北大中文核心期刊国外电子测量技术DOI:10.19652/j.cnki.femt.2204441YOLOX-IM:一种无人机航拍视频的轻量化交通参数提取模型*刘军黎刘晓锋邱洁衣雨玮(天津职业技术师范大学汽车与交通学院天津300222)摘要:在无人机航拍的过程中,背景更广阔,目标的尺寸更小,种类更复杂。提出一种基于YOLOX-s的轻量化无人机航拍目标检测算法YOLOX-IM。首先,为了提高该模型检测小目标的性能,通过使用切片辅助推理(slicingaidedhyperinference,SAHI)算法以及坐标修正矩阵对训练集进行预处理和数据增强。然后,在路径聚合网络(pathaggregationnetwork,PAN)中引入一个浅层特征图以及超轻量级子空间注意模块,并添加一个检测头对小物体进行检测;最后,对边界回归的损失函数进行了优化。在VisDrone2019数据集的消融实验结果表明,所提出的模型检测精度与基础YOLOX-s相比高了8.13%;模型体积4.55MB,相较于原模型下降67.14%。利用该模型在中国天津市渌水道进行实地交通监测的交通参数提取实验,在设定的场景中,当无人机航拍高度为50m时,该模型的交通提取参数精度最高,达到96.14%。关键词:无人机;车辆检测与跟踪;YOLOX模型;深度学习中图分类号:TP391文献标识码:A国家标准学科分类代码:520.604YOLOX-IM:AlightweighttrafficparameterextractionmodelforUAVaerialimagesLiuJunliLiuXiaofengQiuJieYiYuwei(SchoolofAutomotiveandTransportation,TianjinUniversityofTechnologyandEducation,Tianjin300222,China)Abstract:IntheprocessofUAVaerialphotography,thebackgroundisbroaderandthetargetsaresmallerinsize.Inthispaper,weproposealightweighttargetdetectionalgorithmYOLOX-IMforUAVaerialphotographybasedonYOLOX-s.First,toimprovetheperformanceofsmalltargetdecection,thetrainingsetispreprocessedanddataisenhancedbyusingaslicingaidedhyperinference(SAHI)algorithmaswellasacoordinatecorrectionmatrix.Then,ashallowfeaturemapaswellasanultra-lightweightsubspaceattentionmoduleareintroducedinthepathaggregationnetwork(PAN),andadetectionheadisaddedforsmallobjectdetection.Finally,thelossfunctionoftheboundaryregressionisoptimized.TheexperimentalresultsontheVisDrone2019datasetshowthattheproposedmodelhas8.13%higherdetectionaccuracycomparedwiththeYOLOX-s;comparedwiththeoriginalmodel,themodelvolumeissignificantlyreducedto4.55MB,whichis67.1...
