342023RadioEngineeringVol.53No.1doi:10.3969/j.issn.1003-3106.2023.01.005引用格式:任进,李文邦,郭昱汝.基于无人机平台的多目标跟踪算法[J].无线电工程,2023,53(1):34-39.[RENJin,LIWenbang,GUOYuru.Multi-targetTrackingAlgorithmBasedonUAVPlatform[J].RadioEngineering,2023,53(1):34-39.]基于无人机平台的多目标跟踪算法任进,李文邦,郭昱汝(北方工业大学信息学院,北京100144)摘要:针对目前无人机平台多目标跟踪技术的跟踪精确度低、占用内存大的问题,提出了一种基于不同检测器算法和DeepSort算法结合而成的多目标跟踪算法,提高在无人机上对地面行人在跟踪数据集中的效果。使用深度学习的多目标跟踪技术通过构建卷积神经网络(ConvolutionalNeuralNetwork,CNN),用卡尔曼滤波算法实现了对目标轨迹的预测,匈牙利算法则使卡尔曼滤波的预测结果得以分配,使DeepSort算法在保证跟踪效果的同时,也保证了跟踪时的速度。实验结果显示,DeepSort在与YOLOv5x检测器配合后,多目标跟踪精度可提高20%。关键词:无人机平台;多目标跟踪;目标检测;YOLOv5中图分类号:TP391.41文献标志码:A开放科学(资源服务)标识码(OSID):文章编号:1003-3106(2023)01-0034-06Multi-targetTrackingAlgorithmBasedonUAVPlatformRENJin,LIWenbang,GUOYuru(SchoolofInformationScienceandTechnology,NorthChinaUniversityofTechnology,Beijing100144,China)Abstract:Toaddresstheissuesoflowtrackingaccuracyandlargememoryoccupationofcurrentmulti-targettrackingtechnologyonUAVplatform,amulti-targettrackingalgorithmbasedondifferentdetectoralgorithmsandDeepSortalgorithmisproposedtoimprovetheeffectoftrackingdatasetforpedestriansonthegroundonUAV.Themulti-targettrackingtechnologyusingdeeplearningrealizesthepredictionoftargettrajectorywithKalmanfilteralgorithmbyconstructingaConvolutionalNeuralNetwork(CNN).ThepredictionresultsofKalmanfilterisdistributedbytheHungarianalgorithm,sotheDeepSortalgorithmensuresnotonlythetrackingeffect,butalsothetrackingspeed.Theexperimentalresultsshowthatthemulti-targettrackingaccuracycanbeimprovedby20%bythecombinationofDeepSortandYOLOv5xdetector.Keywords:UAVplatform;multi-targettracking;targetdetection;YOLOv5收稿日期:2022-10-0...