Online learning of robust object detectors during unstable tracking

《Online learning of robust object detectors during unstable tracking》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、OnlinelearningofrobustobjectdetectorsduringunstabletrackingZdenekKalalJiriMatasKrystianMikolajczykUniversityofSurreyCzechTechnicalUniversityUniversityofSurreyGuildford,UKPrague,CzechRepublicGuildford,UKz.kalal@surrey.ac.ukmatas@cmp.felk.cvut.czk.mikolajczyk@surrey.ac.ukAbstractdressdire

2、ctlythepost-failurebehaviorandthereforecannotbedirectlyusedinthelong-termtrackingproblem.WereferThisworkinvestigatestheproblemofrobust,long-tothisgroupofalgorithmsasshort-termtrackers.termvisualtrackingofunknownobjectsinunconstrainedClearlythesolutionofthelong-termtrackingproblemenviron

3、ments.Itthereforemustcopewithframe-cuts,requiressomedetectioncapability,tore-detecttheobjectfastcameramovementsandpartial/totalobjectocclu-afteraperiodwhenitisnotinthefieldofvieworaf-sions/dissapearances.Weproposeanewapproach,tertrackingfailure.Tracking-by-detectionmethods[9]orcalledTrac

4、king-Modeling-Detection(TMD)thatcloselymethodsintegratingatrackerandadetector[1,17]addressintegratesadaptivetrackingwithonlinelearningofthetheproblem.However,detectorshavetobedesignedorobject-specificdetector.Startingfromasingleclickinthetrainedbeforetrackingstartsandthuscannotbeusedwhen

5、firstframe,TMDtrackstheselectedobjectbyanadaptivetheobjectofinterestisnotknowninadvance.Thetrain-tracker.Thetrajectoryisobservedbytwoprocesses(grow-ingofthesedetectorseitherrequiresalargehand-labeledingandpruningevent)thatrobustlymodeltheappearancetrainingsets[20],generatesthetrainingset

6、bywarpingtheandbuildanobjectdetectoronthefly.Botheventsmakepatches[9]orextractsthetrainingdatausingsomesophis-errors,thestabilityofthesystemisachievedbytheircan-ticatedmethod[16,18,19].Allthesemethodsstrictlysep-celation.Thelearntdetectorenablesre-initializationofaratethetrainingandtesti

7、ngphasewhichmeansthatap-thetrackerwheneverpreviouslyobservedappearancere-pearancevariabilitynotrepresentedinthetrainingsetneveroccurs.Weshowthereal-timelearningandclassificationisbecomespartofthemodel.achievablewithrandomforests.TheperformanceandtheTheappearancechangeproblemisad