ISPRS Commission III, Vol.34, Part 3A „Photogrammetric Computer Vision“, Graz, 2002
Geometrical descriptors
The geometrical properties of an object in a point p; can be
described intrinsically defining an inertial system centered in
this point. We mustn't forget that we need to know those
properties or descriptors that simplify object segmentation. In
order to reach this result, we note that is desirable that these
descriptors are singular in discontinuities or, in other words,
that they have a peak value. We can't consider sign change of
the descriptors, because of the presence of noise in data.
In the following are described three different descriptors, static
moment, curvature and junction, but it is possible to define a
non-limited number of properties, increasing the dimensions of
the space of descriptors. So, the dimension of the problem is not
only the dimension of the space of the objects O", but the
dimension of the union of this space with the space of
descriptors D":
S = oO" cb}
Static moment: the module of the first order moment (or static
moment) is maximum at the edge and, in a smaller measure, at
a change of curvature. Static moments are defined as:
S =) ME)
S, ES ms tz)
S, 2S tn)
The three static moments can be combined in a comprensive
descriptor that is the total static moment:
SS
EESEREENEUTE
un
jl
i
+ |
i
e
V
| hb
i^
*
MI
VOLU
B
Figure 2 — Total static moments in an airborne laser scanning
data set by TopoSys. In (A) the most elevated values of static
moments are in black. In (B) the black points represent a profile
in the data set, the yellow points the static moment (with sign
changed). Note the peak value in discontinuities.
Curvature: in discrete geometry it is possible to define
curvature in many ways. Fortunately we don't need the exact
value of curvature; a functional one is sufficient, beacuse we
are looking for the location of peak values, not for their
magnitude. In the hypothesis that the point set is a surface, the
ratio
=. Can
ar
is a functional of the ray of curvature. In fact A, is null for a
planar surface and A, is not null; that leads to an infinite
value for p . Increasing curvature, A, Increases and Am, 1s
A - 291
constant, while À … is less than 4,,,. When A; becomes
min
equal to À the two eigenvalues swap, A, becomes
max ?
constant and A4,,, increases with curvature. A functional of
X
curvature is the ratio
A d
Een
Junction: the eigenvalues encode the magnitudes of
orientation (un)certainties, since they indicate the size of the
corresponding ellipsoid. At curve or point junctions, where
intersecting surface are present, there is no preferred
orientation and the eigenvalue A, has peak values. A
junction map that represents the location of the peak value
for À … is very similar to the static moments map of the
figure 2.
min
Data distribution anisotropy
Distribution anisotropy is a problem present in point cloud
data, such as range data, and leads to a classification of the
scanning shape as a real shape. PCA simplify anisotropic
data set processing: principal components are referred to a
spheric neighbourhood, but it is possible to refer
aggregation criteria to an ellipsoidic neighbourhood, whose
semiaxes are defined by RMS coordinates:
0
um max . - mid . as
€x = r $ed = r > Can = r
c. Oo oO
max max max
where 9... , 6, and o,,, are the RMS in the directions of
min ? mid X
a , P o and P. ,
neighbourhood.
With reference to figure 3, that represents a laser scanner
anisotropic data set, we note that the sampling density in the
first principal direction (~7 points/m) is greater than in the
second principal direction. (~1 points/m). Using the
ellipsoidic neighborhood in aggregation, we can take
advantage of the better definition of the measures in the first
principal direction.
and r is the radius of the spheric
[m]
spheric
08 = neighbourhood —
4o 08 08 da 702 Mg Te 04 08 08 18
{m}
Figure 3 - Normalized ellipsoidic neighborood in
anisotropic case; the arrows rapresent the direction of the
principal axis (e data points, A centre of mass). The point
set is from a TopoSys I airborne laser scanner.