Piscataway，NJ，USA:InstituteofElectricalandElectronicsEngineers，2017:658666.［18］樊钰．基于深度学习的安全帽检测系统设计与实现［D］．呼和浩特:内蒙古大学，2019.FANY．Designandimplementationofdetectionsystemofwearinghelmetsbasedondeeplearning［D］．Huhhot:InnerMongoliaUniversity，2019.ResearchonthelightweightdetectionmethodofpersonhelmetwearingZHANGYu-tao，ZHANGMeng-fan，SHIXue-qiang，CHENXiao-kun，RENYao，LIURui(CollegeofSafetyScienceandEngineering，Xi'anUniversityofScienceandTechnology，Xi'an710054，China)Abstract:Theexistinghelmetdetectionalgorithmsbasedondeeplearninghavehighcomputationcomplexityandhighrequirementsforhardwarecomputingcapabilities，resultinginahighcostinactualapplications．Moreover，thediversityoftheshapeandscaleofthedetectiontargetsarenotfullyconsideredinexistingalgorithms，sotherealsoexistcommonproblemsofmissedandfalsedetectionofsmalltargets．Aimingattheaboveproblems，thispaperproposesalightweighthelmetdetectionmodelbasedonthePytorchdeeplearningframeworktoreducetheamountofcalculationofthemodel．Inaddition，adeformablebi-directionaggregationnetworkisproposedtotransfershallowdetailinformationanddeepsemanticinformationinabi-directionalway，therebyimprovingthemodel'sadaptabilitytodetecttargetsofdifferentscales．Andthedeformableconvolutionisintroducedtoimprovethemodel'sadaptabilitytodetecttargetsofdifferentshapes．Wevalidatetheeffectivenessofouralgorithmswithextensiveexperimentsonahelmetdetectiondataset(SafetyHelmetWearingDataset，SHWD)．Weuse6000imagesinSHWDasthetrainingsetand2000imagesasthetestingset．AcomputerwithIntel-Xeon(R)4214CPU(2.2GHz)，64GBmemory，andfourNVIDIAGeForceGTX2080TiGPUsareusedastheexperimentalplatform．Theexperimentalresultsindicatethattherecognitionaccuracyoftheproposedalgorithmisover90%．Theproposeddeformablebi-directionaggregationnetworkemploysdeepsemanticfeaturesandshallowdetailfeatures，andadaptivelyadjuststhereceptivefield，whichcanadapttoobjectsofdifferentshapesandscales，thusimprovingthedetectionaccuracy．Besides，thep...
