发表在CVPR13上的行人检测方法
Figure7.ExperimentalresultsontheCaltech-Testdataset(left)andthePETS2009dataset(right).
groundtruthpedestrianisconsideredasoverlappediftheoverlappingarearatioofhis/herboundingboxwithotherpedestrianislargerthan1%.PedestriansoverlappingwithotherpedestriansarecalledO-pedestriansinthispaper.TheO-pedestriansareinsubsetOverlapped.IsolatedPedestri-ansareinsubsetIsolatedandcalledI-pedestrians.Ourap-proachfocusesonO-pedestrians.NotethatanO-pedestrianisstillapedestrian.Fig.9showsexamplesonI-pedestrianandO-pedestrian.
Inthissection,wespeci callyevaluatethedetectionper-formanceondetectingI-pedestriansandO-pedestrians.Fig.10showstheexperimentalresults.TheimprovementofourapproachoverD-Isolis2%ondetectingI-pedestriansforbothdatasets.Ontheotherhand,theimprovementonde-tectingO-pedestriansis5%fortheCaltech-Traindatasetand7%fortheETHdataset.Thusourapproachout-performsD-Isolonpedestriandetectionmainlybecauseofitsabilityindetectingco-existingpedestrians,i.e.O-pedestrians.Thisexperimentalresultvalidatesthefocusofthispaperinimprovingtheperformanceondetectingpedes-trianswithoverlaps.
An I-pedestrian3 O-pedestrians
Figure9.I-pedestrianversus
O-pedestrian.
Figure10.ExperimentalresultsondetectingIsolatedpedestrians(top)andOverlappedpedestrians(bottom)ontheCaltech-Traindataset(left)andtheETHdataset(right).
6.Conclusion
Thispaperproposesamutualvisibilitydeepmodelthatjointlyestimatesthevisibilitystatusesofmultipleco-existingpedestrians.Startingfromthescoresofconven-tionalpartdetectors,http://www.77cn.com.cnparedwithexistingimage-basedpedestriandetectionapproacheseval-uatedin[9],ourapproachachievesthelowestmissrateontheCaltech-Traindataset,theCaltech-TestdatasetandtheETHdataset.Sincethedeepmodeltakesthepartdetectionscoresasinput,itiscomplementarytonewinvestigationsonfeatures,e.g.colorselfsimilarity,localbinarypattern,motionanddepth,andarticulationmodels,e.g.poselets,multi-objectarticulationmodel.Experimentalresultsonfourpubliclyavailabledatasetsshowthatthemutualvisi-bilitydeepmodeliseffectiveinimprovingpedestriande-tectionresults.
Acknowledgment:ThisworkissupportedbytheGen-eralResearchFundsponsoredbytheResearchGrantsCouncilofHongKong(ProjectNo.CUHK417110andCUHK417011),NationalNaturalScienceFoundationof