Tracking by classifications

Download
2013
Marpuç, Tuğhan
Novel online classifiers are examined and their applications on 2D visual tracking are analyzed in this study. Recently, an emerging class of methods, namely tracking by classification or tracking by detection, achieved quite promising results on challenging tracking data sets. These techniques train a classifier in an online manner during tracking to separate the object from its background. These methods only take input location of the object and a random feature pool; then, a classifier bootstraps itself by using the current tracker state and extracted positive and negative samples. In this thesis, several of these online classifiers are analyzed and a novel tracking system is proposed. A novel feature selection method is introduced to increase the discriminative power of the classifier and a Monte Carlo experiment is setup to observe the performance of the proposed technique. As a final step, a Hidden Markov Model (HMM) is utilized to filter the features that improve the performance during tracking. Moreover, a state of the proposed HMM is allocated to handle occlusions. The proposed tracker is tested on publicly available challenging video sequences and superior tracking results are achieved in real-time.