北大中文核心期刊国外电子测量技术DOI:10.19652/j.cnki.femt.2204450融合特征金字塔和通道注意力的轻量车辆检测算法*张奇陈梦蝶赵杰(西安工程大学电子信息学院西安710048)摘要:车辆检测是智能交通、无人驾驶等系统得以实现的重要支撑性技术。低精度或低速度的车辆检测器应用受限,因此提出了一种快速准确的车辆检测器。首先,前端特征提取网络VGG16由MobileNetV3_Large替代,减少了参数量和计算量,并增加了对高维特征的提取能力;其次,利用特征金字塔思想构建双向加权融合网络,有效融合不同尺度的特征,获取多维度的车辆特征;最后在特征提取层引入高效通道注意力,重新标定不同特征通道的重要性,进一步提高模型性能。与SSD相比,所提出的模型在KITTI数据集和BDD100K数据集上分别将平均精度提高了7.50%和3.50%,并具有实时检测能力(超过40fps),在检测精度和速度方面有更好的平衡,说明了方法的有效性。关键词:车辆检测;SSD;MobileNetV3;特征金字塔;注意力机制中图分类号:TP391文献标识码:A国家标准学科分类代码:510.4LightweightvehicledetectionnetworkfusingfeaturepyramidandchannelattentionZhangQiChenMengdieZhaoJie(SchoolofElectronicsandInformation,Xi'anPolytechnicUniversity,Xi'an710048,China)Abstract:Vehicledetectionisanimportantsupportingtechnologyfortherealizationofintelligenttransportation,autonomousdriving,etc.Pooraccuracyorlowinferencevehicledetectorsarelimitedinapplication,thereforethispaperproposesafastandaccuratevehicledetector.First,thefront-endfeatureextractionnetworkVGG16isreplacedbyMobileNetV3_Large,whichreducesthenumberofparametersandcomputation,andincreasestheabilitytoextracthigh-dimensionalfeatures.Next,thefeaturepyramidideaisusedtoconstructaweightedbi-directionalfusionnetworktoobtainmulti-dimensionalvehiclefeatures;Intheend,introducingefficientchannelattentioninthefeatureextractionlayertore-calibratetheimportanceofdifferentfeaturechannelsandfurtherimprovethemodelperformance.ComparedwithSSD,ourproposedmodelimprovesmAPby7.50%and3.50%onKITTIdatasetandBDD100Kdataset,andwithreal-timeinference(morethan40fps),itreportsabettertrade-offintermsofdetectionaccuracyandspeed,illustratingtheeffectivenessofourmethod.Keywords:vehicledetection;SSD;Mobi...
