9.
gs
a Fadl a # {ses ^ à *
ze ve duet MEAT 20
wu wy rM, c0 NP
ass pe. HUNG di i07 4
US £ nd Bard a dx
re T i . Te SH
aed + 3 MU va,
« & lu Ld
Nr Ms "o
am + ANM vi VE pd
&
+
A
m
dp
zS AM
+
14*
auum
c
3
- zu
Moa
"te Wi
x
+ > ^
pe NE e
> ida:
t A.
+,
+
#
.
arie
y
?
v
+
y TR
SESE
*
T
x
*
A
E
MES
*
*
a
m
+
t
*
a
D4*
he TW
a
a
TAS
Ni
Et
4
cdi
e
few
*
cns
CA
3
+
4
Le
r FOR -
a
TOR
sd
A
A
E
Sort,
ee i ia.
UK. 3 94,44. CE
í , n a
47 47 A; GER 067 4 — am SM 06 DE
Figure 3: A single target scanned from different positions.
The angle that is formed by the z-axis of the scanner and
the surface on which the target belongs is 90" for the first
and 45? for the second image respectively.
For targets scanned with the scanner facing directly the surface
on which the targets belong, a rather strong pattern appears. All
points are classified in the correct classes. When the scanning
angle increases, the classes of higher and medium reflectance
appear to get confused. Furthermore, the points that belong to
these classes are distributed unevenly. This indicates that the
results of the ‘radcent’ algorithm will be poorer, due to the fact
that the weighting will be forcing the centre of the target
towards the centroid of the class that consists of the points with
the highest reflectance.
Another interesting aspect is that the class that consists of the
points with low reflectivity presents the same behaviour in all
cases. Taking all of these into account, two algorithms were
developed.
The first algorithm is named *fuzzypos'. The initial step for this
algorithm is to classify the points of the target according to their
reflectivity. Using the ‘fem’ function, it is required that all
points are classified into three classes. After classification is
completed, the classes are recognized by calculating the mean
value of the points that are assigned to each one of them.
Finally, the coordinates of the centre of the target are derived
by simply calculating the mean position using the two clusters
with the largest mean reflectivity values. This process yields
substantially better results than the ‘radcent’ algorithm, as no
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B5. Istanbul 2004
weighting occurs. In this case, weighting is redundant because
the two classes that are used cover the whole reflective area of
the target.
The second algorithm, named 'fuzzyposfine', uses initially the
*fuzzypos' algorithm to calculate the centre of the target.
Afterwards, a plane is fitted on the surface of the target and
using the parameters of the plane, the ® and ¢ rotations of the
surface are calculated. The origin of the system is transferred to
the calculated centre of the target, and the rotations are applied
in order to transform the points of the target to the XY-plane of
the new system. Then, a square area of Sem x 5cm centered at
the origin of the new system is selected and only the data
contained in that very area are used hereafter. The classification
process is repeated, and the points that belong to the class with
the lowest mean reflectance value are used for estimating the
centre of the target. This class has been found to correspond to
the circular area of low reflectance, which surrounds the centre
of the target. By calculating the centre of this cluster and
transforming back to the original system, the coordinates of the
centre of the target are derived. These algorithms may seem
rather complex but the quality of the results is significantly
better than any other seen in the literature. This is confirmed by
the experiments presented in the following section.
Two more groups of algorithms were developed using grid and
tin surface models. The first group is based on gridding. Using
the Matlab function griddata, a grid is created for the target
using the coordinates of the points of the point cloud. In order
to achieve better results, the data are statistically processed for
the noise to be removed. Specifically, a plane is fitted on the
surface of the target and the standard deviation of the distance
between the points and the surface is calculated, but only the
points that are within £ 1.96 ¢ (95% level of confidence) from
the calculated surface are kept. The spacing of the grid is set to
5mm and the surface model of the target is created. Using the
grid, a model for the reflectance is also calculated. Using the
two grids, two algorithms were created. The first one is named
‘gridrad’ and calculates the centre of the target in the same way
the ‘radcent’ algorithm does, by using the information of the
two grids. The second algorithm is named 'fuzzygridrad' and
applies the *fuzzypos' algorithm for the data of the grid. The
second group of these algorithms is based on delaunay
triangulation. The two new algorithms are named *delrad' and
‘fuzzydelrad’ and the only difference to the previous ones is the
model used.
4. EXPERIMENTS AND RESULTS
Several experiments were conducted to evaluate the
performance of the new algorithms and compare it with the
performance of the algorithms mentioned in literature. In
particular, two series of experiments were designed and
conducted. The first series, involved scanning several targets
that were mounted on the pillars of the EDM internal
calibration baseline at NTUA. Multiple scans of the targets
were obtained from two positions. The second series involved
the scanning of five targets that were mounted on a wall. The
scanning in this case was carried out from various distances and
angles. The data collected for both cases were subdued to
processing in order to evaluate both the internal and external
accuracy of the results produced by the aforementioned
algorithms.
Internati
For the
pillars «
targets \
25m... TI
differen
spacing
targets.
scan for
For the
using th
position
merged
as deriv
of the t:
same [x
mean a
internal
in Tabl
algorith:
This is
These re
experim
scanned
with the
targets t
to creat:
targets
trimmec
exporte
the targ
merged
*fuzzyp
better, i
accurac:
algorith
The sec
external
presente
differen
For the
calculat
fine sca
and B.
targets
registrat
selected
absolute
software
evaluati
In Tabl
accurac
fine sc:
Absolut
Additio:
algorith
datasets
