用于计算机视觉应用的GPU优化立体图像匹配技术(IJMECS-V7-N5-5)

I.J.ModernEducationandComputerScience,2015,5,37-42PublishedOnlineMay2015inMECS()DOI:10.5815/ijmecs.2015.05.05Copyright©2015MECSI.J.ModernEducationandComputerScience,2015,5,37-42GPUOptimizedStereoImageMatchingTechniqueforComputerVisionApplicationsKajalSharmaChosunUniversity,KoreaEmail:[e-mail]—Inthispaper,weproposeagraphicsprocessingunit(GPU)basedmatchingtechniquetoperformfastfeaturematchingbetweendifferentimages.Loweproposedascaleinvariantfeaturetransformalgorithmthathasbeensuccessfullyusedinvariousfeaturematchingapplicationssuchasstereovision,objectrecognition,andmanyothers,butthisalgorithmiscomputationallyintensive.Inordertosolvethisproblem,weproposeamatchingtechniqueoptimizedforgraphicsprocessingunitstoperformcomputationwithlesstime.WehaveappliedGPUoptimizationforthefastcomputationofkeypointstomakeoursystemfastandefficient.Theproposedmethodusedself-organizingmapfeaturematchingtechniquetoperformefficientmatchingbetweendifferentimages.Theexperimentsareperformedonvariousimagestoexaminetheperformanceofthesystemindiverseconditionssuchasimagerotation,scaling,andblurringconditions.Theexperimentalresultsrevealthattheproposedalgorithmoutperformstheexistingfeaturematchingmethodsresultingintofastfeaturematchingwiththeoptimizationofgraphicsprocessingunit.IndexTerms—Featurematching,stereovision,self-organizingmap,graphicsprocessingunitI.INTRODUCTIONFeatureselectionandmatchingisakeycomponentinmanycomputervisiontaskssuchaspathfinding,obstacledetection,navigation,stereovision,andmanyothersapplications[1-6].Severalstrategiesofkeypointdetectorshavebeenproposedintheliterature[7-8].SchmidandMohr[9]usedHarriscornersasinterestpointsinimagerecognitionproblemstomatchthefeaturesagainstalargedatabaseofimages.Thismethodallowsfeaturestobematchedunderarbitraryorientationchanges,butitissensitivetoimagescalechanges.Lowe[10]proposedthescaleinvariantfeaturetransform(SIFT)descriptorfortheextractionofinterestpointsinanimagethatisinvarianttobothscaleandrotation.TheSIFTtechniqueuseda128-dimensionalvectortodescribetheSIFTfeature,whichiscomputationallyintensive.Inarecentresearch[11],wehaveimplementedanefficientfeaturematchingtechniquefasterthanLowe’sSIFTwithself-organizingmap(SOM)thatcanbeusedforreal-timestereomatchingapplications.Inthepresentresearch,weextendedourresearchandimplementourproposedmethodongraphicsprocessingunit(GPU)tofurtheroptimizethetime.Duetotheadvancementsofparallelprocessingtechniques,multi-coreGPUtechniqueshavebeenwidelyappliedtoacceleratethecomputationallyintensivetasks[12].ModernprogrammablegraphicshardwarecontainspowerfulcoprocessorsGPUswithapeakperformanceofhundredsofGigaFLOPSwhichisanorderofmagnitudehigherthanthatofCPUs[13].Foracceleratingtheapplicationsofcomputervisionmanyresearchersarenowexploitingparallelismprovidedbymodernprogrammablegraphicshardwarethatprovidesagreatscopeforaccelerationtoruncomputationsinparallel[14-15].Someresearchersalsoutilizedspecializedhardwareandreconfigurablehardwaretospeedupthesealgorithms[16-18].Oneexampleofabroadareaofapplicationisindiscoveringconcurrencyinparallelcomputing,wherecoloringisusedtoidentifysubtasksthatcanbecarriedoutordataelementsthatcanbeupdatedsimultaneously[19].Anotherexampleofabroadapplicationareaofcoloringistheefficientcomputationofsparsederivativematrices[20].WiththeincreasingprogrammabilityandcomputationalpoweroftheGPU,therecentworkbySinhaetal.[12]acceleratessomepartsoftheSIFTalgorithmusingthehardwarecapacitiesofGPUs.A10xspeedupisobtained,allowingforapplicationsonvideo-sizedimages[12].Avarietyofcomputervisionalgorithmshasbeenparallelized,providingsignificantaccelerationtothecomputation[12,14,15].Inthispaper,anoveltechniqueispresentedthatisdesignedtoachievefastfeaturematchingintheimageswiththeuseofneuralnetworksandGPUs.OurcontributionistheproposalofaGPU-optimizedmatchingtechniquebasedonKohonen’sself-organizingmap(SOM)[21].TheproposedmethodprovidessignificantreductioninthecomputationtimecomparedtoLowe’sSIFT.Inourapproach,thescalespaceforkeypointsextractionisconfiguredinparallelfordetectingthecandidatepointsamongwhichthenumberofkeypointsisreducedwiththeSOMneuralnetwork.ThedescriptorvectorgenerationisacceleratedontheGPUandthematchingisaccomplishedwithcompetitivelearning.ThekeyideaistooptimizethekeypointextractionwithGPUandtoreducethedescriptorsizewiththewinningcalculationmethod.Thesimilarwinningpixelsintheimagesarefoundandassociatedtoaccomplishthefeaturematching.TheproposedmethodoftheGPUisfasterduetotheusageofmulti-processing.38GPUOptimizedStereoImageMatchingTechniqueforComputerVisionApplicationsCopyright©2015MECSI.J.ModernEducationandComputerScience,2015,5,37-42Theremainderofthispaperisorganizedasfollows:SectionIIdescribestheoverviewonstereovision.TheprocedureoffeaturematchingwithGPU-optimizedmethodispresentedinsectionIII.TheexperimentalresultsareshowninsectionIVandconclusionsaredrawninsectionV.II.STEREOVISIONStereovisionisbasedonacquiringthree-dimensional(3D)informationfromdifferentviewsobtainedbyasinglemovingcameraorafixedst