面向自动驾驶的多尺度目标三维检测算法

打开文本图片集

关键词：自动驾驶；单目相机；三维目标检测；多尺度感知；特征融合；注意力机制；机器视觉中图分类号：TN911.73-34；TP391.41 文献标识码：A 文章编号：1004-373X（2026）01-0141-07

Multi-scale object 3D detection algorithm for autonomous driving

LIUMan，CHEN Xiaonan （SchoolofInformationScienceandTechnology，DalianMaritimeUniversity，Dalian116O26，China）

Abstract：Inautonomous drivingscenarios，3Dobject detectionusingamonocularcamera isachallnging task，especilly incomplexroadenvironments，becausethediferencesofobjectscaleandoclusioncanleadtofalsepositivesormised detectionseasily.Inviewof theabove，amonocular3Dobjectdetectionalgorithmbasedonfeaturefusionandenhancementis proposed.Firstly，aFasterNet+isconstructedandtakenasthebackbonenetwork，andtheembeddinglayerandblockstructure areoptimizedtoenhancetheextractionofdetailedinformation，improving theoverallperformanceofthenetwork．Next，a multi-dimensionalfeatureadaptivefusionmoduleisdesignedtoadaptivelyselectandfusehigh-dimensionalandlow-dimensional features，eliminatingthefactsofthesmallobjectinformationlossinhigh-dimensionalfeaturesandthelackofcontextual informatininlow-dimensionalfeatures.Finaly，afatureehancementatentionmoduleisintroducedtoiglighttereasfor specificobjects，furtherimprovingthenetwork'saccuracyinobjectlocalizationandclasification.Experimentalresultsonthe nuScenesdataset demonstratethatthemAPandNDSof theproposed methodisimprovedbyO.O38andO.O35，respectively，in comparisonwiththebaselinemethod，andthattheproposedmethodcandetectobjectsofdiferenttypesandscalesffectively andexhibitsstrongerrobustnessInaword，theproposedmethodprovidesanewideaformulti-dimensionalojectdetectinin autonomous driving scene.

Keywords：autonomousdriving；monocular camera;3Dobjectdetection；multi-scale perception;feature fusion;attention mechanism;machinevision

0 引言

随着科技进步和智能需求增长，自动驾驶、智能交通和机器人等技术成为发展焦点，而目标检测作为其中的关键技术也愈发重要。（剩余9065字）