Adopting Feature-Weighted and Scale Adaptive for Real-Time Visual Tracking

Abstract

Considering poor accuracy and robustness of single-feature object tracking, researchers have proposed numerous multiple-feature integration for tracking, including deep features and traditional hand-crafted features. In this paper, we propose a feature-weighted method of Histogram of Oriented Gradient (HOG) and color names (CN) based on correlation filters. Both features are adopted independently to obtain detection response maps, which are subsequently normalized. Finally, Euclidean distance between normalized map and desired output, which is activated by a negative exponential function, is regarded as the weight of the corresponding response map. Although using feature-weighted correlation filters for tracking has achieved great improvement in accuracy and robustness, many conventional trackers have not solved the problem of scale variation and the current mainstream scale estimation methods have existed the deficiencies of computational redundancy and fixed scale-factor. In view of this problem, firstly adopting three scale-factor to roughly judge the direction of target scale variation, and then the optimal scale-factor is acquired cyclically in this direction. The methods we proposed based on ECO-HC are tested in OTB2013, OTB100 and TC128 datasets and the results demonstrate the effectiveness of our proposed methods.