基于深度神经网络模型的城市交通监控车辆分类模型(IJEME-V6-N1-3)

I.J.EducationandManagementEngineering,2016,1,18-31PublishedOnlineJanuary2016inMECS()DOI:10.5815/ijeme.2016.01.03Availableonlineat:DeepLearning,DeepNeuralNetwork,UrbanSurveillance,VehicleClassification,VehicleDetection.©2016PublishedbyMECSPublisher.Selectionand/orpeerreviewunderresponsibilityoftheResearchAssociationofModernEducationandComputerScience.1.Introduction1.1.VehicleDetectionVehicledetectionisdefinedasdetectingthevehiclesonthebasisofparameterssuchascolor,shapeandsize.Vehiclesaredetectedusuallybyextractingthevehiclequeuesfromthesatelliteimages.Thevehiclescanbedetectedwiththehelpofneuralnetworki.e.convolutionalneuralnetwork.Thecompletesystemistrainedinordertoclassify,locateanddetecttheobjectsinimages.Hencethiscanimprovetheaccuracyofclassification,detectionandlocalization.Thenetworkcanbeappliedatmultiplelocationsintheimageusingtheslidingwindowtechnique.Thenthesystemistrainedtoproducepredictionofthesizeandlocationofboundingbox.Atechniqueisdefinedtoperformobjectlocalizationwithconvolutionalnetworkbasedsegmentation.Thecentral*Correspondingauthor.Tel.:E-mailaddress:ANovelVehicleClassificationModelforUrbanTrafficSurveillanceUsingtheDeepNeuralNetworkModel19pixelofviewingwindowisclassifiedasaboundarybetweentheregions.Whilecategorizingtheregions,semanticsegmentationisperformed[1].Avehiclecrashtestmethodisdefinedinordertoaccessthecrashworthinessofthevehicle.Visualinspectionlikecaraccelerationisusedtoaccesstheoverallsafetyofcar.Butitwascostly,complexandtakesmoretime.ThereforeatechniqueisdefinedtoreproducethekinematicsofcarduringacollisionbyusingNARmodeli.e.Nonlinearautoregressivemodeltodeterminetheestimatedparametersbyusingafeeforwardneuralnetwork.NARmodelisderivedfromNARMAi.e.Nonlinearautoregressivewithmovingaverage.Thismodelisusedtopredictthekinematicresponsessuchasacceleration,velocityanddisplacementofcarduringthecollision[2].Monocularvehicledetectiondefinesthefeatureextractionparadigmwhichisbasedonmachinelearning.Thistechniqueworkswhenthevehicleiscompletelyvisible.Thevehicledetectiontechniqueismainlydividedintotwotypesi.e.appearancebasedandmotionbased.Appearancebasedapproachisusedtorecognizethevehiclesdirectlyfromimagesi.e.frompixelstovehiclesandmotionbasedapproachrequiresthesequenceofimagessoastorecognizethevehicles.Todetecttheobject,clusteringwasusedthatwasimplementedbyusingmodifiedversionofiterativeclosestpointthatcanbedonebyusingthepolarcoordinatestosegmentobjects.Vision-basedvehicledetection,vehicletrackingandbehaviouranalysisonroadisalsodefinedinwhichthevision-basedvehicledetectionisplacedincontextofsensor-basedon-roadperception.Theprogressinvision-basedvehicledetectionisalsodefinedformonocularandstereo-visionsensorconfigurationsandanalyzationoffiltering,estimationanddynamicmodelsisalsodefined.Thestateofartinon-roadbehaviouranalysisisreviewedaddressingthecontextanalysis,longtermmotionclassificationandprediction.Thevehicletrackingisalsodiscussedinwhichvision,detailedimageplaneand3Dtechniquesformodelling,filteringvehicledynamicsontheroadisused[3].1.2.ImageClassificationImageclassificationcanbedefinedasclassifyingtheimageintopixelsandtheimageclassificationsystemscontainstwopartsi.e.bag-of-features(BOF)andspatialpyramidmatching(SPM).TheBOFtechniquerepresentstheimageasahistogramoflocalfeaturesanditisincapableofcapturingtheshapesorlocatinganobject.Thevariousstepsareusedtoclassifytheimagei.e.extractandrepresentpatches,generatethecodewords,encodeandpoolfeatures.TheSPMtechniqueisusedtodividetheimageintofinerspatialsub-regionsandthencomputesthehistogramsoflocalfeaturesfromeverysub-region[4].Imageclassificationisusedtoassignoneormorecategorylabelstotheimage.Theimageclassificationapproachcontainsthreecomputationalstepsi.e.descriptorcoding,spatialpoolingandimageclassification.Indescriptorcoding,eachdescriptorofanimageismappedinordertoformahigh-dimensionalsparsevectorandinimageclassification,imagelevelfeaturevectorisnormalizedandfedintoaclassifier[5].1.3.HDCNNHybriddeepneuralnetworkcanbedefinedasthedetectionofsmallobjectslikevehiclesinsatelliteimages.