resented by rules.
nted in the seman-
relations exist ac-
ition over the links
à condition and an
ew interpretation
> net. If a situation
action is executed
ssed node accord-
tected when all its
instances, is repre-
lete instance
art—of(ng) }
ymplete instance.
rious sets of rules.
or landscape mod-
three conceptual
escribes the scene
sents the objects
"The bottom layer
pecific photomet-
. If more than one
plied accordingly.
cabulary. E.g. the
) — geometry layer
jads, forests, and
Lis composed of a
be modelled sepa-
has to be visible in
ious stripes in the
d2D— region’ and
for segmentation
es respectively.
the objects of the
es of the semantic
its the knowledge
| expected in the
recognition three
shed: hypotheses,
ypotheses are not
tances contain all
parts. Complete
ory parts and con-
Layer
part-of
Vegetation
|
part-of : E part-of |
|
| Grassland | | Forest | |
1 À part-of Ÿ cdpart-of | part-of
| Forest Forest | Road
| Roof Edge | Segment
|
3 3
Wood Wood
3D —Surface
Grass
3D-— Surface
Textured
2D— Region
3D-Stripe
Camera
3
Asphalt
3D-Camera
3D-Stripe :
Homogenous
2D) —S
Fig. 4: Simplified semantic net representing prototypes of landscape objects
Interpretation proceeds primarily top down. Hypotheses are
propagated from the scene layer to the sensor layer to test
them in the sensor data. The propagated hypothesis at the
sensor layer calls methods for segmentation of textured re-
gions or homogenous stripes respectively. The result of the
verification is returned to the superior concept which consec-
utively generates new hypotheses.
If the verification returns competing instances each possible
interpretation is analyzed separately. For this each possible
interpretation is documented by a search node which con-
tains all concepts with their current interpretation state. Each
time competing interpretations occur the search node splits
into child search nodes. The leaves of the resulting search
tree represent the currently competing interpretations. To fo-
cus interpretation on promising search nodes they are judged
and ranked. The judgement computes the compatibility be-
tween expected concept properties and found concept prop-
erties by comparing the range and value slots of attributes. An
A' Algorithm selects the best judged interpretation for fur-
ther investigation.
4.2 Segmentation
The initial concepts in the sensor layer are instantiated by
segmentation of images. Presently two different initial con-
cepts with segmentation methods are available: the concept
‘textured 2D —region’ and homogenous 2D —stripe’.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
4.2.1 Segmentation of Homogenous Stripes: The extraction
of homogenous stripes is based on gradient filters for edge
detection with consecutive conditional local ranking to en-
hance weak contours. Thresholding yields a binary image in
which candidates for homogenous stripes show up as long
parallel lines. Figure 5a shows the segmentation result.
4.2.2 Segmentation of Textured Regions: Forests and grassland
of natural terrain are characterized by different textures. The tex-
ture of a class k, e.g. forest, is assumed as generated by a station-
ary ergodic process. The prerequisite of statistical independence
allows to compute the probability P;(ylk) of the texture process
from the luminance histogram of all intensities y in the learn re-
gion.
Generally the probability for occurrence of a luminance value is
not independent from its neighbours. Hence a texture model of
second order statistics is used. Local mutual dependencies can be
modelled by Gibbs random fields. Couples of neighbored pixels,
named cliques, are inspected. Three measures of co—occurrence
are computed for each clique:
— the probability P2(ylk) of common class membership,
— the probability P3(ylk) of a luminance difference within a
region,
— the probability P4(ylk) of a luminance difference between
different regions.
Gibbs random fields describe the joint probability
=> Vel)
c
Pkly) = ; Z = normalizing constant. (1)
1
7 ©
4
Viylk) = = > MIn(P(ylk)) Q)
i=l