) segments in reality
atical sense). This is
threshold values for
oints and differences
; (see Figure 7).
attributes of features
ributes recorded for
| description of our
define the relation
lations
fa
P
a At P
Lr à NO
ia b ib i At
a ib
pi D
à EN
Puis At fb b
ia
ib
Junction
Figure 8. Attributes of relations - continued
4. ROAD RECOGNITION
Our recognition strategy is simply based on three
assumptions:
e roads have in most cases antiparallel edges (Nevatia &
Babu, 1980);
e roads are elongated objects;
e roads are connected in a network.
In our experience other criteria, such as looking for a
homogeneous texture between the road edges, failed rather
often.
We proceed now with the analysis of the relations of
antiparallism. Perceptual grouping may be applied
recursively, generating different levels of virtual features,
that is, arrangements of the basic primitives. Here the
primitives are pairs of antiparallel segments, grouped in
higher level primitives (would-be road legs) based on a
continuity criterium, namely collinearity and co-
curvilinearity (see Figure 9).
fe
Figure 9. Grouping antiparallel pairs of segments
At the end of this stage we have groups of pairs of
antiparallel lines; the same line may nevertheless belong to
more than one group: this originates a competition among
the groups, which will be solved later on. Looking for a
network structure, a last recursive grouping is performed
between would-be road legs connected by junctions.
Ideally, if there would not be competitions between groups,
this process would end up with all (would-be) disjoint road
networks in the image. Figure 10 depicts the behaviour the
algorithms on an artificial example.
Would-be road- Groups Ambiguities
legs of AP pairs
I a-b 2
a-c
II f-g 1
1-m
Il b-d 4
c-e
IV p-q 2
V d-h 3
e-i
ep
VI m-n 2
VII n-o 1
VIII q-r 1
Junction relations found between would-be legs.
"|
E
E.
V+ VII
Le
II
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996
it
a
I+V
=
VII + VII
=
IV
