Multi-feature weighted cross-fusion target tracking algorithm based on kernel correlation filtering

Abstract

The Kernel Correlation Filtering Algorithm (KCF) is one of the popular algorithms in the field of target tracking. It has been widely used in many fields because of its advantages in tracking effect and tracking speed. However, KCF uses a single directional gradient histogram (HOG) as the feature descriptor, when the tracking background is complex, the target moves too fast, the target rotates or is blocked, the tracking frame drifts and even the target is lost. In order to improve the robustness of the tracking effect, a method of weighted cross-fusing the directional gradient histogram and the color features to form a fusion feature as the new feature descriptor is proposed, which can describe the target feature more accurately, so as to achieve target tracking in more complex scenarios. Finally, the paper tests the tracking effects on multiple OTB standard data sets, and compares them with the KCF using a single feature, which shows that the algorithm using the fusion feature has good robustness in multiple complex backgrounds.