摘要
针对传统反无人机集群方法仅聚焦于检测或打击的单阶段研究,以及目标检测漏检、打击点覆盖不足的问题,提出检测-聚类两阶段反无人机集群打击策略.检测阶段输出的高置信度时空坐标为聚类分析提供可靠输入,聚类阶段生成聚类中心作为关键打击点,形成精准检测-聚类打击的闭环协同机制.首先,检测阶段改进YOLOv8算法,针对无人机集群小目标、密集分布特性,在检测头部分引入多尺度卷积注意力设计,融合Backbone网络提取的多尺度特征以提高检测精度,验证集测试检测精度mAP50达95.77%, mAP50-95达59.66%,较原始模型分别提升1.2%和2.4%;其次,聚类阶段提出改进的快速模糊C均值(AFCM)聚类算法,利用中心筛选以及动态隶属度削减策略将收敛效率较传统FCM提升55%;最后,在自建多场景无人机集群数据集测试,帧率30 fps时,检测-聚类全流程耗时低于2 ms,验证了所提出的两阶段反无人机集群策略以及算法的有效性.
In view of the limitations of traditional anti-drone swarm methods that focus solely on either detection or strike in a single stage,as well as the issues of missed detections and insufficient strike coverage,this paper proposes a two-stage detection-clustering anti-drone swarm strike strategy.In the detection stage,high-confidence spatiotemporal coordinates are output to provide reliable input for subsequent clustering analysis,while in the clustering stage,cluster centers are generated as key strike points,forming a closed-loop collaborative mechanism for precise detection and strike.First,the detection stage improves the YOLOv8 algorithm to address the challenges posed by the small target size and dense distribution characteristics of drone swarms.A multi-scale convolutional attention mechanism is introduced in the detection head to integrate the multi-scale features extracted by the backbone network,thereby enhancing detection accuracy.On the validation set,the mAP50 reaches 95.77%and the mAP50-95 reaches 59.66%,representing improvements of 1.2%and 2.4%,respectively,over the original model.Secondly,in the clustering stage,an improved accelerated fuzzy C-means(AFCM)clustering algorithm is proposed,which employs center selection and a dynamic membership reduction strategy to enhance convergence efficiency by 55%compared to the traditional FCM.Finally,tests on a self-built multi-scenario drone swarm dataset demonstrate that,at 30 fps,the full detection-clustering process takes less than 2 ms per frame,thereby validating the effectiveness of the proposed two-stage anti-drone swarm strategy and its underlying algorithms.
作者
唐杰
张鹏年
周鑫
王秉坤
樊浡昊
马捷
TANG Jie;ZHANG Peng-nian;ZHOU Xin;WANG Bing-kun;FAN Bo-hao;MA Jie(Northwest Institute of Mechanical and Electrical Engineering,Xianyang 712099,China;School of Mechanical Engineering and Automation,Northeastern University,Shenyang 110819,China)
出处
《控制与决策》
北大核心
2025年第11期3263-3272,共10页
Control and Decision
