Robust Feature Points Correspondences for Visual Object Tracking

Abstract

Matching visual appearances of target sample reservoir over consecutive image frames is the most critical issue in sequence-based object tracking. Recent literatures show the effectiveness of the utilization of local feature points set instead of any global feature vectors of patches. A traditional tracking-by-detection framework without taking advantages of geometric information, however, ignores more or less the potential contributions of feature points. This paper proposes a totally novel tracking-by-correspondences framework, a generative approach via an adaptively-selected robust appearance model, a one-step orient motion model based on points correspondences, an automatic scale determination and a clustered online updating target sample reservoir. Extensive experiments validate the accuracy and robustness of the proposed method, and demonstrate the improved performance has been competitive enough to surpass the state of this art.

@inproceedings{yu2013robust,
  author = {Yu, Ning},
  title = {Robust Feature Points Correspondences for Visual Object Tracking},
  booktitle = {Huazhong University of Science and Technology (HUST) Undergraduate Thesis},
  year = {2013}
}