EvaluationofSLAMAlgorithmsinRealisticSensorTestConditionsLanceFangBachelorofEngineering(AerospaceEngineering)(Honours)(RMITUniversity)AthesissubmittedinfulfilmentoftherequirementsforthedegreeofMasterofEngineeringSchoolofEngineeringCollegeofScience,EngineeringandHealthRMITUniversityMelbourne,AustraliaMarch,2018
DeclarationIcertifythatexceptwheredueacknowledgementhasbeenmade,theworkisthatoftheauthoralone;theworkhasnotbeensubmittedpreviously,inwholeorinpart,toqualifyforanyotheracademicaward;thecontentofthethesisistheresultofworkwhichhasbeencarriedoutsincetheofficialcommencementdateoftheapprovedresearchprogram;anyeditorialwork,paidorunpaid,carriedoutbyathirdpartyisacknowledged;and,ethicsproceduresandguidelineshavebeenfollowed.IacknowledgethesupportIhavereceivedformyresearchthroughtheprovisionofanAustralianGovernmentResearchTrainingProgramScholarship.Signed,LanceFangMarch,20181
AcknowledgementWithoutthesupportandhelpfromavarietyofpeople,thisthesiswouldnothavebeenpossible.Iwouldliketoextendmythankstomysupervisors,Dr.AlexFisher,Dr.ReeceClothier,andDr.GrahamWild,astheyprovidedacontinuousamountofsupport,motivation,andpatience.Withouttheirsupport,thisthesiswouldnothavebeenpossible.IwouldalsoliketothankDr.JenniferPalmerandDr.ChaturaNagahawattefortheirsupportbyprovidingthetestenvironmentaswellasmotivation.Duetothegreattimesandforbeingatravelbuddy,sincerestthankstoAshimPanta,withoutyouthereEuropewouldhavebeenboring!MythankstoChungSingLeung,AndrewTedja,andRyanVuforprovidingstress-freenightoutswhereIcouldforgetaboutstudy.IwouldalsoliketothankmycolleaguesandfriendsatRMIT,NicolaKloet,SamPrudden,MattMarino,DavidTennent,ThomasNewnham,JamesKennedy,andtheRMITUASResearchTeamfortheirsupportandgreatsocialexperience.Iwishtoextendmygreatestthankstomyfamily,inwhichsupportandlovewasgiven.Withouttheirsupport,Iwouldhaveneverbeenabletoreachtheendgoal.Ialsowouldliketoextendmythankstomygirlfriend,JessicaBeitz.Throughthetoughestoftimes,youwerealwaystheshiningbeaconthatgavemehope.ThisresearchissupportedbyDST’sStrategicResearchInitiativeonTrustedAutonomousSystems(ProgramTyche).TheauthorswishtoacknowledgeMrMattClementeforhisassistanceintheconductoftheexperiments.IwouldliketothankallthesupportandstaffatDSTfortheircontinuedsupportthroughtheMasterdegree.2
AbstractAutonomousroboticsystemsrelyonSimultaneousLocalisationandMapping(SLAM)algorithmsthatuserangingorothersensorydataasinputtocreateamapoftheenvironment.Numerousalgorithmshavebeendevelopedanddemonstrated,manyofwhichutilisedatafromhigh-precisionranginginstruments.SmallUn-mannedAircraftSystems(UAS)havesignificantrestrictionsonthesizeandweightofsensorstheycancarry,andlight-weightrangingsensorstendtobesubjecttogreatererrorthantheirlargercounterparts.TheeffectoftheseerrorsonthemappingcapabilitiesofSLAMalgorithmswilldependonthecombinationofalgorithmandsensor.Toquantitativelydeterminethequalityofthemap,amapqualitymetricisneeded.ThisthesispresentsanevaluationofthemappingperformanceofavarietyofSLAMalgorithmsthatarefreelyavailableintheRobotOperatingSystem(ROS),inconjunctionwithrangingdatafromvariousrangingsensorssuitableforuseonboardsmallUAS.Tocomparethequalityofthegeneratedmaps,anexistingmetricwasinitiallyemployed,howeverdeficienciesnotedinthismetricledtothedevelopmentoftwonewmetrics.Adiscussionofboththeexistingandnewmapqualitymetrics,andtheadvantagesanddisadvantagesofeach,ispresentedaspartofthisthesis.Toevaluatetheperformanceofalgorithm/sensorcombinations,rangingdatawascollectedfromvarioussensorsinaknownenvironment.Bothsensorposesandthegroundtruthmapwereobtainedusingahighly-accuratemotioncapturesystem.ThemeasuredsensorposeswerethencorruptedwithnoiseanddrifttosimulateodometrymeasurementsrequiredfortheSLAMalgorithms.OftheSLAMalgorithmstested,Gmappingwasfoundtoproducehighqualitymapswithwide-field-of-regardrangesensorsinthepresenceofodometrynoiseanddrift.KartoSLAMproducedsimilarmapstoGmapping(withwidefieldofregardsensors),thoughitdidnotcopeaswellwithodometryerrors.HectorMappingtendstoexcelatcreatingmapswithwidefieldofregardrangingsensors.Keywords:SLAM,ROS,ranging,mapquality3
Contents1Introduction131.1ThesisLayout...............................................152SLAMalgorithmsandRangingsensors162.1SimultaneousLocalisationandMappingAlgorithms.........................162.1.0.1ScanMatching.....................................172.1.0.2LoopClosure......................................182.1.1CommonSLAMtechniques...................................192.1.1.1ParticleFilterSLAMalgorithms...........................192.1.1.2Graph-basedSLAMalgorithms............................202.2MapComparisonMethods........................................202.3RangingSensors.............................................212.3.1Time-of-Flight..........................................212.3.2RotatinglaservsScenecapturingToFsensors........................232.3.2.1Rotating-laserToFsensors..............................232.3.2.2Scene-capturingToFsensors.............................242.3.3VariousToFSensorsystems...................................252.3.3.1TheSingle-PhotonAvalancheDiodeSensor.....................252.3.3.2OtherTime-of-Flightpixels/sensors.........................252.3.3.3LaserRangefinders...................................262.3.4Ultrasonictechniques......................................262.3.5StereoVisionBasedtechniques.................................262.3.6StructuredLighttechniques...................................272.4RangingsensorandSLAMalgorithmcombinations..........................282.4.1Hypothesis............................................282.5Errorsencounteredfromsensordata..................................282.5.1ToFsensors............................................292.5.1.1Multi-pathing.....................................292.5.1.1.1Scenecapturingsensors...........................292.5.1.1.2Rotatinglasersensors............................304