CIP A 2003 XIX th International Symposium, 30 September - 04 October, 2003, Antalya, Turkey 
491 
scanning, 
ural heritage 
ect. In the 
the model 
clouds that 
advantages 
act sections 
cal analysis 
to create a 
i 
% 
,!S Z210 
Italy 
that has 
geometry 
without 
)me very 
complex if the surveyed object has a large number of 
discontinuities (for example, a historical façade with 
columns, eaves, statues and niches on the façade, etc.). 
In this case the only way to be able to carry out a correct 
solid modelling of the object is to use segmentation 
techniques. By segmentation we mean the extraction of 
geometric information from a set of data and their 
partitioning into uniform entities. Each of these entities will 
then be modelled separately. The final form of the model of 
the object will be generated from the set of the various 
portions that have been separately modelled. 
2. SEGMENTATION TECHNIQUES 
2.1 Manual segmentation 
As previously shown, in order to carry out a solid modelling 
starting from a three-dimensional point cloud, it is necessary 
to segment the object into its simplest portions. Modelling 
software usually allows this type of operation to be carried 
out in a completely manual way. The operator chooses and 
selects a portion of a point cloud on the screen and models 
this portion of the points to create meshes or nurbs. 
On other occasions it is possible to manually segment the 
objects using graphic management software such as 
AutoCAD. Therefore the objects are segmented and some 
portions of the object are placed on different layers. 
This type of segmentation which, according to what has been 
described can be defined as manual requires a remarkable 
effort by the operator to identify the portions of the points. 
This is even more the case when the model is dense with 
information and complex in form. 
Today an increasing effort is being made in the field of 
research into the implementation of automatic segmentation 
techniques to resolve this problem. 
2.2 Automatic segmentation 
The problem of segmentation can be found in many scientific 
applications. 
As far as laser technology is concerned, there are many 
researchers who at present are working on comparisons of 
segmentations. These are usually carried out on laser data 
derived from aerial platforms. Some very interesting results 
have already been obtained. 
The case of aerial scanning differs however from that of 
ground laser scanning. The laser data from aerial platforms 
are distributed in planimetry while the most obvious 
discontinuities can be found at heights. 
In the case of ground laser scanning, the data are widely and 
differently positioned in space and the complexity of these 
dispersions depends above all on the complexity of the 
surveyed object. 
Although the field of aerial laser scanning is rather different 
from the field of ground laser scanning (close range), some of 
the segmentation algorithms and techniques that were 
developed for segmentation of DDEMs obtained from aerial 
platforms can easily be applied or adapted for ground 
surveying. 
2.3 Segmentation through the identification of the main 
plains of the object 
In order to be able to model the point clouds acquired with 
ground type laser scanners it is first of all necessary to divide 
the 3D model into its main plains. In the case of a building, 
the main plains can be represented by the individual facades 
or, as is the case for more complex buildings, by portions of 
these facades. 
In order to be able to segment the object according to its main 
plains, the authors have developed specific software which 
functions as follows: 
A point that belongs to the scansion. This point can be 
imposed by the key board or by the operator, or can 
even be extracted by chance. 
A search radius around an initial point is established 
and all the points that fall into to this area are 
considered. The set of these points is used to determine 
a mean plain that passes through the points. This plain 
is calculated using the least square estimation. 
If the parameters of the estimation that derive from the 
calculation of the mean plain that passes through the 
chosen points are considered to be sufficiently correct 
(this decision is up to the operator), or in other words if 
the maximum rejects obtained by the estimation are 
lower than a previously imposed value, it can be 
considered that one of the main plains has been 
identified. After having identified one of the main 
plains, all the points of the scansion that are at a closer 
distance from the plain, according to a previously 
established reference, can be considered as belonging to 
the plain. It is the operator who decides the maximum 
distance between the points and the plain for the 
aggregation. 
o 
o 
o 
• First aggregation point 
® Aggregated points 
© Non aggregated points 
The mean plane 
Figure 2 - An example of a plain considered valid, as seen 
from above. 
In the case in which the parameters of the estimation of 
the plain are not considered valid, that is, when the 
number of rejects is too high, they are considered as not 
belonging to a plain that can be considered as being a 
main plain. 
The aggregation operation is continued. Another point 
to which a layer has not yet been assigned is extracted 
and the operation is repeated. 
All the procedures that have been described are repeated 
until all the points of the scansion have been assigned a layer 
or until a maximum number of plains, as imposed by the 
operator, have been found. 
The described procedure is nothing else but a first very 
rough segmentation, but it allows one to easily separate, for 
example, two facades of the same building that are oriented 
in a different way.