[2009CVPR]Feature based person detection beyond the visible spectrum

《[2009CVPR]Feature based person detection beyond the visible spectrum》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、FeaturebasedpersondetectionbeyondthevisiblespectrumKaiJunglingandMichaelArens¨FGAN–FOM76275Ettlingen,Germanyfjuengling,arensg@fom.fgan.deAbstractproachisabletodistinguishbetweenobjectclasses,whichisimportantsincewearenotinterestedinanymovingob-Oneofthemainchallengesincompute

2、rvisionistheau-ject,likecars(likeamotiondetectionwouldreturniteventomaticdetectionofspecificobjectclassesinimages.Re-inanenvironmentwithastaticcamera),butonlyinpeople.centadvancesofobjectdetectionperformanceinthevisibleOurpersondetectorisbuildonastate-of-the-artfea-spectrumen

3、couragetheapplicationoftheseapproachestoturebasedobjectdetector,introducedbyLeibeetal.in[8].databeyondthevisiblespectrum.Inthispaper,weshowtheWemakeseveralenhancementstothisobjectdetectorandapplicabilityofawellknown,local-featurebasedobjectde-adaptittothespecifictaskofdetecti

4、ngpeopleinthermaltectorforthecaseofpeopledetectioninthermaldata.Wedata.Theapproachoflocal-featurebasedpersondetectionadaptthedetectortothespecialconditionsofinfrareddataisspecificallyapplicableinthermaldata,becausethevaria-andshowthespecificsrelevantforfeaturebasedobjectde-tio

5、ninpersonappearanceisratherlimitedcomparedtovisi-tection.Forthat,weemploytheSURFfeaturedetectorandblewavelengthimagery.Inthermaldata,peoplemostlyap-descriptorthatiswellsuitedforinfrareddata.Weevaluatepearasalightregionwithacontrasttoadarkerbackground.theperformanceofouradapt

6、edobjectdetectorinthetaskAdditionallytojustdetectingpersonsasacompound,weofpersondetectionindifferentreal-worldscenarioswhereshowhowthisfeaturebasedpersondetectorcanbeusedtopeopleoccuratmultiplescales.Finally,weshowhowthisclassifybody-partsofpersons,whichcanbeusedforfur-loca

7、l-featurebaseddetectorcanbeusedtorecognizespe-therinterpretationonthebehaviorofpeople.Weevaluatecificobjectparts,i.e.,bodypartsofdetectedpeople.thepersondetectorinthreethermalimagesequenceswithatotalof2535personoccurrences.Theseimagesequencescoverthecompleterangeofdifficulties

8、inpersondetection,1.Introductioni.e.,peopleappearingatdifferentscales,visib