network which shows the robustness of the proposed algo-
rithm against noise clusters.
The cost calculation procedure for the suggested line
segment in Figure 7 by selecting different ORness values is
depicted in Table 4. It is obvious from the Table 4 that the cost
value considering the average
OWA
value has become lower
compared to
WLC
in Figure 7. However, the value of cost is ac-
ceptable regarding to the selected cost threshold (T
cost
= 1).
T
able
4. C
ost
C
alculation
for
the
L
ine
S
egment
D
efined
in
F
igure
7 S
electing
D
ifferent
OR
ness
V
alues
ORness
0 0.1 0.3 0.5 o.7 0.9 1
OWA 0.020 0.027 0.260 0.460 0.660 0.790 0.840
Ave (OWA)= 0.41
Cost (Ave (OWA))= 0.077
Figure 9a demonstrates how the
OWA
values were modified
by selecting different decision strategies for ten different line
segments in the simulated road image of Figure 8. Regard-
ing Figure 9a, the higher the ORness value is, the higher is
the
OWA
value. The increasing rate is different for each line
segment. The average range of
OWA
values for these ten line
segments are depicted in Figure 9b.
Experiment with Imagery
The road centerline vectorization algorithm proposed in this pa-
per can be used on any binary road image regardless of the road
detection methodology applied. In the following experiments,
the detected binary road image achieved by the combination
of fuzzy C-means clustering (Ameri
et al
. 2008), along with the
hole-filling and area-based filtering methods (Chaudhuri
et al
.,
2012). To verify the performance of the proposed vectorization
algorithm on the real road images, several experiments with
high resolution images were carried out. For different input
binary road images T
cost
=1 and
R=
0.7
d
were considered. The
only parameter that needs to be defined for each image is
K
.
This parameter is selected according to the maximum number
of possible road legs in the places of road junction on the image.
For all data sets the K value was considered as K = 4, and the
cost function was calculated using the average
OWA
value.
Figure 10a presents a pan-sharpened Ikonos image of
Hobart City of Tasmania in an urban area (750 × 650 pixels),
and Figure 10b illustrates the detected binary road image. The
proposed road key point connection algorithm was applied
on the road binary image, and the vectorization results were
shown in Figure 10c. Besides, the results were compared with
Minimum Spanning Tree (
MST
) algorithm for road centerline
extraction (Ameri
et al
., 2008), which its outcomes is present-
ed in Figure 10d. As expected the proposed algorithm vector-
ized the road centerlines more accurately and completely
comparing with the
MST
(see Figure 11).
Figure 12 shows the same procedure on the Ikonos image
of Hobart in a non-urban area (700 × 700 pixels). Figure12c
and 12d illustrate the results of the centerline vectorization
using the proposed methodology and
MST
, respectively, on
the binary road image (Figure 12b). The proposed algorithm
extracts most of the roads perfectly with a few false alarms.
However, the vectorized road centerline by means of
MST
con-
tains several false extractions specifically at place of nearby
parallel roads in the middle of the image (Figure 12d).
As another experiment, Figure 13a depicts a pan-sharp-
ened Ikonos image of Shiraz, Iran. The proposed vectorization
methodology was implemented on the detected road image
(Figure 13b) and the extracted road centerlines are shown
in Figure 13c. There are some missed connections (depicted
by dotted lines) and false extraction (depicted by light solid
lines) where the road detection methodology has failed. On
the whole, the number of false alarms in
MST
results (Figure
13d) are more prominent than the proposed method.
(a)
(b)
Figure 9. Results for ten different line segments of Figure 8: (a) OWA values based on different decision strategies, and (b) average range
of OWA values.
Table 5. Comparison of the Proposed Method with MST for Both Hobart Images
Quality measure
Urban Hobart
Non-urban Hobart
Shiraz
Proposed method MST method Proposed method MST method Proposed method MST method
RMSE (m)
0.97
1.04
1.1
1.3
0.67
0.73
Completeness (%)
95
89
98
94
88
79
Correctness (%)
96
94
97
93
93
86
Junction RMSE (m)
5.24
5.31
3.11
3.35
7.95
8.23
114
February 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
