文章编号：1002-2082(2024)01-0079-10基于改进YOLOv4的低慢小无人机实时探测算法吴璇，张海洋，赵长明，李志朋，王元泽（北京理工大学光电学院，北京100081）摘要：针对低慢小无人机探测任务中精度不高、在嵌入式平台上部署实时性能差的问题，提出了一种基于改进YOLOv4的小型无人机目标检测算法。通过增加浅层特征图、改进锚框、增强小目标，提高网络对小目标的检测性能，通过稀疏训练和模型修剪，大大缩短了模型运行时间。在1080Ti上平均精度（mAP）达到85.8%，帧率(FPS)达75frame/s，实现了网络轻量化。该模型部署在Xavier边缘计算平台上，可实现60frame/s的无人机目标检测速度。实验结果表明：与YOLOv4和YOLOv4-tiny相比，该算法实现了运行速度和检测精度的平衡，能够有效解决嵌入式平台上的无人机目标检测问题。关键词：低慢小无人机；目标检测；YOLOv4；剪枝；嵌入式中图分类号：TN201文献标志码：ADOI：10.5768/JAO202445.0102002ImprovedYOLOv4forreal-timedetectionalgorithmoflow-slow-smallunmannedaerialvehiclesWUXuan，ZHANGHaiyang，ZHAOChangming，LIZhipeng，WANGYuanze（SchoolofOpticsandPhotonics,BeijingInstituteofTechnology,Beijing100081,China）Abstract：Inordertosolvethelowaccuracyinlow-slow-smallunmannedaerialvehicles(UAVs)missiononembeddedplatformanddeploymentproblemofpoorreal-timeperformance,asmallUAVtargetdetectionalgorithmbasedonimprovedYOLOv4wasproposed.Byincreasingtheshallowcharacteristicfigure,improvingtheanchor,enhancingthesmalltarget,andthedetectionperformanceofnetworkforsmalltargetwasimproved,throughsparsetrainingandmodelpruning,themodelrunningtimewasgreatlyreduced.Theaverageaccuracy(mAP)reaches85.8%onthe1080Ti,andtheframerate(FPS)reaches75frame/s,whichachievingnetworklightweight.ThislightweightmodelwasdeployedontheXavieredgecomputingplatform,whichcouldachievetheUAVtargetdetectionspeedof60frame/s.Experimentalresultsshowthat,incomparedwithYOLOv4andYOLOv4-TINY,thisalgorithmachievesthebalanceofrunningspeedanddetectionaccuracy,andcaneffectivelysolvetheproblemofUAVtargetdetectiononembeddedplatform.Keywords：low-slow-smallunmannedaerialvehicles；targetdetection；YOLOv4；pruning；embeddedIntroductionWiththerapiddevelopmen...
