面向反制无人机集群的多目标连续鲁棒跟踪算法  

Multi-object continuous robust tracking algorithm for anti-UAV swarm

作　　者：王传云[1] 苏阳 王琳霖[1] 王田 王静静 高骞 WANG Chuanyun;SU Yang;WANG Linlin;WANG Tian;WANG Jingjing;GAO Qian(College of Artificial Intelligence,Shenyang Aerospace University,Shenyang 110136,China;College of Computer Science,Shenyang Aerospace University,Shenyang 110136,China;Institute of Artificial Intelligence,Beihang University,Beijing 100191,China;China Academy of Electronics and Information Technology,CTEC,Beijing 100041,China)

机构地区：[1]沈阳航空航天大学人工智能学院,沈阳110136 [2]沈阳航空航天大学计算机学院,沈阳110136 [3]北京航空航天大学人工智能研究院,北京100191 [4]中国电子科技集团公司电子科学研究院,北京100041

出　　处：《航空学报》2024年第7期256-269,共14页Acta Aeronautica et Astronautica Sinica

基　　金：国家自然科学基金(61703287,61972016);辽宁省教育厅科学研究项目(LJKZ0218,LJKMZ20220556);沈阳市中青年科技创新人才项目(RC210401);沈阳航空航天大学引进人才科研启动基金(22YB03)。

摘　　要：无人机(UAV)集群作战正朝着智能化、实战化迅猛发展,将在未来战场上造成巨大威胁,面向反制无人机集群的探测与跟踪研究势在必行。针对在复杂场景及远距离探测条件下无人机集群目标之间相互遮挡、无人机为弱小目标等原因造成的检测精度降低和跟踪精度降低问题,本文提出的无人机集群多目标(UAVS-MOT)连续鲁棒跟踪算法可以有效解决。UAVS-MOT模型基于FairMOT模型的多分支无锚框预测结构,将坐标注意力模块与DLA-34网络相结合,构建了全新的主干特征提取网络以提升特征信息的表达能力。此外,引入全新的ArcFace Loss损失函数进行训练以提高模型的收敛速度,并利用BYTE数据关联方法以降低目标漏检率和提高轨迹的连贯性。实验表明,本文提出的UAVS-MOT多目标跟踪算法在UAVSwarm Dataset上的多目标跟踪准确度(MOTA)和目标识别准确度(IDF1)分别为73.4%与76.1%,相比原有FairMOT算法分别提升5.7%与2.9%,可以解决目标的漏检、误检和跟踪精度低的问题,鲁棒性好。Unmanned Aerial Vehicle(UAV)swarm warfare is rapidly developing towards intelligence and practicality,which will pose a huge threat on the future battlefield.Therefore,research on detection and tracking of anti-UAV swarm is imperative.To solve the problems of reduced detection and tracking accuracy caused by mutual occlusion between UAV swarm and weak targets in complex scenes and long-distance detection conditions,this paper proposes a UAVS-Multiple Object Tracking(UAVS-MOT)multi-object continuous robust tracking algorithm.The UAVS-MOT model is based on the multi-branch anchor-free frame prediction structure of the FairMOT model.The coordinate atten⁃tion module is combined with the DLA-34 network to construct a new backbone feature extraction network,so as to im⁃prove the expression ability of feature information.In addition,the new ArcFace Loss function is introduced for training to improve the convergence rate of the model,and the BYTE data association method is used to reduce the target miss rate and improve the track consistency.The experiment shows that the Multiple Object Tracking Accuracy(MOTA)and Identity F1 Score(IDF1)of the proposed algorithm on the UAVSwarm Dataset are 73.4%and 76.1%,respectively,which are 5.7%and 2.9%higher than those of the original FairMOT algorithm.The proposed method can solve the problems of missed and false detection of targets and low tracking accuracy,and has good robustness.

关 键 词：多目标跟踪 无人机集群 FairMOT 注意力机制 反制 

分 类 号：V279[航空宇航科学与技术—飞行器设计] TP391.41[自动化与计算机技术—计算机应用技术]