The design of a robust basis for identifying the objects of interest
on the video scenes and evaluating their trajectories along the
time is essential. The aim is to get a high hit rate in these previous
tasks for making easier the subsequent detection of suspicious
behaviors and alarms in the automated video-surveillance application
for shopping malls presented in this paper. Therefore, the
errors in the video pre-processing and human tracking stages
dramatically affect the effectiveness of the final system. Due to
this, these errors must be reduced as much as possible.
The initial video-surveillance phases defined in this work
are the following: background subtraction, blob fusion,
detections-to-tracks association based on Kalman filtering and a
LSAP solution and, finally, an occlusion management based on
visual appearance. The main goal of these algorithms is minimizing
the error propagation between them and providing useful information
about human trajectories to the final alarm processes for
improving the general performance of the system.