资源描述：
《Robust Real-time Object Detection》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、SECONDINTERNATIONALWORKSHOPONSTATISTICALANDCOMPUTATIONALTHEORIESOFVISION–MODELING,LEARNING,COMPUTING,ANDSAMPLINGVANCOUVER,CANADA,JULY13,2001.RobustReal-timeObjectDetectionPaulViolaMichaelJonesviola@merl.commjones@crl.dec.comMitsubishiElectricResearchLabsCompaqCRL201Broadway,8
2、thFLOneCambridgeCenterCambridge,MA02139Cambridge,MA02142AbstractThispaperdescribesavisualobjectdetectionframeworkthatiscapableofprocessingimagesextremelyrapidlywhileachievinghighdetectionrates.Therearethreekeycontributions.Thefirstistheintroductionofanewimagerepresentationcall
3、edthe“IntegralImage”whichallowsthefeaturesusedbyourdetectortobecomputedveryquickly.Thesecondisalearningalgorithm,basedonAdaBoost,whichselectsasmallnumberofcriticalvisualfeaturesandyieldsextremelyefficientclassifiers[6].Thethirdcontributionisamethodforcombiningclassifiersina“casc
4、ade”whichallowsbackgroundregionsoftheimagetobequicklydiscardedwhilespendingmorecomputationonpromisingobject-likeregions.Asetofexperimentsinthedomainoffacedetectionarepresented.Thesystemyieldsfacedetectionperformacecomparabletothebestprevioussystems[18,13,16,12,1].Implementedo
5、naconventionaldesktop,facedetectionproceedsat15framespersecond.1.IntroductionThispaperbringstogethernewalgorithmsandinsightstoconstructaframeworkforrobustandextremelyrapidobjectdetection.Thisframeworkisdemonstratedon,andinpartmotivatedby,thetaskoffacedetection.Towardthisendwe
6、haveconstructedafrontalfacedetectionsystemwhichachievesdetectionandfalsepositiverateswhichareequivalenttothebestpublishedresults[18,13,16,12,1].Thisfacedetectionsystemismostclearlydistinguishedfrompreviousapproachesinitsabilitytodetectfacesextremelyrapidly.Operatingon384by288
7、pixelimages,facesaredetectedat15framespersecondonaconventional7001MHzIntelPentiumIII.Inotherfacedetectionsystems,auxiliaryinformation,suchasimagedifferencesinvideosequences,orpixelcolorincolorimages,havebeenusedtoachievehighframerates.Oursystemachieveshighframeratesworkingonl
8、ywiththeinformationpresentinasinglegreyscaleimage.Thesealternativeso
