Fast Adaptive Spatial Reduction for Discriminative Correlation Filter Tracking

Abstract

Visual object tracking is a challenging problem in which accurate scale and translation estimation is a difficult problem due to various circumstances. An efficient algorithm should correctly estimate its translation and size deformation under varying conditions. In this paper, we extract multiple channel features and measure its discriminative ability by estimating a weight factor for the object under tracking. Besides, it integrates feature reduction and interpolation techniques to improve the performance of the proposed approach. This helps us to increase the search region of the target area which in turn results in an improvement of accuracy and robustness of the proposed algorithm. Segmentation of the target object is also incorporated to learn an optimised correlation filter. The algorithm uses separate filters for translation and scale estimation. Scale estimation approach and channel weight estimation process are algorithm independent methods and hence, it can be incorporated into any tracking method. Extensive experiments on visual object tracking (VOT) were conducted using the proposed approach on bench mark dataset and have shown promising results. Since only two simple standard feature sets were used, Histogram of Gradient (HoG) and Colour attributes, the novel approach can be integrated into real world applications.