|o
S
f
l-
International Archives of the Photogrammetry, Remote Sensing and Spatial Information éciences, Vol XXXV, Part B2. Istanbul 2004
with Lo: Length of the respective segment. The probabil-
ity p(@ < aj) can now be calculated using the Gaussian
probability density function for a:
"Xp
pla <a) = Flog) = J f(a)da
The Hint Hg allows three interpretations as shown in Tab.
3. In contrast to Hp this Hint also supports —G. As the ob-
jects are assigned to the respective ATKIS segment without
considering the geometric relation it is reasonable to sup-
port —G here.
a [oor P
wa {G} pla = dr ' Qcov * Pcon
Wao [^G] (1-p(a << ap)) ' Qcov * Peon
wags | Oc 1 — p(wa1) — p(waa)
Table 3: Hint HC:
4.3.3 Combining Hints for one ATKIS Segment The
Hints defined in the last two sections refer to the relation
between an ATKIS segment and a segment of the extracted
objects. Applying Dempster's Rule all Hints referring to
one ATKIS segment can be combined. The Hints HR and
HE are thereby computed, representing the overall coinci-
dence of the ATKIS-Segment to the model with respect to
both relations. The frame of discernment © = Op x Og
containing hypotheses whether the ATKIS segment fits to
the model (H^) or not (H?),
5 RESULTS
In this section preliminary results of the introduced ap-
proach are given. In order to investigate whether the qua-
lity of ATKIS objects is reflected by means of Extracted
Road Objects and Linear Local Context Objects some ex-
periments were carried out. Two sets of ATKIS road data
have been prepared: set A) just contains objects with a cor-
rect geometry. For set B) the correct objects have been
rotated in order to obtain incorrect geometries. Each sets
contains 1851 ATKIS segments.
The Extracted Road Objects are obtained by the approach
presented in (Gerke et al., 2004). The parameters are trim-
med for a very strict road extraction, because the influence
from artifically inserted road segments (due to automatic
gap bridging) should be very low. Those gaps are often
caused by vegetation and the intention of the following ex-
periments is to test if explicitely inserted context objects
give adequate evidence. As the road extraction algorithm
is not able to reliably extract roads in built-up areas the ex-
amples are restricted to open landscape areas. The rows
of trees representing a class of Linear Local Context Ob-
jects are captured manually and the parameters for the rows
of trees are uniformly set to wo — 1m, A,o -— 0.2m,
Appo — 2m, Apao — 3m, ayo = 0.6m, peon = 1.
The diagrams in Fig. 3 and 5 show the results in the form of
absolute histograms, keeping in account all assessed seg-
ments. The five histograms per diagram show from left
to right: 1) the support for the ATKIS segments regarding
the topologic relation, 2) and 3) the support and the plau-
sibility for the ATKIS segments regarding the geometric
807
= f | +
/F y | 777%
0
sp E RU EE S.C
ATKIS: Apa = 3m, Awa = 3m, just extracted road objects.
316 ATKIS segments assessed.
ATKIS: Apa = 3m, Awa = 3m, extracted road objects and
rows of trees. 475 ATKIS segments assessed.
ATKIS: A, 4 = Om, Awa = Om, extracted road objects and
rows of trees. 446 ATKIS segments assessed.
Figure 3: Results for Correct ATKIS Data
