3. Istanbul 2004
(GIS road axis]
A e
segment )
4
dere S
end of!
EEE pe
ds
oH
eight variations
5— d
| curvatures
ya:
UEM
raluation
and gaps: con-
low
low
-structures and
—can be mod-
'n road objects,
oncepts "Lane
road network.
ality measures
nents) defining
Model compo-
ce of an object
' evaluating its
1, model com-
lependent from
grouped mark-
sed to create a
ght column are
| to predefined
The fuzzy ag-
indicating the
of all involved
gartner, 2002).
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
4 EXTRACTION AND FUSION OF ROAD OBJECTS
4.1 Extraction of Road Objects — Overview
The extraction strategy inheres knowledge about how and when
certain parts of the road and context model are optimally ex-
ploited, thereby being the basic control mechanism of the ex-
traction process. It is subdivided into three levels (see Fig. 3):
Context-based data analysis comprises the segmentation of the
scene into the urban, rural, and forest area and the analysis of con-
text relations. While road extraction in forest areas seems hardly
possible without using additional sensors, e.g., infrared or LI-
DAR sensors, the extraction in rural areas may be performed with
the system of (Baumgartner et al., 1999). In urban areas, extrac-
tion of salient roads includes the detection of homogeneous rib-
bons in coarse scale, collinear grouping thin bright lines, i.e. road
markings, and the construction of lane segments from groups of
road markings, road sides, and detected vehicles. The lane seg-
ments are further grouped into lanes, road segments, and roads.
During road network completion, finally, gaps in the extraction
are iteratively closed by hypothesizing and verifying connections
between previously extracted roads. Similar to (Wiedemann and
Ebner, 2000), local as well as global criteria exploiting the net-
work characteristics are used. Figures 4 and 5 illustrate interme-
diate steps of extraction and Figs. 6 and 7 show typical results.
For details regarding the extraction we refer the reader to (Hinz
et al., 2001, Hinz and Baumgartner, 2002). The system described
there extracts roads from a single image and uses a DSM and
views from other images to circumvent occlusions. In contrast,
the new version extracts roads from all available images and fuses
them in object space. The next section focuses on this particular
issue.
Segmentation of
context regions
Context-
based da
hased data | urban areas il forest ] [ rural areas ]
analvsts
Analysis of con
text relations
Road extraction
Extraction
of salient Construction of
roads lane segments
using approach
for rural areas
c AT
| Fusion i
J
Road
network
completion
Road Network
completion
Figure 3: Extraction Strategy.
Detected shadow reas Markings outside of shadow area
Figure 4: Examples of intermediate steps during road extraction.
4.2 Fusion of Road Objects
To exploit information from multiple views, an appropriate fu-
sion strategy has been developed, which is especially suitable for
complex environments like urban areas. It can be characterized
by following features: 1) It is based on objects, i.e., parts of the
road network such as lane segments and road segments 2) It is
carried out in object space 3) It is embedded in the system’s con-
cept of self-diagnostic extraction algorithms. From a method-
ological viewpoint, the novelty of this approach mainly relates to
the incorporation and use of self-diagnosis algorithms for fusion.
The first two points, however, accommodate the special proper-
ties of urban scenes and are thus of no minor importance. In the
following comments on each point are given:
Ad 1) Fusion is based on objects because, as mentioned above,
aerial images of urban areas show very high complexity. If fusion
would be based on low level image primitives like raw gray val-
ues or edge structures, either an extremely accurate DSM must
be given (effectively a 3D city model) or the fusion algorithm has
to cope with many ambiguities and many conflicting hypothe-
ses that occur when matching primitives over different images.
Hence, our philosophy is to stay in 2D as long as possible and
to extract objects of large extent and high semantics in each im-
age separately. Matching such kind of objects over images is
much easier and the requirements on a DSM can be relaxed sig-
nificantly. In the case of our road extraction system, the ob-
jects which are subject for fusion are lane segments extracted in
each available image. These are constructed in previous process-
ing steps from groups of markings (i.e., thin bright lines) and
(anti-)parallel road sides (i.e., grayvalue edges) while constrain-
ing them to enclose a homogeneous region or alternatively a ve-
hicle (Hinz et al., 2001, Schlosser et al., 2003).
Ad 2) The main reason for performing fusion in object space is
its natural way in treating each image with equal importance and
not preferring any image a priori. Thus a dependence of the fi-
nal results on the processing order of the images can be avoided.
As a side effect, objects extracted in images of different resolu-
tions may be combined easily and all necessary parameters can
be passed in real-world values.
Ad 3) The fusion algorithm is embedded in the system’s concept
of self-diagnostic extraction algorithms. The idea behind this ap-
proach is that each module used during extraction should attach
its result with a confidence value indicating the quality how well
the job has been done. Our approach to define evaluation crite-
ria from which the confidence values can be calculated is to split’
up the components of the underlying object models into two dif-
ferent types. Model components of the first type are used for
extracting an instance of an object and the components of the
second type serve as criteria for evaluating the quality of the ex-
tracted instance. For guaranteeing an unbiased evaluation, model
components belonging to different types should be independent
from each other. In order to evaluate a certain object, pre-defined
fuzzy functions are used. Since the road model underlying our
Markings inside of shadow areca Verified fanes (top), detected car (bottom)
351
