Robust visual tracking via co-trained Kernelized correlation filters. (September 2017)

• Record Type: Journal Article
• Authors: Zhang, Le; Suganthan, Ponnuthurai Nagaratnam
• Is Part Of: Pattern recognition. Volume 69(2017:Sep.)
• Page Start: 82
• Page End: 93
• Publication Date: 2017-09
• Subjects: Visual tracking -- High speed -- Kernelized correlation filter -- Ensemble tracking -- KCF tracker -- COKCF tracker -- Correlation filter; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.04.004
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: We train a pool of discriminative classifiers jointly in a closed-form fashion for visual tracking. We propose analytic model for datasets of thousands of translated patches. It is able to outperform the baseline by a larger margin. Abstract: Recent advances in visual tracking have witnessed the importance of discriminative classifiers tasked with distinguishing the target from the background. However, a single classifier may fail to cope with complex surrounding environment and large appearance variations of the target. Motivated by multi-view learning, we equip a basic framework to train a pool of discriminative classifiers jointly in a closed-form fashion in this paper. It poses an extra regularization term in ridge regression which interacts with other base models in the ensemble. Through a simple realization of this approach, we show co-trained kernelized correlation filters (COKCF) which consist of two KCF trackers, are able to outperform the KCF tracker by a larger margin and perform favorably against other state-of-the-art trackers on 63 benchmark video sequences.