Rapid Object Detection using a Boosted Cascade of

RapidObjectDetectionusingaBoostedCascadeofSimpleFeaturesOriginalAuthorPaulViola&MichaelJonesIn:Proc.Conf.ComputerVisionandPatternRecognition.Volume1.,Kauai,HI,USA(2001)511–518Speaker:JingMingChiuan(井民全)(movingoractingwithgreatspeed)(increasethestrengthorvalueofSth)OutlineIntroductionTheBoostalgorithmforclassifierlearningFeatureSelectionWeaklearnerconstructorThestrongclassifierAtremendouslydifficultproblemResultConclusionWhathadwedone?AmachinelearningapproachforvisualobjectdetectionCapableofprocessingimagesextremelyrapidlyAchievinghighdetectionratesThreekeycontributionsAnewimagerepresentationIntegralImageAlearningalgorithm(BasedonAdaBoost[5])AcombiningclassifiersmethodcascadeclassifiersSelectasmall#ofvisualfeaturesfromalargersetyieldanefficientclassifiersSpeedupthefeatureevaluationDiscardthebackgroundregionsoftheimageWorkingonlywithasinglegreyscaleimageAdemonstrationonfacedetectionAfrontalfacedetectionsystemThedetectorrunat15framespersecondwithoutresortingtoimagedifferencingorskincolordetectionImagedifferenceinvideosequences384x288onaPentiumIII700MHzThebroadpracticalapplicationsforaextremelyfastfacedetectorUserInterface,ImageDatabases,TeleconferencingThesystemcanbeimplementedonasmalllowpowerdevices.CompaqiPaq2frame/secTrainingprocessforclassifierTheattentionaloperatoristrainedtodetectexamplesofaparticularclass---asupervisedtrainingprocessInthedomainoffacedetection1%falsenegative40%falsepostivieFaceclassifierisconstructedCascadeddetectionprocessThesub-windowsareprocessedbyasequenceofclassifierseachslightlymorecomplexthanthelastAnyclassifierrejectsthesub-window,nofurtherprocessingisperformedTheprocessisessentiallythatofadegeneratedecisiontreeOurobjectdetectionframeworkOriginalImageIntegralImageInordertocomputingfeaturesrapidlyatmanyscalesHaarBasisFunctionsHaarBasisFunctionsHaarBasisFunctionsFeatureEvaluationModifiedAdaBoostProcedureFeatureSelectionLarge#offeaturesSmallsetofcriticalfeaturesCascadedClassifiersStructureFeatureSelectionThedetectionprocessisbasedonthefeatureratherthanthepixelsdirectly.TwoReasons:Thead-hocdomainknowledgeisdifficulttolearnusingafinitequantifyoftrainingdata.ThefeaturebasedsystemoperatesmuchfasterThesimplefeaturesareusedTheHaarbasisfunctionswhichhavebeenusedbyPapageorgiouetal.[9]ThreekindsoffeaturesFeatureSelectionThedifferencebetweenthesumofpixelswithintworectangularregionsTwo-RectangleFeatureTheregionhavethesamesizeandshapeAndarehorizontallyorverticallyadjacentThebaseresolutionis24x24Theexhaustivesetofrectangleislarge,over180,000.Three-RectangleFeaturethesumwithintwooutsiderectanglesubtractedfromthesuminacenterrectangleThedifferencebetweenthediagonalpairsofrectanglesFour-RectangleFeature;0),1(,0)1,(),,(),1(),(),,()1,(),(yiixsyxsyxiiyxiiyxiyxsyxsIntegralImageAintermediatedrepresentationforrapidlycomputingtherectanglefeaturesyyxxyxiyxii'',''),(),(TheintegralimageTheoriginalimageTherecurrencespairforonepasscomputingThecumulativerowsum1253467891254611111420si++13841021112545ii+3149Calculatinganyrectanglesumwithintegralimage1A2A+B3A+C4A+B+C+DRectangleSumD=4-3-2+1AdaBoostlearningalgorithmIsusedtodothefeatureselectiontaskLearningClassificationFunctionsLearningProcessFeatureSetTrainingset1.Positive2.NegativeAvariantAdaBoostprocedureFacenon-FaceThefinalstrongclassifierOver180,000rectanglefeaturesassociatewitheachsub-image2424WeakLearner1WeakLearner2WeakLearner2ThefinalstrongclassifierTheBoostalgorithmforclassifierlearning),(,...),,(),,(2211nnyxyxyxImagePositive=1Negative=0Step1:GivingexampleimagesStep2:Initializetheweightspositives.andnegativesof#theareand,1,0for21,21,1lmylmwiiFort=1,…,T1.Normalizetheweights,2.Foreachfeaturej,trainaclassifierhjwhichisrestrictedtousingasinglefeature3.Updatetheweights:ondistributiprobabityaisthatso,1,,,tnjjtitit.errorlowestwiththe,,classifiertheChoose|)(|,respecttowithevaluatediserrorThettiiijijthyxhwwotherwisecorrectlyclassifiedisif,,,1,,1itititetititwxWeaklearnerconstructorttt1TrainingsetWeaklearnerconstructor圖示解說1w2wnwjfjfjfjfFeaturesOver180,000featuresforeachsubimage123000,180iiijijyxhw|)(|Errorsmin1h2h3h000,180hth.errorlowestwiththe,,classifiertheChoosetthNormalizedtheweights1w2wnwiwmisscorrectcorrectmissttititww1,,1UpdatetheweightsTrainingtheweaklearner圖解說明X(Trainingset))(xfjexFaceexamplesNon-FaceexamplesIffj(x)Xisafaceiiijijyxhw|)(|1)(ijxhFalsepositiveFalsenegativefeatureaissign,inequalitytheofdirectiontheindicating,thresholdais0)(if,1)(jjjjjjjjfPwhereotherwisePxfPxhAdaBoostingPlacethemostweightontheexamplesmustoftenmisclassifiedbytheprecedingweakrulesForcingthebaselearnertofocusitsattentiononthe“hardest”examplesTheBoostalgorithmforclassifierlearning),(,...),,(),,(2211nnyxyxyxStep1:GivingexampleimagesStep2:Initializetheweightspositives.andnegativesof#theareand,1,0for21,21,1lmylmwiiFort=1,…,T1.Normalizetheweights,2.Foreachfeaturej,trainaclassifierhjwhichisrestrictedtousingasinglefeature3.Updatetheweights:WeaklearnerconstructorFinalstrongclassif