CMRT09: Object Extraction for 3D City Models, Road Databases and Traffic Monitoring - Concepts, Algorithms, and Evaluation 
similarly to the method described earlier with the exception that 
the initial curve is defined as a circle around the roundabout 
node so that it must be placed inside the island (Fig. 7a). 
(a) (b) 
Figure 6. Island extraction: (a) Expansion evolution result after 
1330 iterations, (b) selected curve (red) and approximated cubic 
spline (green), (c) other curves resulting from iterative curve 
evolution, and (d) eventual result of expansion evolution. 
The diameter of the circle needs to be less than the threshold 
which dictates whether islands are regarded as point or area 
objects in the topographic database. By experiment, it is safer to 
define a circle with a diameter as one-third of this threshold. 
The expansion result is compared with each group of shrinkage 
results separately, and points that are close enough to each other 
are selected. These points are candidates for ellipse fitting. The 
fitting result for a case with the highest number of points is 
more likely to produce a correct result of island extraction (Fig. 
7f). 
(d) «=4606 (e) (f) 
Figure 7. First sequence for island extraction: (a) initial 
successive circles, 3 located outside central island for shrinking 
evolution and 1 inside for expansion curve evolution; (b), (c) 
and (d) results of the shrinking evolution for exterior circles 
from large to small (n denotes iteration number); (e) result of 
iterative expansion evolution to interior circle; (f) final result. 
Extracted central islands are verified using the existing 
information derived from the topographic database. When a 
roundabout appears in the database as an area object, as shown 
in Fig. 3, the diameter of its central island (D1) obtained from 
the extraction process must only differ from that obtained from 
vector data (D2) by a small amount. In an ideal situation, the 
difference (AD) corresponds to the width of the circulating 
roadway (W), i.e. W=AD. In practice, due to the imprecise 
digitization of roundabouts, polygonal vector data do not 
always lie on the middle axis of the circulating roadway, but 
somewhere within its area. Therefore, AD is expected to be 
within the range of 0 to 2W, i.e. 0< AD<2W. 
In the case where a roundabout appears as a point feature, the 
diameter of the extracted central island must fall within a 
predefined range whose highest value is the threshold below 
which a roundabout is regarded as a point object and whose 
lowest value is the minimum possible diameter for a central 
island. 
3.3 The Snake Model for Roundabout reconstruction 
The snake model, or parametric active contour method (Kass et 
al., 1988), used to delineate the roundabout outline is now 
briefly overviewed to provide a basis for further discussion. 
Further details are provided in Ravanbakhsh et al. (2008) and 
Ravanbakhsh (2008). Snakes are especially useful for 
delineating objects that are hard to model with rigid geometric 
primitives. They are thus well suited to modeling roundabouts 
since the borders are of diverse shape with various degrees of 
curvature. Snakes are polygonal curves associated with an 
objective function that combines an image term (external 
energy) and measurement of the image force (e.g. the edge 
strength). There is also a regularization term (internal energy) 
and a minimization of the tension and curvature of the polygon. 
The curve is deformed so as to iteratively optimize the objective 
function. Traditional snakes are sensitive to noise and need 
precise initialization. Since roundabout borders have various 
degrees of curvature, a close initialization cannot often be 
provided. As a result, traditional snakes can easily get stuck in 
an undesirable local minimum. 
To overcome these limitations, the ziplock snake model was 
developed (Neuenschwander et al., 1997). A ziplock snake 
consists of two parts: an active part and a passive part. The 
active part is further divided into two segments, indicated as 
head and tail, respectively (Fig. 8). The active and passive parts 
of the ziplock snake are separated by moving force boundaries. 
Unlike the procedure for a traditional snake, the external force 
derived from the image is turned on only for the active parts. 
Thus, the movement of passive vertices is not affected by any 
image forces. Starting from two short pieces, the active part is 
iteratively optimized to image features, and the force 
boundaries are progressively moved towards the centre of the 
snake. Each time that the force boundaries are moved, the 
passive part is re-interpolated using the position and direction of 
the end vertices of the two active segments. Optimization is 
stopped when force boundaries meet each other. 
Ziplock snakes need far less initialization effort and are less 
affected by the shrinking effect from the internal energy term. 
Furthermore, their computation is more robust because the 
active part, whose energy is minimized, is always quite close to 
the contour being extracted. This modified snake model can 
detect image features even when the initialisation is far away 
from the solution. However, it can still become confused in the 
presence of disturbances. In high resolution aerial images, such