基于RT-DETR的轻量化交通标志检测算法

打开文本图片集

引用格式：,.基于RT-DETR的轻量化交通标志检测算法[J].现代电子技术,2025,48(18)：57-64.

关键词：交通标志检测；轻量化；RT-DETR；ShuffleNetV2；剪枝；知识蒸馏中图分类号：TN911.23-34；TP391.41 文献标识码：A 文章编号：1004-373X(2025)18-0057-08

Lightweighttrafficsigndetectionalgorithmbasedon RT-DETR

WANG Zexuan,LEI Xuemei

(SchoolofElectronic InformationEngineering,InnerMongolia University,HohhotO1oooo,China)

Abstract：Traffcsigndetection playsanextremelyimportantroleintheautodrivesystem,which isdirectlyrelatedtothe safetyofvehicledrivingandtheaccuracyof traffcrulecompliance.Traffcsigndetectionhashighrequirementsfordetection accuracy,speedandeal-timeperformane.Currentlythereisaproblemof diicultyinbalancingdetetionspeedandauracy inthefieldoftraffcsigndetection.Inalusiontothisproblem,thetraficsigndetectionisconductedbasedonreal-time detection Transformer(RT-DETR)objectdetectionalgorithm,whichhasbeteraverage precisionperformance.Inordertofurther improveefciencyandeal-timeperformance,animprovedRT-DETRalgorithmisproposed torealizethemodellightweightand improve thetargetdetectionspeed.InthelightweightRT-DETR,thelightweightnetwork ShufleNetV2isusedtoreplacethe originalResNetnetworkasthebackbonenetworkofRT-DETR,whichreduces theamountofcomputationand parameters while ensuringthelearningabilityofRT-DETR,andimprovethedetectionspeed.Inordertofurtheroptimize model’sperformance, thefine-tuningisconductedbymeansof thesubsequentchannelpruning，quantization,andknowledgedistilation.The experimental results show that RT-DETR being lightweighted can achieve a mean average precision (mAP @ 50) of 97.1% on the CCTSDB2021dataset,withaninference timeof13.7msandamodel sizeof16.9MB.Incomparison with theRT-DETR model withResNet-5Oasthebackbone network,themodel sizeisreducedby89.5%and theinference timeisincreased by 43.62% （202 underthepremiseofensuringtheaccuracywhich makesthemodel morelightweightandefectivelyimproves thedetection speed.Incomparisonwiththecurrntmainstreamsimilarobjectdetectionmethods,italsohasfasterdetectionspeedand higher detection precision.

KeyWords:trafic sign detection; lightweight;RT-DETR;ShufleNetV2; pruning; knowledge distillation

0 引言

随着社会的日益发展与科技水平的不断提升，人们对日常出行体验的期待与需求同步升级。（剩余16309字）