In. Stilla U, Rottensteiner F, Paparoditis N (Eds) CMRT09. IAPRS, Vol. XXXVIII, Part 3/W4 — Paris, France, 3-4 September, 2009
An extracted road border is assumed to be correct if the
maximum distance between the extracted road border and its
corresponding reference does not exceed the buffer width.
ijÜÜi
%
(a)
(b)
M
H
(c) (d)
Figure 11. Capture of roundabout outline: (a) initial snakes in
black and road arms in white, (b) and (c) evolving curves, and
(d) reconstructed roundabout.
Figure 12. Sample roundabout extraction results for scenes with
varying degrees of complexity including disturbances.
A smaller value of the buffer width can be chosen for an
application that requires more accurate extraction results. A
reference road border is assumed to be matched if the maximum
deviation from the extracted object is within the buffer width.
Based on these assumptions, three quality measures were
adopted, the first being completeness, which is the ratio of the
number of matched reference road borders to the number of
reference objects. The second is correctness, which is the ratio
of the number of correctly extracted road borders to the number
of extracted objects, while the third is geometric accuracy,
which is expressed by the average distance between the
correctly extracted road border and the corresponding reference
border, expressed as a Root Mean Square (RMS) value falling
within the range of [0, buffer width].
Road border extraction results computed with different buffer
width values are shown in Table 1. The completeness of the
road border extraction increased as the buffer width value
increased from 0.5m to 3m, implying that the results are more
complete for higher buffer width values. The geometric
accuracy increase is inversely proportional to buffer width so
that results obtained with a value of 0.5m are more accurate
than those obtained with a larger buffer width. For the buffer
width value 0.5 m, the completeness is rather low. The reason is
that a slight deviation of the extraction results from the true
boundaries exceeding the buffer width frequently occurs due to
disturbances and sometimes also due to road markings.
Buffer width (m)
0.5
1
2
3
Number of road borders
41
41
41
41
Completeness
53%
62%
74%
85%
Geometric accuracy (m)
0.30
0.38
0.50
0.58
Table 1. Evaluation results for road borders.
As seen in Table 2, a favourable evaluation result was achieved
in the extraction of central islands, which proved the robustness
of the proposed method. Central islands of roundabouts were
extracted with high values for completeness and correctness for
the buffer width of 0.5m, implying the effectiveness of the
proposed hybrid evolution strategy. For the buffer width value
1 m, all of central islands were extracted correctly.
Buffer width (m)
0.5
1
Number of central islands
10
10
Completeness
90%
100%
Correctness
90%
100%
Geometric accuracy (m)
0.26
0.35
Table 2. Evaluation results for central islands.
5. CONCLUDING REMARKS
A new snake-based approach to automatic extraction of road
roundabouts has been described and analysed. Under the
approach, the snake’s external force field is modified based on
the shape of the central island to delineate the roundabout
border. The modified snake force field can overcome various
disturbances inside and outside the central island. It was shown
that the use of prior-knowledge derived from an existing
topographic database can considerably enhance the extraction
performance. Furthermore, a level set approach with a hybrid
evolution strategy was proposed to extract central islands. This
produced good results in all 10 test cases, as central islands
were extracted correctly for an assigned buffer width of lm.
Nevertheless, partial occlusion of the central island border by
large trees and shadowing cannot be overcome at this stage
(Fig. 13). There are several possibilities to further enhance the
results obtained so far and to be able to deal with more complex
scenes. The incorporation of high-level prior knowledge about
