Multi-feature Integration with Adaptive Learning Based Correlation Filter for Visual Object Tracking

Abstract

Correlation Filter (CF) based tracking is the most imperative part of computer vision and offers many potential benefits. To get maximum benefits, object trackers need to provide better accuracy in presence of visually challenging scenarios with less computational burden. Therefore, this research aims to develop a robust object tracker to deal with target variations in a real-time environment. At first, the multi-feature descriptor is implemented using the feature fusion technique which combines the response of Histogram of gradient (HOG), saliency, gray level intensities, and Color Naming (CN) features. Afterward, an adaptive learning strategy is integrated by utilizing the Peak-to-Sidelobe Ratio (PSR) to evaluate correlation peaks. The quality of the proposed methodology is validated on challenging datasets. Tracking results reveal that the proposed scheme outperforms the other advanced CF trackers with Distant Precision (DP) scores of 88.2%, 85.9%, and 74.1 % over OTB2013, OTB2015, and TempleColor128 datasets respectively.