International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
3D object extraction by means of a region growing process 
based on all of first/last pulse of range and intensity data. 
2.1.1 3D regions extraction based on morphological 
operators 
In this study we have adapted morphological operators for 
extraction and refinement of the 3D regions (Gonzaless and 
Woods, 1993). The adapted process may be outlined as follows: 
in the first step the initial regions are extracted by a top-hat 
morphological operator: 
Re gions = DSM - (DSM <b) (1) 
where b is the structuring element function, and © denotes the 
opening operator given by: 
fob=(f Ob)@b (2) 
(f ®b)(s,t) = 
min{/ (s —Xx,t-y)- b(x, y) — x), (t-y)e Dy;(x, ye Dy] 
.(f ©Ob)(s,t) = 
min {f(s +Xx,t+y)— b(x, y)Ks TXxL(t-y)e Dpi(x.y) € D, | 
where © and @ denote the grey scale Erosion and Dilation 
operators and D D, are the domains of f and b, respectively. 
The output of this process will be binary data with the values 
one and zero denoting the 3D objects and the background 
respectively. This stage is then followed by the binary cleaning 
and the opening morphological operators. In this way only the 
objects of interest will remain and insignificant objects and 
artefacts are excluded from the extracted regions. 
The 3D regions that are extracted from the range data may be 
quite close and thus morphological operators may fail to isolate 
them as 3D individual objects and hence they may be 
erroneously classified as a single 3D object. This defect is 
resolved by exploiting other information available in the data 
set. That is, the relief and textural information. To utilize these 
information, the extracted 3D regions are mapped into the 
intensity information of LIDAR data. This leads to the 
generation of the preliminary regions. 
2.1.2 Object Extraction Based on Simultaneous Fusion of 
RTS Information 
This stage is designed to extract of all objects of interest in the 
object space, by means of a region growing process. It is 
assumed that a region that belongs to a single object should 
demonstrate a uniform variation of the structural values for all 
pixels included in the region. For example for g 3D region that 
belongs to a tree, the fluctuation of the values of the structural 
components should remain relatively uniform for all pixels on 
thé region. This means that if the structural variation exceeds a 
gertain level, the possibility of the presence of a second object 
^in the region is signalled. To express the relief variations for a 
3D object, a relief descriptor is determined using the following 
strategy: A normal vector is computed for a local surface 
defined by a 3x3 or 5x5 window array constructed around 
Table 1. Linguistic variables and labels of fuzzy reasoning structure in region growing process 
the position of each point on the 3D region. Texture metrics are 
computed over a local collection of facets, and represent how 
the directions of the normals are distributed about the local 
mean normal (Figure 2). 
  
Figure 2. Analytical description of the surface based on normal 
vectors. 
The three components of the normal vector are given by: 
Kj a; 
Di — | f (3) 
[2 2 
M; A I | 
where, @; and /; are the coefficients of the surface given by: 
P(w,h)=a;w+ BP; h+ ri (4) 
This surface is determined within a predefined limit specified 
by the window size, w=1:Width,h=1: Height . Based on these 
vector components, the relief descriptor, k, can be defined as: 
N-I] 
bs 4 (5) 
N-R 
2 
N; Yi [XN N 4 
where R* = > Ki + > 4 + SM; and N is the local 
iz} j=! i= 
surface window size (Besl and Jain, 1988). 
Thus, the value of k quantitatively expresses the overall relief 
variation of a region. The large value of K indicates that the 
region comprises rather uniform relief variations. The large 
value of k, on the other hand, denotes the non uniformity of the 
relief fluctuations (Samadzadegan, 2002). 
Taking into account the complexities and the fuzziness 
behaviour associated with these consistency checks, a fuzzy 
based region growing approach is adapted as follows: The 
region growing starts from the pixel located in the centre of the 
gravity of the 2D regions. For all neighbouring pixels, based on 
a fuzzy reasoning strategy and Mamdani inference type, a 
consistency check is carried out (Zimmermann, 1093). 
The linguistic variables to be fed into the fuzzy reasoning 
module are: (1) the pixels and relief fluctuation values in both 
of intensity and range data (the value of k ), (2) the size of the 
region, and (3) the difference of the pixel values ( TextureDiff) 
and the height value (Relie/Diff). The “Size” item is used to 
exclude the objects that are smaller than a predefined size. By 
the region growing process the regions undergo one of the 
following changes: (a) the region remains unchanged if it 
satisfies the consistency criteria, (b) the region is subdivided 
into two or more regions if consistency criteria are not satisfied, 
(c) different regions are merged if they are consistent. The 
  
Linguistic Variable 
Linguistic Labels 
  
  
  
  
  
  
  
  
  
Texture Solrregular, Irregular, Regular, SoRegular 
Relief Solrregular, Irregular, Regular, SoRegular 
Input Size SoSmall , Small , Medium , Large , SoLarge 
TextureDiff SoSmall , Small , Medium , Large , SoLarge 
ReliefDiff SoSmall , Small , Medium , Large , SoLarge 
Output Grow NotGrow , ProbablyNotGrow , ProbablyGrow , Grow 
  
  
922 
Intei 
lingi 
func 
Tabl 
As a 
the 
obje 
fuzz 
base 
unif( 
two 
2.2 
Asn 
conc 
elem 
need 
(Sam 
be € 
parar 
locat 
struc 
proce 
signa 
The 
shape 
by th 
for a 
indic: 
descr 
reliat 
witho 
As m 
enhar 
paran 
natur 
mathe 
takes 
object 
recog 
appro 
It is 1 
not r 
inforn 
descri 
only « 
structi 
this ti 
first s 
degre«