[CVPR 2012] Contextual Boost for Pedestrian Detection

《[CVPR 2012] Contextual Boost for Pedestrian 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、ContextualBoostforPedestrianDetectionYuanyuanDingandJingXiaoEpsonResearchandDevelopment,Inc.214DevconDrive,SanJose,CA95112{yding,xiaoj}@erd.epson.comAbstract[6,17,28,10,45,44,31,36,18,1].Todeterminewhetheralocalwindowboundingahumanfigure,bothgenera-Pedestriandetectionfromima

2、gesisanimportantandyettiveanddiscriminativeapproacheshavebeendevelopedchallengingtask.Theconventionalmethodsusuallyidentify[23].Thegenerativeapproachesinfertheposteriorprob-humanfiguresusingimagefeaturesinsidethelocalregions.abilityforpedestrianclassusingdiscrete[22]orcontin

3、uousInthispaperwepresentthat,besidesthelocalfeatures,[5,13,49]shapemodels,orcombiningshapeandtexturecontextcuesintheneighborhoodprovideimportantcon-models[26,14].Thediscriminativeapproachesextractim-straintsthatarenotyetwellutilized.Weproposeaframe-agefeaturesinthelocalwind

4、owandconstructclassifiersworktoincorporatethecontextconstraintsfordetection.fordetection.ForthispurposevariousfeatureshavebeenFirst,wecombinethelocalwindowwithneighborhoodwin-proposed,suchasHaarwaveletfeatures[42],gradientbaseddowstoconstructamulti-scaleimagecontextdescripto

5、r,features[6],shapebasedfeatures[34],combinationofmul-designedtorepresentthecontextualcuesinspatial,scaling,tiplefeatures[44,45],automaticallyminedfeatures[10],andcolorspaces.Second,wedevelopaniterativeclassifi-orpose-invariantfeatures[27].Thelocalfeaturesarethencationalgori

6、thmcalledcontextualboost.Ateachiteration,usedtoidentifypedestriansintheclassificationprocessbytheclassifierresponsesfromthepreviousiterationacrossalgorithmssuchasAdaBoost[20].Intheliteraturesthistheneighborhoodandmultipleimagescales,calledclas-processhasbeeneithertargetedatth

7、ehumanfigureasonesificationcontext,areincorporatedasadditionalfeaturesobjectorbasedonpartdetectors.Thepart-basedmethodstolearnanewclassifier.Thenumberofiterationsisde-[15,47,19]treatthehumanfigureasanassemblyofdif-terminedinthetrainingprocesswhentheerrorratecon-ferentbodyparts.

8、Theyrundetectiononindividualpartsverges.Sincetheclassificationcontextincorporatesco