Abstract
Most of existing correlation filter-based tracking approaches only estimatesimple axis-aligned bounding boxes, and very few of them is capable ofrecovering the underlying similarity transformation. To a large extent, suchlimitation restricts the applications of such trackers for a wide range ofscenarios. In this paper, we propose a novel correlation filter-based trackerwith robust estimation of similarity transformation on the large displacementsto tackle this challenging problem. In order to efficiently search in such alarge 4-DoF space in real-time, we formulate the problem into two 2-DoFsub-problems and apply an efficient Block Coordinates Descent solver tooptimize the estimation result. Specifically, we employ an efficient phasecorrelation scheme to deal with both scale and rotation changes simultaneouslyin log-polar coordinates. Moreover, a fast variant of correlation filter isused to predict the translational motion individually. Our experimental resultsdemonstrate that the proposed tracker achieves very promising predictionperformance compared with the state-of-the-art visual object tracking methodswhile still retaining the advantages of efficiency and simplicity inconventional correlation filter-based tracking methods.