发表在CVPR13上的行人检测方法
ModelingMutualVisibilityRelationshipinPedestrianDetection
WanliOuyang1,2XingyuZeng2XiaogangWang1,2
1
ShenzhenkeylabofComp.Vis.&Pat.Rec.,
ShenzhenInstitutesofAdvancedTechnology,CAS,China
2
DepartmentofElectronicEngineering,TheChineseUniversityofHongKong
wlouyang@ee.cuhk.edu.hk,xgwang@ee.cuhk.edu.hk
Abstract
Detectingpedestriansinclutteredscenesisachal-lengingproblemincomputervision.Thedif cultyisaddedwhenseveralpedestriansoverlapinimagesandoc-cludeeachother.Weobserve,however,thattheocclu-sion/visibilitystatusesofoverlappingpedestriansprovideusefulmutualrelationshipforvisibilityestimation-thevis-ibilityestimationofonepedestrianfacilitatesthevisibil-ityestimationofanother.Inthispaper,http://www.77cn.com.cnparedwithexistingimage-basedpedestriandetectionapproaches,ourapproachhasthelowestaveragemissrateontheCaltech-Traindataset,theCaltech-TestdatasetandtheETHdataset.Includingmutualvisibilityleadsto4% 8%improvementsonmultiplebenchmarkdatasets.
Without mutual
visibility
Compatible
(a)
(b)
1.Introduction
Pedestriandetectionisachallengingtaskduetotheintra-classvariationofpedestriansinclothingandarticula-tion.Whenseveralpedestriansoverlapintheimageregion,somewillbeoccludedbyothersandtheexpectedvisualcuesoftheoccludedpartsarecorrupted,resultingintheaddeddif cultyindetection.TheexamplesofoverlappingpedestriansinFig.1(a)oftenappearinrealworldapplica-tions.
Pedestrianswithoverlapsaredif culttodetect,however,weobservethatthesepedestrianshaveusefulmutualvisi-bilityrelationshipinformation.Whenpedestriansarefoundtooverlapintheimageregion,therearetwotypesofmutualvisibilityrelationshipsamongtheirparts:
http://www.77cn.com.cnpatiblerelationship.Itmeansthattheobservationofonepartisapositiveindicationoftheotherpart.
There
1
Figure1.(a)Mutualvisibilityrelationshipofpartsamongpedes-triansand(b)detectionresultscomparisonoftheapproachwith-outmodelingmutualvisibilityin[25]andourapproachmodelingmutualvisibility.Withmutualvisibilitymodeledinourapproach,thefalsepositivewindowontheleftlegissuppressedandmissedpedestrianontheleftisfoundbymodelingthevisibilityrelation-shipamongparts.
aretwoparts,i.e.left-halfpartandright-halfpart,foreachpedestrianinFig.1(a).InFig.1(a),giventhepriorknowl-edgethattherearetwopedestrianco-existingsidebyside,theright-halfpartoftheleftpedestrianiscompatiblewiththeright-halfpartoftherightpedestrianbecausethesetwopartsoftenco-existinpositivetrainingexamples.Thecom-patiblerelationshipcanbeusedforincreasingthevisibilitycon denceofmutuallycompatiblepedestrianparts.TakeFig.1(b)asanexample,ifapedestriandetectordetectsbothAlice1ontheleftandBobontherightwithhighfalsepositiverate,thenthevisibilitycon denceofAlice’sright-1’Alice’
and’Bob’areusedasplaceholdernamesinthispaper.