GASiam: Graph Attention Based Siamese Tracker for Infrared Anti-UAV

Unmanned aerial vehicle (UAV) is widely used in all walks of life, which not only brings convenience but also poses a threat to security. Although infrared UAV tracking has many advantages, it still faces many challenges such as target rapid movement, scale variation, etc. In order to realize the accurate tracking of UAV targets, this paper proposes a graph attention based Siamese tracker, called GASiam. Specifically, aiming at the loss caused by occlusion or rapid movement of UAV targets, SiamR-CNN is used as the baseline because of its three-stage redetection mechanism. Aiming at the lack of semantic information of infrared UAV targets, a graph attention module (GATM) is designed to enhance the information embedding ability of the local tracking network. This module can reduce the influence of background while retaining more foreground information, and improve the ability of the tracker to face the challenges of cluttered background and scale variation. Finally, a switching strategy is designed to select local tracking and global redetection adaptively according to the target state, in order to realize the robust and stable tracking of the infrared UAV target. Experiments on the IR-anti-UAV dataset show that GASiam has achieved leading performance in the field of anti-UAV.