联合图卷积和聚类的红外无人机集群多目标跟踪算法  

A Multi-target Tracking Algorithm Based on Graph Convolution and Clustering for Infrared UAV Swarm

作　　者：李琦 席建祥 杨小冈 卢瑞涛 谢学立 LI Qi;XI Jianxiang;YANG Xiaogang;LU Ruitao;XIE Xueli(College of Missile Engineering Rocket Force University of Engineering,Xi'an 710000 China)

机构地区：[1]火箭军工程大学导弹工程学院,西安710000

出　　处：《电光与控制》2025年第3期15-20,共6页Electronics Optics & Control

基　　金：国家自然科学基金(62176263,62276274);陕西省杰出青年科学基金(2021JC-35)。

摘　　要：针对红外无人机集群多目标跟踪场景中个体间外观特征稀少,且同质化严重、集群内个体相互遮挡、平台晃动等挑战问题,提出了一种基于图卷积神经网络(GCN)与聚类算法的融合跟踪算法。首先,引入自注意力特征掩码以增强GCN对轨迹聚合的效果;其次,结合交并比(IoU)和可能性C均值聚类,以增强对运动特征的提取和集群内相邻目标的区分能力;最后,采用轨迹连接模型和高斯平滑插值算法对跟踪结果进行进一步优化。所提算法融合了短时轨迹聚合和长时轨迹匹配的能力,仅利用运动信息和交互信息就能实现红外无人机集群多目标跟踪。在红外无人机集群多目标跟踪数据集上进行实验,结果表明:与其他先进跟踪算法相比,所提跟踪算法具有更高的性能指标,MOTA与IDF1分别达到84.9%与80.2%;在目标相互遮挡、平台晃动等复杂场景下也具有优越的跟踪效果。In the scenario of multi-target tracking of infrared UAV swarm there are challenges such as limited appearance features and severe target homogeneity mutual occlusion of individuals in the swarm and platform jittering.To address the problems this paper proposes a fusion tracking algorithm based on Graph Convolutional neural Network(GCN)and clustering algorithms.Firstly a self-attention feature mask is introduced to enhance GCN s trajectory aggregation.Secondly IoU and likelihood-based C-means clustering are adopted to improve motion feature extraction and adjacent target differentiation in the swarm.Finally optimization of the tracking results is further achieved through a trajectory connection model and a Gaussian smoothing interpolation algorithm.The proposed algorithm integrates short-time trajectory aggregation and long-time trajectory matching and achieves infrared UAV swarm multi-target tracking by only using motion information and interaction information.The experiments are conducted on infrared UAV swarm multi-target tracking dataset.The experimental results demonstrate its superior performance compared with that of other advanced tracking algorithms.MOTA and IDF1 of the proposed algorithm reach 84.9%and 80.2%respectively and it has excellent tracking effects even in complex scenarios such as mutual occlusion of targets and platform jittering.

关 键 词：无人机集群 红外目标跟踪 图卷积神经网络 时空联合约束 轨迹片段聚合 

分 类 号：TN911.73[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]