In: Paparoditis N., Pierrot-Deseilligny M.. Mallet C.. Tournaire O. (Eds). IAPRS. Vol. XXXVIII. Part 3A - Saint-Mandé, France. September 1-3. 2010 
Figure 5. First Base Frame 
Figure 6. Frames difference 
Difference between base frames 
1 tt; ijp 
PpHfeK 
To evaluate the performances of the techniques for background 
detection, some information retrieval measurements are 
generally used, based on the number of pixels correctly detected 
by the algorithm. For this aims, “ground-trutlf’ frames are 
obtained, by manually highlighting the foreground. 
It can be observed that, when the described methods are directly 
used for background updating and foreground detection, some 
drawbacks are present: 
all background algorithms are sensitive to 
environmental noises; 
algorithms that adapt more slowly (MF, KF, RGA, 
MoG) have better performance than those that adapt 
quickly, but, in case of sudden changes, produce 
“ghost objects”; 
For RGA and MoG, a rapid variation in global 
illumination, after a long stationary period, can turn 
the entire frame into foreground, due to the very small 
variances of the background components. 
In our case, two operations are carried out for the current frame 
processing: after the Frame Difference, both morphological 
operation and non maxima suppression are performed. To 
update the base frame the following criterion has been adopted: 
the pixels corresponding to the black ones after the current 
frame processing (i.e. the pixels not belonging to the candidate 
vehicles), are used to update the base frame. With reference to 
the described example, the pixels of figure 2 corresponding to 
the white ones of figure 4 are used to update the corresponding 
ones of figure 1. In fact, figure 1 is not a real frame, but it is the 
frame of figure 5 after the updating. 
2.3 Vehicle recognition and tracking 
The blobs obtained and labelled are used to recognize the 
entering vehicles and those already present in the previous 
frames, to determine their trajectories. The How chart of the 
algorithm used to build the compatibility table, along with some 
examples, is shown in (Artese, 2008). The adopted strategy is 
the following: 
- in the first frame the coordinates of the moving vehicles are 
known; these vehicles have been opportunely labelled; 
- for every region detected in the second frame, a vector is built, 
which elements are the coordinates of the barycentre, both 
orthogonal (pixel) and polar with reference to the centre of the 
roundabout, along with other characteristics (area, average RGB 
and HSI values, bounding box); 
- every vehicle of the first frame is confronted with every 
region of the second frame and there are obtained the distance, 
the march direction (clockwise or counter-clockwise) and the 
ratio between the areas; 
- a table is obtained in which, for every vehicle, the regions of 
the second frame compatible for trajectory and dimension are 
reported. The radiometric characteristics are then compared and 
a figure of merit is obtained for every couple vehicle-region. 
- by using the table, the number of the corresponding vehicle 
should be assigned to every region; in case of incoming 
vehicles, the number of the last vehicle increased by one will be 
assigned. 
In several cases, the correspondence between vehicles and blobs 
is not trivial. If we exclude the entering or outgoing vehicles, 
and the case of compatibility with only a region, several 
possibilities should be investigated. By comparing figures 3 and 
4, we can observe that the coach has been divided into two 
regions due to the light pole, while the blue car in the middle of 
the frame has been divided into three blobs. 
In this case, the sum of the areas should be compared with the 
vehicle having compatible barycentre. 
The obtained coordinates of the vehicles allow to obtain both 
trajectory and velocity, once known the time interval between 
consecutive frames. 
2.4 The use of Kalman filter 
The prediction of the barycentres of the vehicles can be very 
useful to improve the results of the comparisons above 
described. For this aims, the use of Kalman filter (Kalman, 
R.E., 1960) has been foreseen. 
The Kalman filter addresses the general problem of trying to 
estimate the state of a discrete-time controlled process that is 
governed by the linear stochastic difference equation: