International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
both x and v directions, (c) Roof classes results 
without gap filling, (d) Roof classes results after gap 
filling and refinement. 
x» & 3 A 
E " - 
H. xd 
E : A 
3 i 
4 ) 
1| 3 x i sd i 
] Bn m a. m 7 x = 
N | I 
ias + | ' 
   
   
x RA 
(a) Aerial image 
  
       
  
(c) Slope in y 
^. Unclassified Cells (pixels) 
o 
^ 
i» 
Heigh: intercept 
  
Height intercept 
  
. tas “0 s ; 0 
45 t9 z ;,9lepeinx s 05 — - as Slope in x 
Slope in y 0s : Slopeiny © 5 
(e) Preliminary Clustering result (f) Refined clustering result 
UU Um 
(e) Raw segmentation results (f) Refined segmentation result 
Figure 4: Roof facets clustering in 3D search space. 
4. EXTRACTION OF PLANE-ROOF POLYGONS 
In this section, the procedure of translating the irregular roof 
facet regions to typical vectorized polygons is discussed. This 
procedure contains many steps to get the desired 3D polygons 
of the roof. Extracted roof regions will be transferred to 2D 
polygons first through raster to vector conversion that includes 
line extraction, connecting, joining, trimming, and segment 
adjacency determination. The geometrical plane-roof 
parameters, inclination and height, are then estimated based on 
the irregular LIDAR points inside each polygon. This enables 
the configuration of 3D roof facet polygons. 
4.1 Extraction of plane-roof regions outlines 
As a result of the region growing segmentation, roof facet 
regions were segmented and labeled as shown in the above 
section. 
4.1.1 Simple roof structure 
Simple roof structures mean here that the breakline between 
roof segments is uncomplicated and is parallel to one of the two 
dominant directions of the building footprint. In such buildings, 
the polygon extraction algorithm that was discussed in 
(Alharthy and Bethel, 2002) is applied to obtain roof segments 
outlines. The only constraint to this algorithm is that it can only 
extract lines in the two dominant directions of the building. 
However, the algorithm was very useful since all intermediate 
steps such as line extraction, connecting, trimming, and 
polygon formation, are embeded in it. And its main advantage 
and strength is the ability to preserve the squaring property of 
the extracted polygons. In general, the performance of this 
algorithm was excellent. Results of this step which show the 
extracted polygons (black lines) overlaid on the segmented roof 
regions are shown in figure 5. 
  
50 
  
  
  
Yinm 
3 
Yinm 
  
  
  
  
  
  
19 20 30 40 50 60 zu 30 40 50 60 7% 8 9 100 110 120 
Xinm Xinm 
Figure 5: Extracted roof segments polygons 
4.1.2 Complex roof structure 
In complex roofs, breaklines between roof segments are not 
limited to be parallel to the dominant directions of a building, 
instead they might take any direction and roof segments might 
be in any shape. Based on that, the previous algorithm of 
polygon extraction would not work here. So, a modified 
prismatic algorithm was used to refine the segmentation results. 
A data driven model was used to connect and generalize these 
roof planar surfaces in order to extract standard roof polygons. 
The approach is a modified version of an approach that was 
presented in a report in 1995 by U. Weidner, Institute of 
Photogrammetry, Bonn University, Germany. The approach 
treats each region segment individually. It starts with the 
boundary points by sorting them in clockwise mode starting 
from the upper left point as shown in figure 6(a). In addition to 
its position (x,y), each region boundary point will be given two 
labels, the first one tells to which roof segment this point 
belongs, and the second label tells its order among the boundary 
points of the segment. Now points will be considered as the 
polygon vertices that make polygons in vector format. In order 
to minimize the number of vertices, unnecessary points will be 
deleted. Then the procedure of eliminating discretization noise 
continues by testing the significance of each point in shaping 
the polygon. First, in order to keep only significant points and 
delete points on straight lines, all points with altitude close to 
zero will be eliminated. In previous similar approaches in 
(Douglas and Peucker, 1973; Weidner, 1995), the computed 
altitude was used directly as a criterion of point significance. 
    
    
    
   
    
   
   
   
  
  
    
   
   
   
  
    
    
    
   
  
  
   
    
    
  
  
   
  
  
  
   
   
  
   
  
   
   
   
   
   
   
    
   
  
   
   
Interna 
Conseq 
their cc 
Howev 
in this 
The coi 
triangle 
and m 
imperfe 
procedi 
comes 
from ze 
the star 
thresho 
damage 
shown 
segmen 
: 
(a) Oi 
As shov 
not con 
First, vi 
were g 
external 
was det 
the ext 
borders. 
between 
computi 
and if tl 
will be 
were gr 
  
For eacl 
are estin 
irregular 
reweight 
paramet 
intercept 
inside ez 
i.c, each