Abstract
Discriminative correlation filtering (DCF)-based trackers have demonstrated remarkable results in the field of visual tracking. Nevertheless, in most DCF-based trackers: (i) they apply correlation operations across the entire search area features without discriminative weights, rendering them highly susceptible to background interference; and (ii) the fixed aspect ratio scale search strategy is inadequate for accurate scale estimation under irregular scale variations. To overcome these challenges, this study proposes a new correlation filter through saliency-driven localization and cascaded scale estimation (SDCS-CF). Specifically, a U-like network is devised to generate a pixel-wise saliency map. This map is then multiplied element-wise with search area features to accentuate the target attribute, mitigating distractors and increasing localization confidence. Furthermore, a cascaded scale estimation model consisting of three one-dimensional filters is designed to refine the scale estimation process and improving the robustness. Extensive experimental results on three public datasets demonstrate that the proposed SDCS-CF outperforms most DCF-based trackers.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call