International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
value of the road segment in the junction area. In addition, an
angular constraint is applied forcing the angle of the two
reconstructed roads at the reconstructed junction point similar
to that defined by the corresponding VEC25 roads at the
VEC25 junction point. Furthermore, in the angular constraint
higher priority is given to higher class roads or continuous
roads. In such formulation, the junction point is correctly
located, while the shape of the roads at the junction point is
well formed: complying with the reality and following the road
design rules.
With the extracted roads and road junctions, the road network is
obtained. The results inherit other attributes from the VEC25
data with the road lengths updated, and road widths appended.
The number of lanes can be inferred from the known road width
and also possibly roadmarks. The 3D information permits the
derivation of other useful attributes like horizontal and vertical
curvatures and road slope.
4.6 Performance evaluation
We introduce two types of measure for self-diagnosis of the
extraction results: an overall quality measure for the whole
road, and measures for the road segments. If a result does not
pass the overall quality test, a further test is conducted to find in
which segments the errors occur. The overall quality of the
extraction result can be obtained from the following criteria:
e the lengths of the extraction result and the VEC25 road
should be similar
e the shape difference between the extraction and the VEC25
road should be small
e total length of PRSPs should cover a large part of the
extraction result
We also define an internal quality measure for each road
segment using the shape similarity measure between the
segment and the corresponding VEC25 road, and the
information in the classification result and the nDSM in the
segment area (Zhang, 2003b). For higher class roads, the
assessment of the reliability of the extraction results is also
conducted through comparison of the results by using parallel
edges and roadmarks.
External evaluation is done by comparing the extracted results
with precise reference data. The quality measures aim at
assessing completeness and correctness as well as geometrical
accuracy (Heipke et al, 1998; Zhang, 2003b). Additional
measures have been defined, e.g. to evaluate the shape quality,
but have not been used in the test results presented here.
5. RESULTS
The described system has been implemented as a stand-alone
software package with a graphic user interface running both on
UNIX SGI and PC Windows XP. The system imports color
stereo or ortho-imagery, the existing road database and other
input data, and outputs the extracted roads in 3D Shapefile
format that is readily imported by existing GIS software. Other
data formats can be easily accommodated. The system has been
tested using more than 20 models in various landscapes. Some
reports of the system performance can be found in Zhang and
Baltsavias (2002) and Zhang (2003b). A benchmark test has
been conducted independently by our project partner using new
flight imagery, in the test site Thun, Switzerland. The terrain
height ranges from 550 m to 2200 m. Almost all road types in
Switzerland can be found in this area. The images were
acquired in October 2001, and the image data have the same
specifications as described in Section 1. During the test, our
system is only applied to extract roads in rural areas, while
roads in urban and forest areas are not processed. Fig. |
presents a portion of 3D road extraction and road network
generation (only the left image is shown). The landscape of Fig,
I includes open rural, forest areas and small settlements.
Figure 1. Extracted 3D roads and road network in test site Thun,
Switzerland.
The details of automatic 3D road extraction and junction
generation in rural areas with varying complexity are presented
in Fig. 2, where the VEC25 roads are shown in white lines and
the extracted roads in black lines. Note that the road junctions
are also well extracted and modeled.
Recently, the system has been under extensive test by our
project partner. Around 10,000 km of roads have been
processed. Almost all roads in rural areas were correctly
extracted using visual check.
Fig. 2. Details of road extraction and junction generation with
varying complexity in test sites in Switzerland. The
extracted roads are shown in black lines and the VEC25
roads in white lines.
Table 1 summarises the external evaluation of the extraction
results in Fig. 1 using the reference data measured by L+T at an
analytical plotter. It can be seen that our system achieves very
good results. The completeness and correctness are very high.
The accuracy of the extracted road network is about 0.5 m both
in planimetry and height, fulfilling the accuracy requirements of
the project ATOMI.
Quality Measures
Completeness 04.2%
Correctness 96.9%
Length of reference (km) 12.4
Length of extraction (km) 11.69
dx 0.40
RMS (m) dy 0.33
dz 0.58
Table 1. Quality measures for the test dataset in Fig.l.
1056
Int
A
sp!
oct
ext
inv
ext
the
We
Na
arc
DS
got
ma
ach
res,
The
wh
mat
and
par
hav
ster
pro
acc
Dut
Mai
vari
sem
colc
1:25
situ:
sout
The
the
imag
qual
shov
are
obse
very
occli
very
are (
RMS
natic
Neth
othei
used
are fi
have
Durii
Fig.
areas
4, the
datas
Refer
The ;
each
extra
datab
i$ ca
deper
perfo;
