Abstract
The rapid growth of drones presents potential threats to public security and personal privacy, and it is vital to effectively detect the intruding drones. Prior work on visual-based drone detection using convolutional networks regards the drone detection task as a regression problem on a large set of human-defined components, i.e., proposals and anchors. These components bring a huge number of predictions to be selected, which pose the challenges to drone detection. In this paper, we propose a Deformable DETR-based drone detector with visual transformer, which eliminates the human-defined components to pursue high-accuracy detection performance. Specifically, to detect remote drones at a higher accuracy, the resolution of the features in backbone is enhanced. Meanwhile, two data augmentation methods including illumination jittering and multi-blurring are developed to cope with the time-varying illumination and the changeable weather, based on which the environmental robustness of the proposed detector is thus maintained. The field experiments are carried out, and it is demonstrated that a higher detection accuracy is achieved for the proposed drone detector.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call