Surrounding-Aware Correlation Filter for UAV Tracking with Selective Spatial Regularization
Abstract
The great advance of visual object tracking has provided unmanned aerial vehicle (UAV) with intriguing capability for various practical applications. With promising performance and efficiency, discriminative correlation filter-based trackers have drawn great attention and undergone remarkable progress. However, background interference and boundary effect remain two thorny problems. In this paper, a surrounding-aware tracker with selective spatial regularization (SASR) is presented. SASR tracker extracts surrounding samples according to the size and shape of the object in order to utilize context and maintain the integrality of the object. Additionally, a selective spatial regularizer is introduced to address boundary effect. Central coefficients in the filter are evenly regularized to preserve valid information from the object. While the others are penalized according to their spatial location. Under the framework of SASR tracker, surrounding information and selective spatial regularization prove to be complementary to each other, which actually did not draw much attention before. They managed to improve not only the robustness against various distractions in the surrounding but also the flexibility to catch up with frequent appearance change of the object. Qualitative evaluation and quantitative experiments on challenging UAV tracking sequences have shown that SASR tracker has performed favorably against 23 state-of-the-art trackers.
This paper corresponds to the research project I’ve done in Vision4Robotics Group.
Visual tracking means to estimate the trajectory of an arbitrary object in a video, given the bounding box of the object in the first frame. I reviewed the development of the state-of-the-art algorithms, and was soon attracted by the elegant mathematical formula and superlative performance of “Correlation Filter-based methods”. Although I quickly had a new idea, things did not go out smoothly as the framework I designed did not work at first. Nevertheless, I dived deep into the formulas for an answer. Through days of analysis, inside the matrices and vectors, I found an implicit incompatibility of two modules in my framework. After solving it, rich experiments on UAV tracking videos demonstrated the good tracking ability of my framework.
