Attending to the evaluation, the most effective background subtraction
approach for the surveillance scenario studied in this
paper is the proposed by Odobez and Yao (2007). It yields high
detection rates even in challenging conditions as low illumination
or pixelation, which are very common in several low-cost
video-camera pre-installed systems in shopping malls. However,
the method suggested by Odobez and Yao (2007) is an expensive
algorithm in computational cost terms, and its real-time performance
in a multi-camera implementation only can be deployed
in systems with a very high processing capacity, which is not our
goal. For low computational cost solutions capable of working in
real-time, the method proposed by Zivkovic (2004) is a reasonable
alternative and it is selected for our expert video-surveillance
application, because it also provides effective results (which are
improved in the tracking stage) and allows managing more cameras
in real-time.
As a final task for this stage, the segmented images obtained
with background subtraction are filtered applying the dilation
() and erosion () iterations described in Eq. (1), where S is the
segmented image, F is the resulting filtered image and K is a
3 3 kernel with the anchor at its center. The objective of this
operation is, firstly, filling spaces in granulated foreground objects