多特征融合和交叉核SVM的车辆检测方法.docx

多特征融合和交叉核SVM的车辆检测方法
Abstract
Vehicledetectionisasignificantresearchproblemincomputervision.Inthispaper,weproposeanovelmethodforvehicledetectionusingmulti-featurefusionandcross-kernelSVM.OurproposedmethodistestedontheCaltech-USAdataset,andtheresultsshowthatitoutperformsstate-of-the-artmethodsinvehicledetectionaccuracy,precision,andrecall.
Introduction
Vehicledetectionisacriticalproblemincomputervisionasithasnumerouspracticalapplicationsinthefieldsoftrafficmanagement,surveillance,andautonomousdriving,etc.Thefundamentaltaskofavehicledetectionsystemistoidentifyvehiclesfromagivenimageoravideo.Thechallengeinvehicledetectionisthatvehiclescanhavevarioussizes,shapes,andorientations,andcanbeobscuredbyotherobjectsorpartiallyoccluded.Inthispaper,weproposeanovelmethodforvehicledetectionthatusesmulti-featurefusionandcross-kernelSVM.
RelatedWorks
Manytechniqueshavebeenusedforvehicledetection,withvaryingdegreesofsuccess.Somepopulartechniquesincludetemplatematching,Haarcascades,feature-basedmethods,andmachinelearning-basedmethods.Templatematchingsimplycomparesanimagepatchwithatemplaterepresentingavehicle.Thismethodcanbeveryaccurate,butitisonlyusefulfordetectingvehicleswithaspecificshapeandorientation.Haarcascadesuseasetofhaar-likefeaturestodetectobjectsofdifferentshapesandsizes.Thismethodiscomputationallyefficientbuthasloweraccuracy.Feature-basedmethodsusehand-craftedorlearnedfeaturesextractedfromtheimagetodetectvehicles.Thesemethodsperformreasonablywell,buttheymaynotberobusttochangesintheappearanceofvehicles.Machinelearning-basedmethodstrainadetectorusingasetoflabeledimages.Thesemethodscanbeveryaccuratebutrequirelargeamountsoflabeleddataandmaygeneralizepoorlytounseendata.
ProposedMethod
Inourproposedmethod,weusemulti-featurefusionandcross-kernelSVMtoimprovetheaccuracyandrobustnessofvehicledetection.Weusefourdifferentfeatures,includingHOG,LBP,SIFT,andColorHistogram,tocapturevariousvisualinformationintheimage.Sincedifferentfeaturesaresensitivetovariousaspectsoftheimage,usingallfourfe