CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation
Through the use of shape description parameters such as
curvature computed from the snake vertices, another force can
be added to the GVF force field. This is the so-called balloon
force, which lets the contour have a more dynamic behaviour
(Cohen, 1991), thereby addressing the two described problems.
This new force, which makes the contour act like a balloon,
applies an inflating effect to the contour to localize the concave
part of the roundabout outline:
F = k\n(s) (16)
where n(s) is the normal unitary vector of the curve at point
F(.v) and k\ is the amplitude of the force. The combination of
the GVF force field and the balloon force modifies Eq. 15 to the
form
V [,] = (K + r /)"' * ir V l '-' ] - K G(u,v) v | iH] n(s)) (17)
The balloon force is activated when the snake’s passive and
active parts are approximately straight, i.e. their overall
curvature, which is defined as the sum of the absolute
curvatures along the curve, is below a threshold. It is applied
only on the passive part of the curve. This is regarded as lying
outside the roundabout’s border, whereas the snake at the active
parts is assumed to be on the right track. The direction in which
the balloon force is applied is towards the roundabout central
area. However, in order to be able to delineate the roundabout
outline, the balloon force has to be applied in two different
directions, central island inwards and outwards (Fig. 9a).
The answer to the question of when and in which direction the
balloon force needs to be applied differs for different samples.
As a result, several parameters need to be tuned on an ad hoc
basis to address this question, which is not a desirable
requirement. To resolve this, the external force field of the
snake approach described so far is modified based on the shape
of the central island. As the shape of the roundabout outline
corresponds to the shape of the enlarged central island, the
island is enlarged to an extent depending on the width of the
circulating roadway (Fig. 9b). Subsequently the snake external
force field is modified based on the enlarged central island. The
external force field in the enlarged central island is replaced
with the GVF of an intensity-step image (Fig. 9c) whose main
characteristic is that its external force points directly from the
centre outwards so that snakes situated in this area are drawn
toward the outline of the roundabout.
The intensity-step image is generated from a signed distance
function. To generate this function, the border of the enlarged
central island is taken as the reference (Fig. 9b). Successive
concentric layers at a specific distance interval from the
reference to the centre point are then defined. Conversely,
proportional to the distance of each layer to the reference, an
intensity value is calculated and assigned to the respective
layer, i.e. layers closer to the reference curve are brighter and
vice versa.
The obtained intensity-step image has a gradual increase of
intensity values from the centre point towards the reference
curve. Consequently, its GVF field points directly outward. The
modified force field pulls the snakes toward the outline even if
the initialization is far away from true borders. Furthermore,
with this modified force field, problems created by the presence
of various kinds of disturbances such as trees and vehicles
within and outside the central island are overcome. An example
illustrating the improved result using the proposed modified
force field is shown in Fig. 10. The complete reconstruction of a
roundabout using the proposed modified snake model is shown
in Fig. 11, along with intermediate results.
(a) (b) (c)
Figure 9. (a) Two directions in which the balloon force is
applied; (b) reference for the signed distance function (white
curve) computation and concentric regions (black curves); (c)
intensity-step image from the signed distance function.
(a) (b) (c)
Figure 10. The effect of the modified external force field: (a)
intersection lines (black) from initial snakes, (b) results from
unmodified GVF, and (c) improved results with modified GVF.
4. RESULTS AND EVALUATION
The proposed approach was tested using 0.1m GSD
panchromatic aerial orthoimagery covering rural and suburban
areas. The Authoritative Topographic Cartographic Information
System of Germany (ATKIS), which nominally corresponds to
a mapping scale of 1:25,000, was used as the source of external
vector data. Roads are modelled as linear objects in ATKIS.
Tests were conducted on 10 roundabouts. Sample results that
highlight the capabilities of the proposed approach are shown in
Fig. 12, where is can be seen that the method can deal with a
variety of disturbances inside and outside the central island.
Also, most of the roundabout borders were captured correctly.
However, in areas where the curvature of the outline was too
high, as is the case in the top-left example (lower border) and
top-right image (right border), roundabout borders were
extracted with some deviation.
In order to evaluate the performance of the approach, the
extracted roundabout areas were compared to the manually
plotted roundabouts used as reference data. The comparison
was carried out by matching the extracted road borders
resulting from the connection of the roundabout to its associated
road arms to the reference data using the so-called buffer
method (Heipke et al. 1998). Although the buffer width can be
defined using the required accuracy of ATKIS, which for a road
object is defined as 3m, it was decided to set the buffer width
within the range of 0.5 m to 3 m, i.e. 5 pixels to 30 pixels, in
concert with the image resolution of 0.1 m. This allowed
assessment of the relevance of the approach for practical
applications that demand varying degrees of accuracy.
24