Title: Tracking via Classification


Abstract:

The talk will be focused on classifiers fusion for target tracking. More in detail, 
an ensemble of classifiers can be exploited to build on-the-fly a model to represent 
the appearance of an object. The combination of experts can be constantly updated, to 
learn variations in the appearance of the target through the frames providing a means 
for detection and tracking.

The problem of how multiple classifiers can be fused together, that is the choice of 
the rule to merge their outputs to reach a final consensus, will be discussed. Three 
criteria to improve the performance of a classifier ensemble will be proposed; to test 
them, the classification problem as a means to get a better target tracking performance 
on real-world multimedia video sequences will be examined.