基于DSP的通过局部特征实时物体识别嵌入式系统xWord文件下载.docx
《基于DSP的通过局部特征实时物体识别嵌入式系统xWord文件下载.docx》由会员分享,可在线阅读,更多相关《基于DSP的通过局部特征实时物体识别嵌入式系统xWord文件下载.docx(10页珍藏版)》请在冰点文库上搜索。
Abstract
Inthelastfewyears,objectrecognitionhasbecomeoneofthemostpopulartasksincomputervision.Inparticular,thiswasdrivenbythedevelopmentofnewpowerfulalgorithmsforlocalappearancebasedobjectrecognition.So-called‘‘smartcameras’’withenoughpowerfordecentralizedimageprocessingbecamemoreandmorepopularforallkindsoftasks,especiallyinthefieldofsurveillance.Recognitionisaveryimportanttoolastherobustrecognitionofsuspiciousvehicles,personsorobjectsisamatterofpublicsafety.Thissimplymakesthedeploymentofrecognitioncapabilitiesonembeddedplatformsnecessary.Inourworkweinvestigatethetaskofobjectrecognitionbasedonstate-of-the-artalgorithmsinthecontextofaDSP-basedembeddedsystem.Weimplementseveralpowerfulalgorithmsforobjectrecognition,namelyaninterestpointdetectortogetherwithanregiondescriptor,andbuildamedium-sizedobjectdatabasebasedonavocabularytree,whichissuitableforourdedicatedhardwaresetup.Wecarefullyinvestigatetheparametersofthealgorithmwithrespecttotheperformanceontheembeddedplatform.Weshowthatstate-of-the-artobjectrecognitionalgorithmscanbesuccessfullydeployedonnowadayssmartcameras,evenwithstrictlylimitedcomputationalandmemoryresources.
Keywords DSP;
Objectrecognition;
Localfeatures;
Vocabularytree
1.Introduction
Objectrecognitionisoneofthemostpopulartasksinthefieldofcomputervision.Inthepastdecade,bigeffortsweremadetobuildrobustobjectrecognitionsystemsbasedonappearancefeatureswithlocalextent.Forsuchaframeworktobeapplicableintherealworldseveralattributesareveryimportant:
insensitivityagainstrotation,illuminationorviewpointchanges,aswellasreal-timebehaviorandlarge-scaleoperation.Currentsystemsalreadyhavealotofthesepropertiesand,thoughnotallproblemshavebeensolvedyet,nowadaystheybecomemoreandmoreattractivetotheindustryforinclusioninproductsforthecustomermarket.
Inturn,recentlyembeddedvisionplatformssuchassmartcamerashavesuccessfullyemerged,however,onlyofferingalimitedamountofcomputationalandmemoryresources.Nevertheless,embeddedvisionsystemsarealreadypresentinoureverydaylife.Almosteveryone’smobilephoneisequippedwithacameraand,thus,canbetreatedasasmallembeddedvisionsystem.Clearlythisgivesrisetonewapplications,likenavigationtoolsforvisuallyimpairedpersons,orcollaborativepublicmonitoringusingmillionsofartificialeyes.Inaddition,thelowpriceofdigitalsensorsandtheincreasedneedforsecurityinpublicplaceshasledtoatremendousgrowthinthenumberofcamerasmountedforsurveillancepurposes.Theyhavetobesmallinsizeandhavetoprocessthehugeamountsofavailabledataonsite.Furthermore,theyhavetoperformdedicatedoperationsautomaticallyandwithouthumaninteraction.Notonlyinthefieldofsurveillance,butalsointheareasofhouseholdrobotics,entertainment,militaryandindustrialrobotics,embeddedcomputervisionplatformsarebecomingmoreandmorepopularduetotheirrobustnessagainstenvironmentaladversities.EspeciallyDSP-basedembeddedplatformsareverypopularastheyarepowerfulandcheapCPUs,whicharestillsmallinsizeandefficientintermsofpowerconsumption.AsDSPofferthemaximuminflexibilityofthesoftwaretoberun,comparedtootherembeddedunitslikeFPGAs,ASICorGPU,theircurrentsuccessisnotsurprising.
Forthereasonsalreadymentioned,recognitiontasksareaveryimportantareaofresearch.However,inthisrespectsomeattributesofembeddedplatformsstrictlylimitthepracticabilityofcurrentstate-of-the-artapproaches.Forexample,theamountofmemoryavailableonadevicestrictlylimitsthenumberofobjectsinthedatabase.Therefore,forbuildinganembeddedobjectrecognitionsystem,onegoalistomaketheamountofdatatorepresentasingleobjectassmallaspossibleinordertomaximizethenumberofrecognizableobjects.Anotherimportantaspectisthereal-timecapabilityofthesesystems.Algorithmshavetobefastenoughtobeoperationalintherealworld.Theyhavetoberobustanduser-friendly;
otherwise,aproductequippedwithsuchfunctionalityissimplyunattractivetoapotentialcustomer.Forexample,inaninteractivetourthroughamuseum,objectrecognitiononamobiledevicehastobefastenoughtoallowforcontinuityinguidance.Formallyspeaking,weconsiderthistobeanapplicationrequiringsoftreal-timesystembehavior.Clearly,thisisjustoneexample,andtheexactmeaningofthetermreal-timeisdependentontheconcreteapplication.Westillconsideranobjectrecognitionsystemasbeingreal-
timecapable,ifitisabletodeliveratleastoneresultpersecond.Thisalreadyservesenoughformanyapplicationsliketheexampleoftheinteractivemuseumintroducedabove.However,itisclearthatthisdefinitiondoesnotmeetotherapplications,andthatanimprovementinthroughputisneededforobjectrecognitionatframerate,forinstanceincombinationwithobjecttracking.Tosummarize,buildingafull-featuredrecognitionsystemonanembeddedplatformturnsouttobeachallengingproblemgivenallthedifferentaspectsandenvironmentalrestrictionstoconsider.
Inthiswork,wedescribeamethodtodeployamediumsizedobjectrecognitionsystemonaprototypicalDSPbasedembeddedplatform.Tothebestofourknowledge,wearethefirsttoextensivelyinvestigateissuesrelatedtoobjectrecognitioninthecontextofEmbeddedSystems;
bynowthisistheonlyworkstudyingtheinfluenceofvariousparametersonrecognitionperformanceandruntimebehavior.Wepickasetofhigh-levelalgorithmstodescribeobjectsbyasetofappearancefeatures.AsaprototypicallocalfeaturebasedrecognitionsystemweusedifferenceofGaussian(DOG)keypointsandprincipalcomponentanalysisscaleinvariantfeaturetransform(PCASIFT)descriptorstobuildcompactobjectrepresentations.Byarrangingthisinformationinaclevertreelikedatastructurebasedonk-meansclustering,aso-calledvocabularytree,real-timebehaviorisachieved.Byapplyingadedicatedcompressionmechanism,thesizeofthedatastructurecanbetradedoffagainsttherecognitionperformanceandtherebyaccuratetuningthepropertiesofarecognitionsystemtoagivenhardwareplatformcanbeperformed.Asitisshowninextensiveevaluationsbyconsideringboth,specialpropertiesofthealgorithmsanddedicatedadvantagesofspecialhardware,considerablegainsinrecognitionperformanceandthroughputcanbeachieved.
Theremainderofthispaperisstructuredasfollows.InSect.2wegiveanoverviewaboutdevelopmentsinbothareasthatwearebringingtogetherinourwork.Ontheonehandwelistanumberofreferencesinthecontextofobjectrecognitionbycomputervision;
ontheotherhand,weciteanumberofpublicationsfromtheareaofembeddedsmartsensors.Adetaileddescriptionofthemethodsinvolvedinbuildingourobjectrecognitionalgorithmisgiveninpart3.InSect.4weoutlineourframeworkandgivedetailsabouttrainingandimplementationofoursystem.Wecloselydescribeallstepsindesigningourapproachandgivesidenotesonalternativemethods.InSect.5,weexperimentallyevaluateoursystemonachallengingobjectdatabaseanddiscussrealtimeandreal-worldissues.Furthermore,weinvestigatesomespecialfeaturesofourapproachandelucidatethedependenciesofseveralparametersontheoverallsystemperformance.TheworkconcludeswithsomefinalnotesandanoutlookonfutureworkinSect.6.
2.RelatedWork
Inthefollowingwewillgiveashortintroductiontothetopicoflocalfeaturebasedobjectrecognition.Duetothehugeamountofliteratureavailable,wewillfocusonthemostpromisingapproachesusinglocalfeatures,andrefertothosealgorithmswhicharesomehowrelatedtoourwork.Wewillalsogiveashortoverviewaboutobjectrecognitioninthecontextofembeddedsystems,which,duetothesparsenessofexistingapproaches,containbothglobalandlocalmethods,aswellasalgorithms
implementonFPGAandDSP-basedplatforms.
Local-appearancebasedvisualobjectrecognitionbecamepopularafterthedevelopmentofpowerfulinterestregiondetectorsanddescriptors.Earlyfull-featuredobjectrecognitionsystemsdealingwithalltheindividualalgorithmicstepsandtheirrelatedproblemswereproposedbySchmidandMohr,andSchieleandCrowley.Themainideabehindlocalfeaturebasedobjectrecognitionismaintainingobjectrepresentationsfromcollectionsoflocallysampleddescriptions.Inotherwords,theappearanceoflocalpartsofasingleobjectisencodedindescriptors,andasetofthesedescriptorsformsthefinalobjectrepresentation.Forfindingthedistinguishableregions,so-calledinterestregiondetectorsareused,whichfindregionsorpointsofspecialvisualdistinctiveness.Theneighborhoodofsuchregionsissubsequentlyencodedusingaspecialtransformtobuildadescriptioninherentlyprovidingseveraldesirableproperties.Besideinsensitivityagainstilluminationchangesandpartialviewpointinvariance,representationsassetsoflocaldescriptorsofferrobustnessagainstbackgroundclutterandpartialocclusions.Needlesstosaythatasocalledbagofdescriptorsrepresentationcanbebuiltusingonesingleorseveralcombinationsofdifferentdetectorsanddescriptors.
Thecollectivityofalldescriptorsfrommultipleobjects(i.e.,bagsofdescriptors)isusedtobuildadatabase.Giventhisdatabaseandanewrepresentationofanobjecttobe