融合遮挡信息的改进DDETR无人机目标检测算法.pdfVIP

2362024，60（1）ComputerEngineeringandApplications计算机工程与应用

融合遮挡信息的改进DDETR无人机目标检测算法

12，31

周建亭，宣士斌，王婷

1.广西民族大学电子信息学院，南宁530006

2.广西民族大学人工智能学院，南宁530006

3.广西混杂计算与集成电路设计分析重点实验室，南宁530006

摘要：针对无人机航拍图像中目标场景复杂、小目标多、遮挡严重的问题，提出了一种融合目标遮挡信息的改进

DDETR（deformableDETR）的无人机目标检测算法。提出模型用SwinTransformer代替DDETR模型中残差网络

来获得更丰富的多层次语义特征；增加DDETR模型对低层次特征的使用来提高对中小目标的检测效果；利用提出

的遮挡程度估计模块来辅助模型解决遮挡问题，使模型能更好地检测出遮挡严重的目标。在VisDrone数据集上达

到32.3%的平均准确度均值（meanaverageprecision，AP），比标准DDETR模型AP值提高了3.3个百分点，与主流无

人机航拍图像目标检测方法相比，达到了当前先进水平。

关键词：无人机目标检测；深度学习；交叉注意力；可变形卷积

文献标志码：A中图分类号：TP391doi：10.3778/j.issn.1002-8331.2208-0025

ImprovedDDETRUAVTargetDetectionAlgorithmIncorporatingOcclusionInformation

12,31

ZHOUJianting,XUANShibin,WANGTing

1.CollegeofElectronicInformation,GuangxiMinzuUniversity,Nanning530006,China

2.CollegeofArtificialIntelligence,GuangxiMinzuUniversity,Nanning530006,China

3.GuangxiKeyLaboratoryofHybridComputationandICDesignandAnalysis,Nanning530006,China

Abstract:Fortheproblemofcomplextargetscenes,manysmalltargetsandsevereocclusioninUAVaerialimages,an

improvedDDETR(deformableDETR)UAVtargetdetectionalgorithmthatincorporatestargetocclusioninformationis

proposed.TheproposedmodelreplacestheresidualnetworkintheDDETRmodelwithSwinTransformertoobtainricher

multi-levelsemanticfeatures;increasestheuseoflow-levelfeaturesintheDDETRmodeltoimprovethedetectionof

smallandmedium-sizedtargets;andusestheproposedocclusiondegreeestimationmoduletoassistthemodelinsolving

theocclusionproblem,sothatthemodelcanbetterdetecttheheavilyoccludedtargets.Themeanaverageprecision(AP)

of32.3%isachievedontheVisDronedataset,whichis3.3percentagepointshigher