the number of
r-threshold and
s determined to
| a unsupervised
| gives a better
ing data set is
ed training data
ass number and
'ervised method
are known. An
adequate when.
' each object is
ted using an
remote sensing
ine detector are
road segments,
ave only found
s two methods
ased screening.
and one for
ween each line
nspection. The
5 lateral offset
udinal gap. It
any of these
air is rejected.
hecks contrast
is whiter than
n background,
connected and
3. Istänbul 2004
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
3. RESULT & DISCUSSIONS
31 Line Feature Extraction Result
According to Steger, the relationship between o and the road
width is described as follows.
oz w/ A3 (2)
Table 1 shows how many line are detected when o is fixed.
Seeding threshold and linking threshold are determined where
results seem best for the human eye. Correctness constantly
increases as Gaussian Kernel Parameter s become bigger. It
suggests that if the kernel is too small, it picks up much noise
and results in many false road segments.
Completeness, on the other hand, does not decrease constantly.
It has a peak around s=1.8. Considering relationship between
Gaussian Parameter and line width, this result suggests that
width of dominating line segments in the image is less than 6
pixels (=15m in ground resolution). As many of roads in the
scene are dual lane road and each lane is about 4-6m width, the
result indicates s=1.8 may be best starting points for road centre
extraction.
Table 2 shows another result of the centre line detector.
Gaussian Kernel Parameter is fixed at 1,80. Change of link
threshold has little effect while increase of seed threshold
slightly improves correctness. This result implies that seeding
threshold of 4.0 — 5.0 and linking threshold of 1.0-1.2 is best for
reliable road segment extraction.
3.2 False Line Elimination Result
Table 3 shows the grey scale threshold result by ATC.
Gaussian Kernel Parameter, seceding threshold and linking
threshold are set to 1.80, 5.0 and 0.5 respectively in all cases.
ATC separates grey level histogram into 4 classes. Range of
each class is, 22-30, 31-55, 56-88, 89-197. Then grey level of
roads that are within the buffer are picked up. At last, the total
length of picked up GIS roads is calculated. The column
"Correctly extracted road segments" indicates how many of
"extracted road segments" is parallel to *GIS road".
Then completeness can be defined as “correctly extracted GIS
road” / “GIS road” and indicate how complete is the extracted
road. Correctness can be defined as “correctly extracted road
segment” / “extracted road segment” and indicates how the
Seed Link Num. of | Correct- | Complete-
threshold threshold Line ness (%) ness (%)
3.0 0.8 1464 10.8 56.9
4.0 0.8 1244 12.0 61.2
5.0 0.8 1120 12.3 59.4
6.0 0.8 1018 12:1 56.5
7.0 0.8 917 13.1 56.4
5.0 0.4 1119 ]1.9 59.1
3.0 0.6 1119 11.9 59.1
5.0 0.8 1120 11.9 59.4
5.0 1.0 1121 11.9 60.1
5.0 12 1119 11.9 60.3
5.0 1.4 1118 12.0 59.9
5.0 1.6 1113 11.9 59.]
Table 2. Result of centre line detector. o is set to 1.8 in all cases.
result is reliable.
Correctness is improved if dark and / or bright threshold are set.
It suggests that introduction of multilevel threshold is
reasonable to eliminate false road segments. However,
completeness slightly decline if dark threshold was applied.
This is because the dark threshold cuts out shadow region on
roads. Shadows of buildings have elongated shapes like roads
and the centre line detector recognizes these shadows as line
segments. Fig.2 shows the effect of darker region filtering.
Though some false line segments that go through building
shadow are efficiently removed by the filtering, shadows cast
on a road are also removed.
As both correctness and completeness are dissatisfying low in
a 30 and 89 are used for dark and bright thresholds. grey scale threshold, utilising multi-spectral images sensor is
69.1 The column “GIS road” means the total length of true road tested. ee.
: contained in the scene. This data was created manually on the The unsupervised classification tool in ERDAS was used for
34.7 GIS software, ArcView. The column “Extracted road segment” the classification. Figure 2 shows a classification result. Six
58.2 means the total road length obtained by the centre line detector. categories are obtained. Red, green, deep blue, light bluc and
47.3 The column “Correctly extracted GIS road” indicates how grey (dark & light) represent buildings, vegetation, road, road
56.1 many “GIS road” is extracted from the line segments obtained X building and shadows respectively.
58.4 by the method. It is calculated as follows. At first 2.0 pixels
58.6 buffer is created around the extracted road segments. Then GIS
52.9 | ; : :
= Threshold (dark) 0 30 0 30 3.3 Line Grouping Result
48.4 Threshold (bright) 235 233 89 89 In the rule based screening section, two characters of
52.5 | Num. Of lines 1173 1023 1137 984 line pairs were checked to judge connectivity of the
43.2 a. GIS road (pixel) 3724.9 3724.9 3724.9 3724.9 pairs.
46.5 b. Extracted road 22649.6 | 20105.6 | 21399.0 | 18859.3 For testing geometric information, three properties
were inquired based on thresholds suggested in the
2291.1 2181.3 original paper. Maximum angular difference between
lines concerned and the line connecting the gap is set to
5 degree. Maximum Transverse gap, which represents
how much the pair of lines is off to the side, is set to 3
pixels. Maximum Longitudinal gap, which represents
how far both end points are without offset (transverse
gap), is set to 1.5 times the length of shorter segment.
For testing photometric information, the author
introduced "Contrast reversal test" in the original
L| segment (pixel)
c. Correctly extracted | 2284.1 2170.9
GIS road (pixel)
d. Correctly extracted | 2614.9
road segment (pixel)
Correctness (b/d) 11.5% 12.3% 12.1% 12.9%
Completeness (a/c) 61.3% 58.3% 61.7% 58.6%
Table 3. Grey scale threshold result. Masking brighter area effectively
reduces false line segments while masking darker arca
ılidity of each 2469.0 | 2580.9 | 2435.0
on's deviation
slightly lessens completeness.
405
