Exploiting the Anisotropy of Correlation Filter Learning for Visual Tracking

摘要

Correlation filtering based tracking model has received significant attention and achieved great success in terms of both tracking accuracy and computational complexity. However, due to the limitation of the loss function, current correlation filtering paradigm could not reliably respond to the abrupt appearance changes of the target object. This study focuses on improving the robustness of the correlation filter learning. An anisotropy of the filter response is observed and analyzed for the correlation filtering based tracking model, through which the overfitting issue of previous methods is alleviated. Three sparsity related loss functions are proposed to exploit the anisotropy, leading to three implementations of visual trackers, correspondingly resulting in improved overall tracking performance. A large number of experiments are conducted and these experimental results demonstrate that the proposed approach greatly improves the robustness of the learned correlation filter. The proposed trackers performs comparably against state-of-the-art tracking methods on four latest standard tracking benchmark datasets.