Assembling similar tracking approaches in order to strengthen performance

In this paper we present a novel ensemble of two similar tracking approaches, which independently present good performance for different video sequences. We propose that by combining the response of these tracking approaches, we can strengthen their detecting capability and therefore increase the tracking performance of the ensemble. The Tracking-Learning-Detection (TLD) and the LocalTLD are the approaches we chose for building our ensemble. Our main motivation for assembling these two approaches is that both approaches focus on particular instances of an object and also manage different object representation, for instance, the TLD works reasonably well for planar rigid objects due to the global classifier it includes, meanwhile the LocalTLD focuses on invariant local features and is able to overcome the planar assumption. Combining these approaches, we are able to take advantage of their best qualities and overcome their biggest problems. For introducing our method, we first need to review the principal components of the two chosen approaches, and then we finally introduce the ensemble. The proposed ensemble is compared against results of the independent approaches using a data set of 10 video sequences, showing, in general, a significant improvement. © 2014 Springer International Publishing.

Assembling similar tracking approaches in order to strengthen performance Chapter in Scopus