第39卷第4期2024年4月Vol.39No.4Apr.2024液晶与显示ChineseJournalofLiquidCrystalsandDisplays基于主干增强和特征重排的反无人机目标跟踪郑滨汐*,杨志钢,丁钰峰(哈尔滨工程大学信息与通信工程学院,黑龙江哈尔滨150001)摘要:视频图像中面向无人机的目标跟踪是反无人机任务中的重要一环。无人机低空飞行背景复杂,同时在视频图像中目标像素占比较小,都给目标跟踪增加了难度。针对以上问题,以SiamRPN++为基础,提出了一种引入改进的主干网络和特征重排的孪生神经网络目标跟踪算法(SiamAU)。首先,在主干网络中加入ECA-Net注意力机制网络,同时对激活函数进行改进,以提升复杂背景下的特征表征能力;然后,对主干网络输出的浅层特征进行浅层降维并与后三层深层特征进行融合,得到更适合无人机等小目标跟踪的改进深度融合特征。在DUTAnti-UAV数据集上,SiamAU算法的成功率和精确率达到了60.5%和88.1%,相比基准算法提升了5.6%和8.1%。在两个公开数据集上的测试结果表明,在反无人机场景中SiamAU算法的跟踪表现优于目前主流的算法。关键词:反无人机;目标跟踪;孪生网络;注意力机制;特征重排中图分类号:TP391文献标识码:Adoi:10.37188/CJLCD.2023-0150Anti-UAVobjecttrackingwithenhancedbackboneandfeaturerearrangementZHENGBinxi*,YANGZhigang,DINGYufeng(CollegeofInformationandCommunicationEngineering,HarbinEngineeringUniversity,Harbin150001,China)Abstract:Objecttrackingfortheunmannedaerialvehicle(UAV)invideosisanimportantpartoftheAnti-UAVtask.Thecomplexbackgroundduringlow-altitudeflightandthesmallimagingsizearetwodifficultiesforUAVobjecttracking.ASiameseneuralnetworkobjecttrackingalgorithm(SiamAU)isproposed,whichisbasedonSiamRPN++incombinationwithanimprovedbackboneandafeaturerearrangementtechnique.Firstly,ECA-Netattentionmoduleisintegratedintothebackbonenetwork,whiletheactivationfunctionisimprovedtoenhancetherepresentationabilityofconvolutionfeaturesincomplexbackground.Then,channelnumberofthelastthreeconvolutionfeaturesisrearrangedinordertomakefulluseoflow-levelfeaturesthatareconduciveforsmallobjecttracking.Therearrangedfeathersarefurtherfusedtoobtaintheimprovedfeaturemap.Finally,OntheDUTAnti-UAVdataset,SiamAUalgorithmachievess...