International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004 
Therefore, its 3D modeling is enabled by the 
unification of all flat parts in the 3D model. Next, 
the break-lines provide the object edges and, 
ridgelines, and these can used to accurately 
determine the points to be modeled of the TIN along 
with the flatness classification results. The technical 
outlines are presented later. 
3.2 Classification of The Flat and Non-Flat Areas 
In order to develop a robust filtering method for 
topographic surveying, 3D point cloud data for a 
topographic scene was acquired using a terrestrial 
laser scanner. A small mask with 30*30 cm area is 
used instead of the 3D information samples. After a 
3*3 point mask is generated around an interest point, 
the mask size expands to 30*30 cm by computing 
the plane coordinates for the neighbor points.The 
mask is then transformed so that in the first step, a 
normal vector for the mask becomes parallel to the 
Z-axis (Figure 2). In the next step, the Standard 
Deviation (S.D.) is computed for the interest point. 
The threshold value should be considered while 
classifying the interest points into flat (ground 
surface, structure walls, etc.,) and non-flat areas 
(trees, bushes, sky, ctc.). The threshold value is 
determined on the basis of the measured data. 
z 
! 
ISN < Normal 
Vector 
N 
   
Fig.2 Coordinate Transformation 
However, the ends of the big slopes are classified as 
non-flat points. Figure 3 shows example of flat and 
non-flat points in the big slopes. In order to resolve 
the second issue, the following procedures were 
created, and the authors termed this process as "S.D. 
Saving". 
The S.D. Saving Process: 
* If an interest point is detected as topographic data, 
the Z values of all the points in the mask are 
compared with the S.D. values. 
+ If the Z values of each point are smaller than the 
S.D. values of the interest point, then these points 
are recorded as topographic data. 
  
  
   
Treated 
as Flat 
Point 
© -JFlat-Points 
O ONon-Flat Points 
Fig. 3 Examples of Flat and Non-Flat Points 
3.3 Derivation of the Break-Lines 
The break-lines (e.g., object edges, ridge-lines) 
provide important morphological information. 
Although these are indispensable features for DTM 
generation, city and object modeling, problems with 
automatic detection of the break-lines still persist. A 
technique for automatic detection of the break-lines 
using the flatness values was developed?. The 
algorithm is closely related to edge- preserving 
smoothing. however, only the flat area is 
smoothened, and points with a larger S.D. within the 
non-flat area are emphasized. A small mask was 
used for smoothening. A mask size of 30*30 cm was 
sufficient, and an interest point was smoothened 
using the following equation: 
Dp "Agri 
Sm Is (1) 
> Pi 
where, 
£ ;: S.D. for an interest point, 
Ag; : Difference in S.D. between an interest point 
and its neighboring points;, 
defined as Ag ij- Zi 8p 
p ; Weight of the i point; where the weight is 
defined as the square of the values between the 
interest point and each neighboring point. 
In order to detect the break-lines, smoothening of 
only the flat areas is repeated. Repeating this three 
times was sufficient, and the points with a larger 
standard deviation within the non-flat areas were 
emphasized. 
As a result, the break-lines were derived, and these 
are shown in Figure 4, and figure 5 presents break- 
   
  
    
  
  
   
  
  
  
  
  
  
  
  
  
  
   
   
  
  
  
  
  
  
    
  
  
   
   
   
  
  
   
  
   
    
  
   
   
    
    
  
   
   
  
   
  
   
   
   
  
  
lin