摘要
为解决反无人机过程中常见的背景变换、定位不佳、飞出取景范围和长时跟踪等问题,以SiamCAR为基础,提出结合注意力机制和细化模块的目标跟踪算法SiamGR以及融合目标检测算法的变体SiamGR-FR。首先,在主干网络中加入全局注意力机制(global attention mechanism,GAM)来提升复杂背景下的特征表征能力;然后,使用Alpha-Refine模块对跟踪器输出的结果进一步细化以获得无人机目标更精确的定位;最后针对长时跟踪以及无人机飞出取景范围等情况,融合目标检测算法Faster-RCNN对目标进行重定位并重置跟踪器来显著提升跟踪效果。在DUT Anti-UAV数据集上,SiamGR以及SiamGR-FR算法的成功率和精确率分别达到了61.5%、84.2%和66.2%、94.6%,相比基准算法分别提升了5.3%、3.4%和10%、13.8%,在反无人机场景中优于目前主流的算法。研究结果可为视频反无人机目标跟踪提供参考。
To address the common issues such as background change,poor localization,targets flying out of the frame,and long-term tracking in the anti-drone process,we propose the object tracking algorithm SiamGR,which combines attention mechanism and refinement module based on SiamCAR,as well as its variant SiamGR-FR that integrates object detection algorithms.Firstly,the global attention mechanism(GAM)is added to the backbone network to enhance the feature representation capability in complex backgrounds;then,the Alpha-Refine module is used to further refine the result output by the tracker for more precise localization of the UAV target;finally,in response to the issues of longterm tracking and UAV flying out of the frame,the object detection algorithm Faster-RCNN is integrated to re-localize the target and reset the tracker,significantly improving the tracking performance.On the DUT Anti-UAV dataset,the SiamGR and SiamGR-FR algorithms achieve success rate and precision rate of 61.5%,84.2%and 66.2%,94.6%,which are 5.3%,3.4%,and 10%,13.8%better than the baseline algorithms,outperforming the current mainstream algorithms in Anti-UAV scenarios.The conclusions of this paper can provide a reference for video-based anti-drone object tracking.
作者
丁钰峰
杨志钢
郑滨汐
DING Yufeng;YANG Zhigang;ZHENG Binxi(College of Information and Communication Engineering,Harbin Engineering University,Harbin 150001,China)
出处
《应用科技》
CAS
2024年第5期235-242,共8页
Applied Science and Technology
基金
航空科学基金项目(201801P6002)。
