56 
In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part ЗА - Saint-Mandé, France. September 1-3, 2010 
radiometric filters are used. The problems due to the variation 
of background are shown, along with the chosen solution. 
In order to effectively perform the detection of vehicles and 
their trajectory, a prediction of their positions is obtained by 
using data previously collected and a Kalman filter. 
2. METHODOLOGY 
2.1 Candidate vehicles detection and labelling 
A base frame is chosen, with no vehicles in the roundabout 
(Figure 1). For every frame (Figure 2), the difference of the 
base image is performed (Figure 3). By applying a non-maxima 
suppression, the pixels with null or negligible radiometric 
absolute differences are black, and those for which variations 
have been detected have non zero values. The following 
operations are then carried out: pixels are grouped through a 
segmentation and some regions are detected, whose edges are 
extracted with classic techniques (Canny, 1986); open lines are 
connected through an expansion of the edges, and a clustering 
is performed by filling the regions surrounded by closed lines; 
an erosion of the edges is then performed, in order to restore the 
original dimensions of the filled regions. A filter is used to 
eliminate the regions having a width less than a selected 
threshold; every obtained area, corresponding to a candidate 
vehicle, is finally labelled (Figure 4). In this way, small noises 
are eliminated, along with the false foregrounds (known as 
ghost objects). A counterpart of this procedure is the 
elimination of some real foregrounds: in figure 4 the region 
corresponding to the red car on the left side of figure 2 is 
Figure 3. Frames difference before non-maxima suppression 
6r 
Figure 4, Detected regions labelling for the frame 78880 
present in figure 3, but it was eliminated by non-maxima 
suppression performed using a high threshold. Kalman filter 
will be useful in this case, as described later. 
Generally, to have the view of a whole roundabout, the camera 
should be positioned in a high place; light poles can be used, 
but in this case oscillations are expectable, the confronted 
frames are translated and rotated, and several “false’' areas will 
be detected (Artese, 2008). 
In order to solve this problem, a registration should be 
performed; automatically detectable targets (Fraser, 1997), or 
known points external to the road area, can be effectively used. 
2.2 Base frame updating 
Background subtraction is used primarily to identify image 
regions that contain foreground information. The ideal 
background model is dynamic and should handle slight 
variations in the background conditions. For this reason, an 
updating must be performed. The changes in the background 
are often not predictable, above all in a partially cloudy day. 
Also in sunny days the changes due to the motion of the 
shadows of poles and buildings are fast. 
Figure 5 shows a base frame obtained only 30 minutes before 
the frame of figure 1. It is possible to observe some big 
differences: the shadow of the building on the left side, the 
shadow of the central pole and the colour of the road. In fact, 
the surface of the road was rapidly drying, and, consequently, 
the radiometric characteristics were changing. If we perform 
the difference between figure 2 and figure 5, we obtain a result 
(Figure 6) very different from Figure 3. In Figure 7 the 
difference between the base frames is shown. It is evident that 
few minutes are sufficient to change the background.