基于AdaBoost的人體檢測算法.pdf

1、東南大學(xué)碩士學(xué)位論文基于AdaBoost的人體檢測算法姓名：錢成申請(qǐng)學(xué)位級(jí)別：碩士專業(yè)：計(jì)算機(jī)科學(xué)與技術(shù)；圖像處理與科學(xué)可視化指導(dǎo)教師：羅立民；鮑旭東20110415東南大學(xué)碩士學(xué)位論文Title：Author：Supervisor：School：AbstractHumanDetectionBasedonAdaBoostQIANChengProfLU0LiminSoutheastUniversityThisprojectstudiedh

2、umandetectioninstaticimageandvideo(cameraf僅ed)respectivelyAdaBoostalgorithmwasintroducedtodetecthumaninstaticimageForthedetectionwithfixedcamera，foregroundWasextractedfirst，followedbyapplyingAdaBoostclassifiertoitTrainin

3、gmethod：TrainingmethodofAdaBoostWasimprovedinthiswork珊1iletrainedresultswerekeptthesame，timecomplexityoftrainingdecreasedfromO(MN2)toO(MNlogⅣ)，whereMwasnumberoffeaturetypesandNWasnumberoftrainingsamplesExperimenton13，000

4、samplesshowedthatimprovedalgorithmtook1／500ofthetimecomparedwitllthetraditionalone，gettingthesameclassifierFeaturecomputation：11leparallelcomputationoffeaturesbasedonGPU(GPGPU，usingCUDA)andCPU(multithread，usingOpenMP)Was

5、studiedThewaybasedonGPUWaschosenbecauseofitsrapiddevelopmentTheparallelprogram，whichusedthevideocardwittlonly16streamprocessing(SP)units，Was06timesfasterthantheserialoneTheacceleratingeffectwouldbecomemuchmoresignificant

6、ifabettervideocardwasavailableForegroundextraction：ThebackgroundwasmodeledusingGaussianMixtureModel(OMM)ClassicGMMtendedtoextractseveralsmallregiOIlSononesingleobjectThisworkproposedamethodusingcolorinformationtoincrease

7、extractionaccuracyBesides，amethodbasedonthethecorrespondenceamongvideoframeswasintroduced，decreasingthecomputationalcostImprovedalgorithmprocessedoneframeofa640X480videoin208msonaverage，whilethetraditionalmethodneeded195