Changxin Gao1, Feifei Chen1, Jin-Gang Yu1,2, Rui Huang1, Nong Sang1
1National Key Laboratory of Science and Technology on Multispectral Information Processing,
School of Automation, Huazhong University of Science and Technology, Wuhan, 430074, China
2Department of Computer Science and Engineering,
University of Nebraska-Lincoln, Lincoln, NE, 68503, USA
Tracking-by-detection has become an attractive tracking technique, which treats tracking as a category detection problem, and trains a detector to separate the object from the background. While most tracking-by-detection methods take the task in tracking as searching for an object category as which in detection, we argue that the task in tracking should be searching a specific object instance. Based on this viewpoint, a novel tracking framework based on specific object exemplar detectors is proposed for visual tracking. To build a quite specific and discriminative model to separate the object exemplar from background, the proposed method trains an exemplar-based linear discriminant analysis (ELDA) classifier for the object exemplar, using a single object instance and massive negatives. To improve its adaptivity, we use an ensemble of ELDA detectors and update them during the tracking, to cover the variations. Extensive experimental results show that, even on a large scale dataset, the proposed method outperforms the other state-of-the-art trackers, which demonstrate the effectiveness of our tracking algorithm.
Matlab code (zip, github). (off-line background model)
result locations on 51 videos, Videos results .
An Enhancement version was proposed with CNN Features and Adaptive Model Update (project).
Related paper:
_______________________
If you have any questions, please contact Changxin Gao, cgao at hust dot edu dot cn