ee ae
1] DTM (0,5m resolution) |
118- | Ground laser points |
qe Low non-ground laser points, Y wk
Wot Non-ground laser points | 4*
1 x
114 : ;
4 fi 4 2
112 : Mey, ;
3 49r Vb. vi x
= 4 C ys Ma ^i x e
= 1104 one. Clin, x
= VO Ru f i) X
© J être
— 108 :
106 + V = OR S
4
104
102 st
T : T : T x T T T y T T T T 1
0 20 40 60 80 100 120 140
Along Profile (m)
Figure 6: profile of both the final DTM (in gray) and the clas-
sified laser points over an other location of the Roujan data set.
red circles, vertical blue dashes and green crux are respectively
ground, low non-ground and non-ground laser points.
of laser data (more than 3.10° points). The propagation mech-
anism efficiency is directly linked to the point density and as a
result to the neighborhood dimension.
6 DISCUSSION
The algorithm presented in this paper classifies raw laser data
into three classes: ground, non-ground, and an intermediary one
with points that could belong to one of the previous ones. We
did focus on the quality of the resulting classification as well as
on the accuracy of the DTM. Unfortunately, we did not have in
situ measurements for any of the laser data sets we tested. As a
result, ground is estimated only with regard to laser points and it
was not possible at the time of this study to provide any index of
the classification quality. But we planed to compare classification
results over urban areas with the related cadastral map.
As far as the classification part is concerned, we may point out
interesting behaviors of the algorithm. At first, the lower bound
of the low non-ground class influences the final 3D point label.
This parameter can be tuned with an a priori knowledge of the
landscape composition (in a urban environment, cars (low non-
ground) are generally lower than 1.5 m whereas in a rural area,
non-ground vegetation begins 20 cm over the ground). It is of
importance because points classified as low non-ground ones will
not be considered to be attractors when it comes to compute the
deformable model. Therefore, the final DTM will not be as accu-
rate as it should have been.
Secondly, the ground estimation is performed using a set of laser
points included into a defined neighborhood. The shape of this
neighborhood do not influence the classification. But for maxi-
mizing the continuity of the initial surface, the overlapping ratio
must be large enough (up to 7096). The neighborhood size must
take into account the point cloud density (card(V) > 20 pts)
as well as local variations of the topography (C « 20 m). It
is always better when a real ground point is included into the
neighborhood since laser points will be classified with regard to
computed ground points and propagated neighboring values.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
"wo | - Affined DTM (0,5m resolution)!
s | ——— Coarse DTM (3m resolution) |
110 - jo ud
e
4 =
x.
— 108- a
= x
= + \
5 1064 \ /
D
X X /
A 1j
104 1
N /
4 X /
V
102
T T T T T " T 7 T T T T T 7 =
0 20 40 60 80 100 120 140
Along Profile (m)
Figure 7: Black line depicts Sin, (DTM after the classification
process). Gray line is the final DTM at the top end of the process.
Figure 8: Resulting 0.5 m resolution DTM of the classification of
a mountainous area near Montmirail, South of France. These two
pictures represent the same laser strip split up for the presentation
concerns. Itisa 16 km long and 2.3 km wide strip.
As a matter of course, the ICM algorithm will not make the curve
converge toward a global energy minimum. But we must keep
in mind that the laser point is already an integration of the real
backscattered energy of the laser impact. Therefore, the result-
ing punctual altitude is slightly noisy and the final attractors may
not belong exactly to the true ground. We thereafter just need
to compute a local minimum energy associated to the final bald
Earth model.
The propagation strategy ensures a coherent diffusion of the alti-
tude information. As it is presented in Figure 9 right (Amiens),
roads are first computed following the lowest neighboring aver-
age altitude, then buildings are tackled. The chosen route depends
on the initial seed. It is an on-going development to check the dis-
crepancies of two point clouds classified from a different starting
point.
Interna
Figure
Montn
cessing
cessed
We ha
fying ¢
into th
points.
of the
deforn
high re
feature
In a ne
ular pc
We we
derive
ods de
The ai
affiliat
GREF
jan as
discus
Axelss
ans ap
Sensir
Bretar
ing int
Vol. X
mote |
Eckste
digital
cal En
Dimer
Elma
laser s
