基于改进YOLOv8n-seg的非机动车道场景实例分割算法

打开文本图片集

关键词：自动驾驶；非机动车道；实例分割；YOLOv8n-seg；DEAB；CARAFE；CBAN 中图分类号：TN911.73-34；TP391.41；U463.67 文献标识码：A 文章编号：1004-373X（2026）07-0170-10

Non-motorized lanesceneinstancesegmentationalgorithm based on improved YOLOv8n-seg

LIU Zizhao’， ZHANG Yuting1，TENG Guifa1，²， ZHU Guopeng1 （1.Collegeof Information Scienceand Technology，Hebei Agricultural University，Baoding O71Ooo，China; 2.HebeiProvincial KeyLaboratory ofAgricultural BigData，BaodingO71ooo，China）

Abstract：Inviewof theenvironmentalcomplexity，severeobjectocclusion，andimbalancedcategorydistributionin non-motorizedlanes，thispaperproposesanon-motorizedlanesceneinstancesegmentationalgorithmbasedonimproved YOLOv8n-seg，andthealgorithmistermedDCC-YOLOvn-seg.Firstly，adynamiceficientattentionblock（DEAB）isintroduced intotheetwork.Dynamicfocusingatpixellevelisrealizedbydetail-enhancedconvolutionandcontent-guidedattention mechanism.Alightweightupsampling operatorCARAFEisincorporated toexpand thereceptive field.Meanwhile，upsampling kernelsareadjusteddynamicallytocapturesemanticinformationofdiferentscalesmorecompletely.Furthermore，the convolutional blockattentionmodule（CBAM）isembeddedintothebackbonenetwork tofurtherenhance thedetection performanceofthemodelincomplexenvironments.Themodel'sfocusonoccludedregionsandcriticaltargets isenhanced by chanel-wiseandspatial-wisefeaturemapweighting.Finally，alossfunctioncombiningCIoUandFocalLossisdesignedto optimize boundingboxregressionaccuracyandmitigateclassimbalance.Experimentsonaself-builtdatasetdemonstratethatthe precision of the improved algorithm reaches 76.8%，and its mask mean average precision （mAP reaches 52.1% ，which areimprovedby3.2%andO.9%，respectively，incomparisonwiththoseoftheoriginalalgorithmYOLOv8n-seg.Themodel exhibitsstrongrobustnessinscenarioswithdenseroadsidevehiclesandmulti-objectoccusion，soitcanprovidenoveltechnical support for urban road environment perception and intelligent parking decision-making of autonomousdriving vehicles.

Keywords：autonomous driving; non-motorized lane; instance segmentation;YOLOv8n-seg;DEAB; CARAFE;CBAM

0 引言

随着自动驾驶技术的快速发展，提升系统的环境感知能力已成为现代交通领域的研究重点。（剩余15058字）