Multi-scale Kernel Correlation Filters for Visual Tracking with Fusion of multi features and multi templates

Although the correlation filter (CF) based trackers have currently achieved brilliant results in terms of both accuracy and robustness, they are not capable of effectively handle the scale variation. To tackle this problem and solve the drifting issue, we propose a CF based tracker by selecting and fusion of multiple scales, multiple features, and multiple templates. Firstly, to deal with the problems of the fixed template size, a set of possible scales is considered to estimate the scale of a target object. Secondly, in order to relieve the drifting issue, a set of candidate templates, which are affected by significant appearance changes is carefully selected and learned as filter templates to jointly capture the target appearance variation. Finally, the HOG and the color-naming features are integrated to improve the overall tracking performance. The experiments are done on CVPR2013 dataset. The proposed tracker successfully tracked the target objects in experimented sequences and performed well in terms of accuracy and robustness through the state-of-the-art trackers.

Keywords:

Visual Tracking , Correlation Filter , Candidate Templates , Scale Variation , Drifting Problem