Rapid object detection using a boosted cascade of simple features--viola_cvpr2001

《Rapid object detection using a boosted cascade of simple features--viola_cvpr2001》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、RapidObjectDetectionusingaBoostedCascadeofSimpleFeaturesPaulViolaMichaelJonesviola@merl.commichael.jones@compaq.comMitsubishiElectricResearchLabsCompaqCambridgeResearchLab201Broadway,8thFLOneCambridgeCenterCambridge,MA02139Cambridge,MA02142Abstracthi

2、ghframerates.OursystemachieveshighframeratesworkingonlywiththeinformationpresentinasinglegreyThispaperdescribesamachinelearningapproachforvi-scaleimage.Thesealternativesourcesofinformationcansualobjectdetectionwhichiscapableofprocessingimagesalsobein

3、tegratedwithoursystemtoachieveevenhigherextremelyrapidlyandachievinghighdetectionrates.Thisframerates.workisdistinguishedbythreekeycontributions.ThefirstTherearethreemaincontributionsofourobjectdetec-istheintroductionofanewimagerepresentationcalledthe

4、tionframework.Wewillintroduceeachoftheseideas“IntegralImage”whichallowsthefeaturesusedbyourde-brieflybelowandthendescribethemindetailinsubsequenttectortobecomputedveryquickly.Thesecondisalearningsections.algorithm,basedonAdaBoost,whichselectsasmallnum

5、-Thefirstcontributionofthispaperisanewimagerepre-berofcriticalvisualfeaturesfromalargersetandyieldssentationcalledanintegralimagethatallowsforveryfastextremelyefficientclassifiers[5].Thethirdcontributionisfeatureevaluation.MotivatedinpartbytheworkofPapa

6、-amethodforcombiningincreasinglymorecomplexclassi-georgiouetal.ourdetectionsystemdoesnotworkdirectlyfiersina“cascade”whichallowsbackgroundregionsofthewithimageintensities[9].Liketheseauthorsweuseasetimagetobequicklydiscardedwhilespendingmorecompu-offe

7、atureswhicharereminiscentofHaarBasisfunctionstationonpromisingobject-likeregions.Thecascadecanbe(thoughwewillalsouserelatedfilterswhicharemorecom-viewedasanobjectspecificfocus-of-attentionmechanismplexthanHaarfilters).Inordertocomputethesefeatureswhichu

8、nlikepreviousapproachesprovidesstatisticalguar-veryrapidlyatmanyscalesweintroducetheintegralim-anteesthatdiscardedregionsareunlikelytocontaintheob-agerepresentationforimages.Theintegralimagecanbejectofinterest.Inthedomainoffacedetectionthesystemcompu