基于深度卷積神經(jīng)網(wǎng)絡的地鐵行人檢測算法關鍵技術研究

基于深度卷積神經(jīng)網(wǎng)絡的地鐵行人檢測算法關鍵技術研究摘要

隨著城市化進程的不斷加速，地鐵成為了現(xiàn)代城市中最為重要的公共交通工具之一，人口密度也不斷的增加。為了提高地鐵行人的安全性和系統(tǒng)的智能化程度，本文提出了一種基于深度卷積神經(jīng)網(wǎng)絡的地鐵行人檢測算法。首先，本文通過對目標檢測方法的研究，將FasterR-CNN應用于地鐵行人檢測。然后本文針對在地鐵場景下行人檢測中所存在的問題，如人群密集、光線條件不穩(wěn)定等，提出了一系列關鍵技術，包括RetinaNet的改進和用于光照不均和強烈陰影區(qū)域的自適應閾值選擇等。最后，使用KSTSubwayPedestrian數(shù)據(jù)集對所提出的算法進行了評估，并分別與基于YOLO、SSD和FasterR-CNN等方法進行了比較。實驗結(jié)果表明，本文所提出的方法具有較高的檢測精度和實時性能，在地鐵行人檢測領域具有一定的推廣價值和應用前景。

關鍵詞：地鐵行人檢測，深度卷積神經(jīng)網(wǎng)絡，F(xiàn)asterR-CNN，RetinaNet，自適應閾值選擇。

ABSTRACT

Withtheaccelerationofurbanizationprocess,subwayhasbecomeoneofthemostimportantpublictransportationtoolsinmoderncities,andthepopulationdensityisalsoincreasing.Inordertoimprovethesafetyofsubwaypedestriansandtheintelligenceofthesystem,thispaperproposesasubwaypedestriandetectionalgorithmbasedondeepconvolutionalneuralnetwork.Firstly,basedonthestudyofobjectdetectionmethods,FasterR-CNNisappliedtosubwaypedestriandetection.Then,inviewoftheproblemsexistinginpedestriandetectioninsubwayscenarios,suchascrowddensityandunstablelightingconditions,aseriesofkeytechnologiesareproposed,includingtheimprovementofRetinaNetandtheadaptivethresholdselectionmethodusedforunevenilluminationandstrongshadowareas.Finally,theproposedalgorithmisevaluatedusingKSTSubwayPedestriandatasetandcomparedwithmethodsbasedonYOLO,SSDandFasterR-CNN,respectively.Theexperimentalresultsshowthattheproposedmethodhashighdetectionaccuracyandreal-timeperformance,andhascertainpromotionvalueandapplicationprospectsinsubwaypedestriandetectionfield.

Keywords:Subwaypedestriandetection;deepconvolutionalneuralnetwork;FasterR-CNN;RetinaNet;adaptivethresholdselectionSubwaypedestriandetectionisacriticaltaskinthefieldoftransportationsafety.Inrecentyears,deeplearning-basedmethodshaveachievedremarkableprogressinpedestriandetection.However,existingapproachesstillfacechallengessuchaslowdetectionaccuracy,highcomputationalcomplexity,andslowreal-timeperformance.

Toaddresstheseproblems,thispaperproposesanadaptivethresholdselectionstrategytoimprovethedetectionaccuracyofRetinaNet,astate-of-the-artdeepconvolutionalneuralnetwork.Theproposedmethoddynamicallyadjuststhethresholdaccordingtotheinputimage,whicheffectivelyimprovesthedetectionaccuracyofsmallobjectsandreducesfalse-positivedetections.

Toevaluatetheperformanceoftheproposedalgorithm,experimentsareconductedontheKSTSubwayPedestriandataset,andtheresultsarecomparedwithmethodsbasedonYOLO,SSD,andFasterR-CNN.Theexperimentalresultsdemonstratethattheproposedmethodachieveshighdetectionaccuracywhilemaintainingreal-timeperformance.Moreover,theproposedmethodoutperformsexistingstate-of-the-artmethodsintermsofbothdetectionaccuracyandreal-timeperformance.

Inconclusion,theproposedalgorithmisapromisingapproachforsubwaypedestriandetection.TheadaptivethresholdselectionstrategycaneffectivelyimprovethedetectionaccuracyofRetinaNet,makingitsuitableforreal-timedetectionapplications.FutureworkcouldfocusonexpandingthedatasetandfurtheroptimizingtheproposedalgorithmtoachieveevenbetterperformanceInadditiontofurtheroptimizationoftheproposedalgorithm,futureworkcouldalsofocusonaddressingsomeofthelimitationsandchallengesthatariseinsubwaypedestriandetection.Oneofthemainchallengesisthepresenceofocclusions,whichoccurwhenpedestriansarepartiallyorfullyhiddenbyotherobjectssuchaspillarsorboxesinthesubwaystation.Thiscanleadtofalsenegativesorinaccuratedetectionresults.Toaddressthisissue,researcherscouldexploretheuseofmulti-viewcamerasandsensorfusiontechniquestocombineinformationfrommultiplesourcesandimprovetheaccuracyofdetection.

Anotherchallengeisthevariabilityinlightingconditions,whichcanaffectthequalityofimagesandmakeitdifficulttodetectpedestrians.Onepossiblesolutionistouseinfraredcamerasorothersensorsthatarelesssensitivetolightingconditions.Additionally,researcherscouldinvestigatetheuseofdeeplearningmodelsthatarelesssensitivetolightingvariations,suchasthosethatincorporateattentionmechanismsoradversarialtrainingtechniques.

Furthermore,assubwaysystemsbecomemoreubiquitousincitiesaroundtheworld,researcherscouldfocusonoptimizingtheproposedalgorithmfordifferenttypesofsubwaystationsandsettings,suchasthosewithhighpassengervolumesorthosethatoperateindifferentweatherconditions.Thiscouldinvolvecollectingmorediversedatasetsfromarangeofsubwaysystemsandenvironments,andtrainingthealgorithmonthesedatasetstoimproveitsrobustnessandgeneralizationability.

Inconclusion,whiletheproposedalgorithmshowspromisingresultsforsubwaypedestriandetection,therearestillmanychallengesandlimitationsthatneedtobeaddressedinordertoimproveitsoverallperformanceandscalability.Bycontinuingtoexplorenewtechniquesandapproaches,researcherscanhelptoadvancethefieldofcomputervisioninthecontextofsubwaytransportation,andcontributetothedevelopmentofsaferandmoreefficientsubwaysystemsOnelimitationoftheproposedalgorithmisitsrelianceoncolorinformationforpedestriandetection.Thiscanposechallengesinsituationswherelightingconditionsorclothingcolorsmayvarywidelyamongpedestrians.Toaddressthis,futureresearchshouldexploretheintegrationofothermodalities,suchasdepthorthermalimaging,toimprovedetectionaccuracy.

Anotherchallengeistheneedforreal-timeperformanceinasubwayenvironment,wherefast-movingpedestriansandtrainsrequirequickandaccuratedetection.Thishighlightstheimportanceofoptimizingthealgorithmforspeedandefficiency,suchasthroughtheuseofparallelprocessingorhardwareacceleration.

Finally,thealgorithmwasevaluatedprimarilyondatasetscontainingpedestrianswalkinginastraightline,whichmaynotfullyrepresentthediversityofpedestrianmovementinasubwaysetting.Futureworkshouldevaluatethealgorithmonmorediversedatasets,includingscenariossuchascrowdedplatformsandstaircases.

Overall,whilethereisstillworktobedonetoimprovetheperformancea