find the best rectilinear 
polygon but is only applicable on rectangular polygons. First 
1st approximation 1st approximation - Model of Surplus 
Original (Superset) Original Regions 
1st approximation - model 
of surplus regions 
2nd Approximation Original - 2nd 
(Subset) Approximation 
Figure 4: Iterative process of MBR for building outline approxi 
mation 
Figure 5: Building approximation result 
the main orientations of the building edges are determined using 
a Hough transform. The iterative process of applying MBR’s is 
shown in figure 4. The process stops if the remaining unmodeled 
details are neglectable. A result of such a MBR approximation 
is shown in figure . If the analysis in Hough space indicates that 
there is more than one main orientation (cf. .6) the second tech 
nique is used. The example shown in figure illustrates that the left 
building has a single main orientation represented by the red lines 
and the right building has two main orientations represented by 
red and blue lines. Accordingly, outline polygons are extracted 
and approximated with MBR or the RANSAC method. 
To generate the 3D model from 2D polygons the 2 component 
of the polygon nodes is extracted from the DTM and averaged. 
A representative height of the building is found by averaging the 
heights of the LIDAR points inside the boundary polygon. Next, 
the polygons relating to the walls and floor of each building are 
formed. All 3D polygons are overlaid on DTM to create LODI 
representation.