基于高分辨率孪生网络的无人机目标跟踪算法
UAV target tracking algorithm based on high resolution siamese network
王殿伟 1张池 1房杰 1许志杰2
作者信息
- 1. 西安邮电大学 通信与信息工程学院,西安 710121
- 2. 哈德斯菲尔德大学 计算机与工程学院,哈德斯菲尔德HD13DH
- 折叠
摘要
针对无人机(UAV)目标跟踪任务中目标尺寸小、尺度变化明显和视点改变频繁等问题,本文提出一种基于高分辨率孪生网络的无人机目标跟踪算法.首先,利用改进高分辨率网络作为特征提取主干网络,并且采用动态多模板策略挖掘视频的帧间信息;然后,构建多帧特征融合模块,得到利于目标定位的融合特征;最后,选取无锚框策略定位目标位置,得到精确的跟踪结果.实验结果表明:本文算法在DTB70数据集测试的成功率和准确率分别为66.0%和84.7%,在UAV123数据集测试的成功率和准确率分别为65.7%和84.3%,有效地提升了目标跟踪性能.
Abstract
Unmanned Aerial Vehicle(UAV)target tracking tasks are often suffer from small target size,large scale variance and frequent viewpoint change.To address these issues,in this paper,an UAV target tracking algorithm based on high-resolution siamese network is proposed.Firstly,a high-resolution network is improved as the feature extraction backbone network(Lite-high resolution network,L-HRNet),and a dynamic multi-template strategy is used to mine the inter-frame information of the video.Secondly,a multi-frame feature fusion module is constructed to obtain fusion features that are beneficial to target localization.Finally,an anchor-free strategy is selected to locate the target position and obtain accurate tracking results.The experimental results show that the success rate and accuracy of the proposed algorithm are 66.0%and 84.7%on DTB70 dataset,65.7%and 84.3%on UAV123 dataset respectively,which improves the target tracking performance effectively.
关键词
计算机视觉/目标跟踪/无人机/高分辨率孪生网络/多帧特征Key words
computer vision/target tracking/unmanned aerial vehicle/high-resolution siamese network/multi-frame feature引用本文复制引用
基金项目
国家自然科学基金青年基金(62201454)
西安邮电大学研究生创新基金(CXJJLY2021053)
出版年
2024
NETL
NSTL
万方数据