Towards Robust Visual Tracking for Unmanned Aerial Vehicle with Spatial Attention Aberration Repressed Correlation Filters

Abstract

In recent years, correlation filtering has been widely used in the field of UAV target tracking for its high efficiency and good robustness, even on a common CPU. However, the existing correlation filter-based tracking methods still have major problems when dealing with challenges such as fast moving targets, camera shake, and partial occlusion in UAV scenarios. Furthermore, the lack of reasonable attention mechanism for distortion information as well as background information prevents the limited computational resources from being used for the part of the object most severely affected by interference. In this paper, we propose the spatial attention aberration repressed correlation filter, which models the aberrations, makes full use of the spatial information of aberrations and assigns different attentions to them, and can better cope with these challenges. In addition, we propose a mechanism for the intermittent learning of the global context to balance the efficient use of limited computational resources and cope with various complex scenarios. We also tested the mechanism on challenging UAV benchmarks such as UAVDT and Visdrone2018, and the experiments show that SAARCF has better performance than state-of-the-art trackers.