International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
2. GENERAL STRATEGY
The developed system makes full use of available information
about the scene and contains a set of image analysis tools. The
management of different information and the selection of image
analysis tools are controlled by a knowledge-based system. In
this section, a brief description of our strategy is given. We refer
to Zhang (2003b) for more details. The initial knowledge base
is established by the information extracted from the existing
spatial data and road design rules. This information is formed in
object-oriented multiple object layers, i.e. roads are divided into
various subclasses according to road type, land cover and
terrain relief. It provides a global description of road network
topology, and the local geometry for a road subclass. Therefore,
we avoid developing a general road model; instead a specific
model can be assigned to each road. This model provides the
initial 2D location of a road in the scene, as well as road
attributes, such as road class, presence of roadmarks, and
possible geometry. A road is processed with an appropriate
method corresponding to its model, certain features and cues are
extracted from images, and roads are derived by a proper
combination of cues. The knowledge base is then automatically
updated and refined using information gained from previous
extraction of roads. The processing proceeds from the easiest
subclasses to the most difficult ones. Since neither 2D nor 3D
procedures alone are sufficient to solve the problem of road
extraction, we make the transition from 2D image space to 3D
object space as early as possible, and extract the road network
with the mutual interaction between features of these spaces.
3. CUE EXTRACTION
When a road from VEC25 is selected, the system focuses on the
image regions around the road, defined using the position of the
road and the maximal error of VEC25. Then, according to the
road attributes a set of image processing tools is activated to
extract features and cues. 3D straight edge generation is a
crucial component of our procedure because the road sides are
among them. The 3D information of straight edges is
determined from the correspondences of edge segments between
stereo images. An image classification method is implemented
to find road regions. With the DSM and DTM data, the above-
ground objects and ground objects are separated. We also
exploit additional cues such as roadmarks to support road
extraction.
3.1 3D straight edge extraction
The edges are extracted by the Canny operator in stereo images.
For each edge, the edge attributes are computed, including the
geometrical description of the edge and the photometrical
information in the flanking regions of the edge. The epipolar
constraint is applied to reduce the search space. We then
compute a similarity measure for an edge pair by comparing the
edge attributes. The locally consistent matching is then
determined through structural matching with probability
relaxation using the similarity measures as prior information.
We refer to Zhang and Baltsavias (2000) for the detailed
matching strategy and qualitative performance evaluation. The
matched edges are then transformed to object space by finding
the corresponding pixels within each matched edge pair (Zhang
and Baltsavias, 2002). Finally, 3D straight edge segments are
fitted to the 3D edge pixels.
3.2 Image classification for road region detection
We implemented the ISODATA algorithm to classify the color
images and separate road regions from other objects. The
success of image classification also depends on the input data,
The original RGB color image is transformed into different
color spaces and is also used to compute several artificial
bands/indices to enhance features such as vegetation and
shadows so that they are more isolated in feature space. The
following 3 bands are selected for image classification: (1) the
first component of principal component transformed image, (2)
a band calculated with R and G bands in RGB space as (G-
R)/(G+R), (3) S band from HSI color space. We then determine
5 classes corresponding to road regions, green objects, shadow
areas, dark roofs and red roofs.
3.3 DSM and DTM analysis
The DTM or DSM has been used in our system to reduce search
space for straight edge matching. They are also used to verify if
a 3D straight edge or a region is on the ground. Because a DSM
ideally models the man-made objects as well as the terrain,
subtracting the DTM from DSM results in the so-called
normalized DSM (nDSM) which enables the separation of
above-ground objects (buildings and trees) and ground objects
(roads, etc.). Since in ATOMI, the DTM data is not very
precise, we extract above-ground objects directly from the DSM
using a multiple height bin method presented in Baltsavias et al.
(1995). By combining the information of nDSM with image
classification data, our system creates redundancy to confirm
the existence of roads. Furthermore, it can partly compensate
the missing and wrong information in the classification.
3.4 Roadmark and zebra crossing extraction
Roadmarks and zebra crossings are good indications of the
existence of main roads and roads in urban areas. Both of them
have distinct color (usually white or yellow). Usually roadmarks
give the road direction and often the road centerline, while the
zebra crossings define the local road width. Thus, they can be
used to guide the road extraction process or verify the extraction
results. In addition, in many cases the correct road centerlines
can be even derived directly from present roadmarks and/or
zebra crossings. This is especially useful when the road sides
are occluded or not well-defined, such as in cities or city
centers.
The roadmarks are detected as white straight lines using an
image line model in which the shape of an image line is
presented as a second order polynomial (Zhang, 2003a). The
extracted straight lines are then transformed into object space by
our developed structural matching method. Only those that are
on the ground (as defined by nDSM), belonging to the road
region (as determined by the classification) and in the buffer
defined by VEC25 are kept as detected roadmarks. Zebra
crossings are composed of several thin stripes. Using color
information, the image is first segmented. Morphological
closing is applied to bridge the gaps between zebra stripes. We
then obtain several clusters by connected labeling. Only the
clusters with a certain size are kept, while the small ones are
discarded. Then, the shape of the cluster is analyzed. The
rectangle-like clusters are selected as zebra crossings. The
center, the short and long axes of the detected zebra crossings
are computed using spatial moments.
1054
In
co
rei
an
Or
in
E
in
sic
rex
set
fir
us
ex
me
4.]
Th
Or
as
are
ch
if
mi
bei
tw
ed;
mt
mu
cla
exí
col
ha
Po
iS |
COI
4.2
No
des
in
De
mo
roa
cla
mis
pro
roa
(RS
knc
me.
img
reli
usii
PR
of]