this 
rban 
this 
oint 
oint 
rage 
the 
liew 
and 
odel 
(k) 
iture 
ctor, 
rmal 
n in 
eter 
A: is 
is 
L,3D 
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XXXIX-B4, 2012 
XXII ISPRS Congress, 25 August — 01 September 2012, Melbourne, Australia 
Where the prime and double prime denote to the first and the 
second derivatives. According to the above equation we can 
analyse and then extract an object from a point clouds data. The 
algorithm for extracting an object from the point clouds will be 
compare the magnitude and the singe of the curvature at the 
point p and at the neighbouring points, when a significant 
change has been found the object will be isolated and then 
extracted. 
The second operator is defining the curvature of each three 
points which build a surface. Let S be the polyhedral surface 
which the point clouds as its vertices (S c R?, p € S), therefore 
the directional curvature function of surface S at point p is a 
quadratic form as expressed in the following (Taubin 1995): 
(uy (A IN Ce, 
WORT (8 75) ©) (84.5 
Where k,(T) is directional curvature of S at p in the direction 
of the unit length vector T (the tangent to S at p). Also T — 
t,T, + t, T, which {T, , T,} is an orthonormal basis of the 
tangent space to S at p. Finally kj! — k,(T,) , k2? — ky (T;), 
and kj? — k$' . In this study we assume kJ? = k2! = 0 , that 
means the vector of (T, , T5] is principle directions of S at p. 
k,(T) = (s)' ki 0 (t) (Eq. 4) 
p Tta 0 k2? t; aq 
With analysing the singe of kj! and k2? one can recognise the 
changes of the curvature and then the object can be extracted 
and isolated. 
2.1 Proposal for extracting buildings 
Firstly, no attempt was made in order to develop a template as a 
framework for detecting and extracting building from the point 
clouds in this study. Some studies assumed the buildings have a 
rectangular shape and mostly focused on detecting and 
extracting rectangular shapes from the images (Tupin and Roux 
2003). We believe that there is no reason to assume all buildings 
have a rectangular shape. In generic term we logically accept 
that all buildings have a closed and regular geometric shape and 
most of buildings’ roof has an elevation greater than 2 meters 
from the ground, as well as edges of roofs in most cases remain 
at the one level. According to the above knowledge, the 
operator which was developed based on either equation 2 or 
equation 4 searches for buildings within a defined area. It has to 
be noted that when speaking about detecting and extracting 
buildings in this study, it means to detect and extract the roof of 
buildings. After extracting buildings from the point clouds, all 
detected buildings will be checked and corrected before 
transform and register on the image. Then the buildings will be 
converted to raster format and then transformed and registered 
on the image. For transformation, an initial and rough 
orientation will be carried out. Then the algorithm will extract 
the corresponding building from the image according to the 
geometric information were obtained from the point clouds and 
using feature-based matching approach. Then each extracted 
building from the image will transform and register on the 3D 
models, the parameters of orientations were computed 
individually for each object as Homainejad (20112) explained. 
2.2 Proposal for extracting trees 
For extracting trees from the point clouds, the algorithm mostly 
focuses on the Laser Scanner facts. New generation of Airborne 
Laser Scanners are able to record up to four backscattered 
waveforms returned from trees. Each emitted signal after hit the 
tree will penetrate to the ground, and each time when the signal 
hit the trec’s branches a reflected waveform will return to the 
Laser Scanner. Unlike the returned waveform from a surface 
which forms a trace of a straight line, the returned waveform 
from trees scatters in an area. Therefore, for extracting a tree, or 
group of trees a grid is designed. The size of each cell of the 
grid is equal to the size of Laser Scanner footprint. Then the 
origin of the grid and bearing of its main direction will be 
defined. The algorithm will search according to the equations 2 
or 4 and above initialisation inside each cell for (i) checking the 
density of point clouds inside of each cell, (ii) testing the 
variation of the curvature between points. If algorithm 
recognises the variation of the curvature is changing inside of 
each cell and the density is different with a defined criterion, 
then it will recognise and extract trees. 
2.3 Proposal for extracting roads 
Extracting road requires following initialisation. At the 
beginning a number of points on the road will be captured for 
initialising. The points include the start, the end, and changing 
directions points of the roads. Then the maximum and the 
minimum width of the roads will be measured. The algorithm 
according to these data and one of the equations 2 or 4 will 
detect and extract the road from the point clouds. Then the 
extracted road will be checked and corrected. Then the extracted 
roads will converted to raster data and will transform and 
register on the image for extracting the corresponding roads 
from the image. Finally, the extracted road from image will be 
transformed and registered on the 3D model. 
2.4 Proposal for extracting crown land 
Each object after extraction will be removed from the point 
clouds and finally after removing all extracted object a bare 
point clouds will be available. The bare point clouds will 
represent the crown land. In this case, the point clouds consists 
many gaps regarding to the extracted object that can be easily 
filled by different methods such as interpolation method, but it 
has to be aware sometimes the interpolation will fill an area 
which basically cut off for a development. A supervised 
interpolation has to be carried out for achieving the best result 
and reducing any error. 
3. STUDY AREA 
As mentioned earlier, the author is participating in an ISPRS 
test project on urban classification and 3D building 
reconstruction. The project focuses on reconstruction 3D model 
by using aerial image and Laser Scanner data. The test data set 
was acquired over Vaihingen in Germany. Three areas have 
been chosen for participants, nevertheless, each area has slightly 
different from two other areas, but each area includes building, 
road, and trees. Digital aerial images were acquired by 
Intergraph/ZI DMC camera with ground resolution of 8cm. 
Point clouds has been captured by Leica ALS50 system with 
point density of 4 pts/m?. In this paper the results from two 
areas will be shown. The first area is in the centre of city with is 
characterized by dense development consisting of historic 
buildings having rather complex shapes; also the area includes 
trees. The second area includes multistorey buildings. 
4. IMPLEMENTATION AND ANALYSIS 
The focus of this research study is primarily on extracting 
objects from DTM or DSM which are generated from point 
clouds or the point clouds itself. Then transforming and 
registering the extracted object on the aerial image for assisting 
181