Fast-SeqSLAM A Fast Appearance Based Place Recognition Algorithm

《Fast-SeqSLAM A Fast Appearance Based Place Recognition Algorithm》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、2017IEEEInternationalConferenceonRoboticsandAutomation(ICRA)Singapore,May29-June3,2017Fast-SeqSLAM:AFastAppearanceBasedPlaceRecognitionAlgorithmSayemMohammadSiamandHongZhangAbstract—Loopclosuredetectionorplacerecognitionisahasdemonstratedbetterperformanceinrecogni

2、zingplacesfundamentalprobleminrobotsimultaneouslocalizationandthatunderwentsevereappearancechangesthantheothermapping(SLAM).SeqSLAMisconsideredtobeoneofthemostsuccessfulSLAMalgorithmslikeFAB-MAP2.0[5].Mostsuccessfulalgorithmsforloopclosuredetectionasithasbeenrecen

3、tly,MilfordandShen[5]havedemonstratedsignificantdemonstratedtobeabletohandlesignificantenvironmentalconditionchangesincludingthoseduetoillumination,weather,improvementoftheviewpointinvarianceofSeqSLAMbyandtimeoftheday.However,SeqSLAMreliesheavilyongeneratingsyntheti

4、cviewpointsusingthestate-of-the-artexhaustivesequencematching,acomputationallyexpensivedeeplearningtechniques[6].However,SeqSLAM,whichprocessthatpreventsthealgorithmfrombeingusedindealingisbasedonexhaustivesequencesearch,iscomputationallywithlargemaps.Inthispaper,

5、weproposeFast-SeqSLAM,costlywhenthenumberofimagesornodesinthemapisanefficientversionofSeqSLAM.Fast-SeqSLAMhasamuchreducedtimecomplexitywithoutdegradingtheaccuracy,andlarge.Thisisunfortunateasarobotneedstoperformallthethisisachievedbyusinganapproximatenearestneighbo

6、rcomputationforlocalizationincludingloopclosuredetection(ANN)algorithmtomatchthecurrentimagewiththoseonlineinaSLAMsystem.intherobotmapandextendingtheideaofSeqSLAMtoThekeymotivationforourworkistodevelopaplacegreedilysearchasequenceofimagesthatbestmatchwithrecogniti

7、onalgorithmforalargescaleandlongtermop-thecurrentsequence.WedemonstratetheeffectivenessofourFast-SeqSLAMalgorithminappearancebasedloopclosureeration,basedonsequencematchinglikeSeqSLAMbutdetection.computationallyefficient.Toachieveahighcomputationalefficiency,wefirsts

8、toretheimagedescriptorsofamapI.INTRODUCTIONinatreestructure.ThenweuseanapproximatenearestThecomputationtaskofbuildingorupdatingamapandneighbor(ANN)algor