Background: Although correlation filtering is one of the most successful visual tracking
frameworks, it is prone to drift caused by several factors such as occlusion, deformation and
Objective: In order to improve the performance of correlation filter-based trackers, this paper proposes
a visual tracking method via online reliability evaluation and feature selection.
Methods: The main contribution of this paper is to introduce three schemes in the framework of correlation
filtering. Firstly, we present an online reliability evaluation to assess the current tracking
result by using the method of adaptive threshold segmentation of response map. Secondly, the proposed
tracker updates the regression model of correlation filter according to the assessment result.
Thirdly, when the tracking result based on a handcrafted feature is not reliable enough, we propose a
feature selection scheme that autonomously replaces a handcrafted feature used in the traditional
correlation filter-based trackers with a deep convolutional feature that can re-capture the target by its
powerful discriminant ability.
Results: On OTB-2013datasets, the Precision rate and Success rate of the proposed tracking algorithm
can reach 84.8% and 62.5%, respectively. Moreover, the tracking speed of proposed algorithm
is 19 frame per second.
Conclusion: The quantitative and qualitative experimental results both demonstrate that the proposed
algorithm performed favorably against nine state-of-the-art algorithms.