CI PA 2003 XIX th International Symposium, 30 September — 04 October, 2003, Antalya, Turkey
492
The software that was developed for this kind of
segmentation was prepared in Visual Basic 6 environment.
The choice of the parameters that are necessary for the
search for the aggregation points is left to the operator. The
choice of these parameters is of fundamental importance for
a correct identification of the plains and it greatly depends
on the density of the point cloud and on the size and
complexity of the model.
• First aggregation point
The mean plane
Figure 3 - An example of a plain considered not to be valid
as seen from above.
It is possible to obtain aggregations that are completely
different from each other by varying these parameters.
The software that has been developed by the authors helps
remove the necessity of making a chance choice of the
starting point.
The operator can a priori choose from which point to start
the aggregation procedure.
2.4 Segmentation using region growing and analysis of
the main components
One of the models that were developed for the segmentation
of the laser point clouds from aerial platforms is the region
growing technique and analysis of the main components.
This method considers the problem of grouping sets together
of single entities that have certain properties defined by the
partitions in the n-dimensional space of the objects. The
algorithms of this method look for the solution using
clustering techniques that originate from image analysis
techniques. The function principle is based on the hypothesis
of the knowledge of certain geometric properties, called
describers, whose particulars one wishes to study to identify
the partitions in a space. The solution of this problem can be
obtained through a simple implementation: the aggregation of
the point entity under study or region growing. The space of
the describers is defined through the local analysis of the
main components, which allow the local geometric properties
of the object to be defined in an intrinsic reference system.
The connection algorithm of the growing region begins the
aggregation starting from a chance seed, which can be any
entity which has not yet been aggregated; this point entity pi
is memorised in a 0 level of a tree structure. Each point is
identified by an indicator, which refers to the data base.
Starting from this chance seed, the algorithm searches for the
nearest points in an a priori defined neighbourhood in the
space that unites the spaces of the object with the describers.
The neighbourhoods, that is, the entities contained in such
neighbourhoods, are aggregated and memorised in level 1 of
the structure, as a new branch of the tree. The algorithm
continues the aggregation starting from the entity contained
in the last level, and so on until the terminal branches of the
structure are obtained. The terminal branches are determined
by the occurrence of at least one or two events. The first is
when there are no other non aggregated entities in a
neighbourhood; the second, when the branch contains a
particular. The algorithm finishes the aggregation of an object
when all the branches of the tree have reached a terminal
point; it then starts another object starting from a new chance
seed.
The analysis of the main components (PCA) can be used to
study the structure of a set of data. The PCA searches a
vector base for a (sub) space that:
maximises the variance of the distributions
minimises the correlations (which is equivalent)
The local properties of the objects in the three-dimensional
case or in the more generic multi-dimensional case can be
described using some describers such as:
inertia moments and tensors
curvature
static moments
The software used for the elaboration utilises two different
types of algorithms that combine region growing techniques
with an analysis of the main components (Roggero, 2002).
3. EXAMPLES OF SEGMENTATION
3.1 Identification of the main plains of the object
The data relative to the façade of a building were treated to
validate the efficiency and actual correct functioning of the
proposed method.
Figure 4 - Segmentation of a building façade. Test area of the
Politecnico di Torino - Department of Georesources and
Land.