1
11
i
j
“|
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Figure 5: Lanes extracted in different images (a,b) and their superimposition (c) calculated using DSM (d: grayvalue coded DSM)
system is designed in a hierarchical way (roads consist of lanes
which again consist of markings and road sides, etc.), the confi-
dence measure of each object are used in three different ways:
a) Confidence propagation: Confidence values of lower level ob-
jects (e.g., groups of markings and road sides) are combined us-
ing the principles of fuzzy-set theory and propagated to the next
level of the model hierarchy (e.g. lanes).
b) Autonomous evaluation: According to our model for self-
diagnosis, at each level, object knowledge not used for extrac-
tion or evaluation at lower levels is incorporated, e.g., each lane
should have a parallel counterpart (one lane roads are not con-
sidered). Note that this evaluation is independent of propagated
confidence values (therefore autonomous”).
¢) Consistency check: The score of autonomous evaluation of a
higher level object are used to test the consistency of lower level
objects. Consider, for instance, a hypothesis of a two-lane road
segment (i.e., the higher level object) of which the first lane is ex-
tracted correctly but the other one is extracted only in fragments,
e.g., due to inhomogeneities of the pavement. The latter lane
hypothesis has consequently a low rating through autonomous
evaluation, however, from the higher level point of view, there is
strong evidence that this particular hypothesis is correct. Hence,
such a hypothesis would pass the consistency check and is kept
for further processing. In general, this means for the implementa-
tion that a hypothesis—regardless of its autonomous evaluation—
is kept as long as the next level in the model hierarchy is com-
pletely processed and evaluated.
Implementation: This concept is also applied and implemented
for fusion of road information from multiple views. Lanes are
extracted in each image separately (see Fig. 5) and projected on
a fairly accurate DSM (grid size and accuracy ca. 2m). In case
of overlapping lanes, the lane having the best (propagated) con-
fidence value is selected first and its mutual overlap with other
lanes 1s computed. The score for autonomous evaluation of such
a lane is calculated from the overlap ratios of lanes extracted in
other images including weights for their deviation in position and
direction. After deleting redundant parts of lanes the lane with the
second highest confidence value is selected, and so forth. Thus a
unique set of fused lanes is achieved. In the next hierarchy level,
road segments are constructed from the fused lanes, i.e., parallel
and collinear lanes are merged. Note, that the individual lanes of
a road segment may be fragmented as long as a parallel lane pro-
vides a connection from one lane fragment to another fragment.
The average degree of fragmentation of a road segment serves as
consistency check for the fused lanes, i.e., lanes are rejected if
not enough evidence is given for grouping them into larger road
objects.
>
Tests with less accurate DSMs have shown that the use of lanes
as objects to be fused may lead to matching ambiguities. Hence,
an alternative version of the system (Hinz, 2003) uses the object
"road segment"—Aan object with more semantics (see the model
hierarchy in Fig. 1 a)—for fusion.
352
ir
5 EVALUATION OF THE RESULTS AND DISCUSSION
Figures 6 and 7 illustrate the final result of road extraction in two
parts of the Zurich Hoengg dataset (Baltsavias et al., 2001). As
can be seen, major parts of the road networks have been extracted
in spite ofthe high complexity ofthe scenes. The system is able to
detect shadowed road sections or road sections with rather dense
traffic (see e.g. Fig. 7 a and b). The results have been evaluated
by matching the extracted road axes to manually plotted reference
data. Table 1 summarizes the numerical values according to the
definition of (Wiedemann, 2003). As can be seen, we achieve
a completeness of more than 75 % and a correctness of about
95 % regarding the extracted road axes that could be linked into a
network. Also the evaluation of the network characteristics yields
satisfying results since for all evaluation criteria (detour/shortcut
factor, topological completeness, topological correctness) values
close to the optimum are reached.
| Evaluation criteria || Data setI: | Data set IL: |
Completeness [%] 76.6 81.6
Correctness [%] 98.8 95.0
RMS-Error [m] 1.3 25
Mean detour factor [ ] 1.04 1.05
Mean shortcut factor [ ] 0.95 0.95
Topological completeness [96] 100.0 84.0
Topological correctness [%] 96.2 100.0
Table 1: External Evaluation of extracted road axes.
However, it must be noted that some of the lane segments have
been missed or have been linked incorrectly (Fig. 7 b). This is
most evident at complex road junctions and crossings in both im-
age parts, where only spurious features for the construction of
lanes have been extracted. Another obvious failure can be seen at
the right branch of the junction in the central part of Data Set II
(Fig. 7 a). The tram and trucks in the center of the road have been
missed since our vehicle detection module is only able to extract
vehicles similar to passenger cars. Thus, this particular road axis
has been shifted to the lower part of the road where the imple-
mented parts of the model fit much better. As a consequence, the
RMS-value drops down from acceptable 1.3m in Data Set I to
poor 2.5m in Data Set II. The interested reader may be referred to
the much more exhaustive evaluation carried out in (Hinz, 2003).
In summary, the results indicate that the presented system extracts
roads even in complex environments. An obvious deficiency ex-
ists in form of the missing detection capability for vehicle types
as busses and trucks. However, the main bottleneck of our sys-
tem is the (still) weak model for complex junctions. Hence, be-
sides the aforementioned improvement of verifying connection
hypotheses, one of our next steps will be directed towards the
modelling and reliable detection of road junctions. As a final re-
mark regarding the percentages of correctness and completeness
we would like to mention that, in spite of the definitely encour-
aging results, it would be unfair to disregard the fact that these
percentages can be achieved only due to the expertise of the sys-
tem developers in setting the parameters correctly (as it is surely
Internati
Figure 6
extractic
true for ¢
this fielc
there are
so that z
operatio
6 |
In the le
those ob
tions. W
monitor
data sim
is desigr
trast to r
upon loc
6.1 Ca
To mod
wirefran
ometric
region.
during e
