Anchor-free Object Tracking Algorithm Combining RFB and Dual Attention

Main Article Content

Yucheng Wang

Abstract

In view that the Siamese trackers could not capture the long-range dependence and vulnerable to problems similar to objective factors such as interference of background, this paper proposes a lightweight dual attention module, it can be increased under the condition of less amount of calculation to efficiently capture the attention of space dimension and the channel dimension. Meanwhile, we introduced the RFB module to strengthen the feature.
Specifically, our method is mainly divided into two steps in the feature fusion stage: (i) Use the RFB module to strengthen the single-scale template features and search area features. (ii) Use the dual attention module to make the network focus on the distinguishing and robust features of the target adaptively to eliminate similar targets' interference. In the dual attention module, this study uses cascaded two Criss-Cross attention modules to model
spatial dimension attention. Extensive experiments on OTB2015 and LaSOT datasets show that this object tracking algorithm embedded with the RFB module and the lightweight dual attention module not only achieves good performance compared to the most advanced tracking algorithms and still runs at a real-time speed of 45FPS.

Article Details

How to Cite
Yucheng Wang. (2021). Anchor-free Object Tracking Algorithm Combining RFB and Dual Attention. CONVERTER, 2021(6), 89-101. Retrieved from https://converter-magazine.info/index.php/converter/article/view/373
Section
Articles