Report - Microgeo
Transcript
Report - Microgeo
Measurement accuracy of Lidarbased SLAM systems Kaarta Report 001 Abstract We determine the accuracy of a lidar-based mapping system by determining differences in point cloud measurements. We use two methods to determine accuracy. The first method involves computing derived points and the second method compares point clouds by first aligning the point clouds. The results are that the lidar-based SLAM system does very well in comparison to stationary measurements with a high-end lidar. September, 2016 Kaarta 5001 Baum Blvd, Suite 430, Pittsburgh, PA 15213 1 KAARTA Table Of Contents Overview..............................................................................................................................3 Background.........................................................................................................................3 Accuracy Specifications and Factory Calibration........................................................4 Assessing Accuracy....................................................................................................................4 A Comparison of Two Lidar-based Systems.................................................................5 Acquiring the Point Clouds........................................................................................................5 Derived-Point to Derived-Point Measurements......................................................................6 Comparing Point Clouds...........................................................................................................10 Summary............................................................................................................................14 References............................................................................................................................15 CopyrightKaarta2016 2 KAARTA Overview What is the accuracy of a 3D mapping system? Manufacturers often state the accuracy of the system is the same as the sensor used to collect the data, but this is only part of the answer since subsequent calculations affect the answer. Even the geometry and surfaces of the local environment can affect the measurements. Also, accuracy is often confused with precision. They aren’t the same. In this paper, we examine the problem, define accuracy, and provide some answers. Background Real Earth has developed a mobile 3D mapping system that creates a map of its environment even while moving continuously. It creates the map using only the range measurements it acquires along the way and a low cost MEMS IMU. No GPS is required and there is no prior map or other environmental infrastructure required. The accuracy of our point cloud-based 3D mapping system is determined by both the accuracy and precision of the sensing device and by the software used to create the map. Since a global 3D map is computed by incorporating a sequence of local maps, any noise or inaccuracy in the data propagates into downstream calculations. Figure1:TheStencilproductfromRealEarthisontheleftandtheFAROFocus3DX330ontheright.Relativesizeisnottoscale. Therobustnessoftheprocessformergingpartiallyoverlappingpointcloudsdependsonthepresenceof strongfeaturessuchasedgesintheenvironmenttoestablishcorrespondencesbetweenthepoint clouds.Ifthesensorismovingwhileacquiringpointclouds,thealgorithmsusesensormeasurementsto maptheenvironmentandlocatethesensorwithinthatmap.Abenefitofdeterminingthesensor positionisdeterminingtheoutwarddirectionofthesurfacenormal.FortheRealEarthproducts,every pointmeasurementinthecloudismadewiththesensorinadifferentpositionandorientation.This processforanalyzingtherangedatatobuildamapanddeterminelocalizationisknownassimultaneous localizationandmappingorSLAM. CopyrightKaarta2016 3 KAARTA Variouscalculationsandassumptionsaboutmeasurementerrorareincludedwhencreatingthe resultantmap.Conditionsaffectingdatafromrealworldmeasurementsincludelighting,atmospheric effects,environment,surfacematerialsandtextures,and,inthecaseoflidar,laserbeamdivergence. Divergenceformsalargerspotatgreaterdistances.Ononehandtheresolutionandprecisionofthe instrumentmaybeverygood;ontheotherhandthereisavastamountofdata,hundredsofthousands ofpoints,whicharelow-passfilteredtomitigatethenoise.Theaccuracyspecificationoftheinstrument isonlyastartingpointinunderstandingtheaccuracyofthemeasurementprocess. Accuracy Specifications and Factory Calibration Alidartypicallyproducesdataintheformat<angle><angle><range>.Theremayalsobereflectancedata and,insomecases,achoiceofsignalorfirstreturn.Theseareusuallydifferent.Factorycalibration definesthelocationandorientationofaninstrumentcoordinatesysteminsideofthelidarsensorhead. Thedirectionandrangetoatargetinthatcoordinatesystemisdeterminedfromencoderpositionsthat definethelaserbeamdirection.Ifthecalibrationisaccurate,scannedflatsurfaceswillbemeasuredto beflat,andwewillseeaccuratemeasurementsofdistancesbetweenderivedpointssuchascorners, radiiofcurvature,distancesbetweenparallelsurfaces,andtheanglesbetweenflatsurfaces,etc. However,rangenoisecanpropagateintofittedparameterssuchassurfaceorientationorsurface curvature.This,inturn,leadstouncertaintyintheresult.Theresolutionneededtodetectafeatureisa separatetopic. Factorycalibrationisstatic,thoughthesensormaybemovedbetweenmeasurements.SincetheReal Earthproductprocessesastreamofrangemeasurementswhilemoving,theaccuracyoftheresultant mapneedstobecomparedwithgroundtruthinameaningfulway. Tomeasuretheaccuracy,wecomparedoutputsoftwolidarunits.OneunitistheRealEarthStencil productandtheotherisaFAROunit.WeusedaFAROFocus3DX330lidartodefinegroundtruth.To assessaccuracy,wecomparedderived-pointtoderived-pointmeasurements,andthencomparedthe pointcloudsgeneratedbythetwosystems.ClearlytheFARO,whileanexcellentinstrument,isnottrue groundtruth–ittoohasameasureableaccuracythatthemanufacturerprovides. Assessing Accuracy Theaccuracyofanysensorisdeterminedincomparisontogroundtruth.Importantly,withapoint cloudsystem,specificfeaturepointsonasurfacearenotmeasured,butderived.Inotherwords,the laserbeamfromalaserrangefinderorarayfromacamerapixeldoesnotintersectthesurfaceata specificallytargetedlocation.Thus,derivedpointsarecomputed,notdirectlymeasured.Examplesof derivedpointsincludethecenterofaspherewhichiscomputedfromdatapointscollectedfromthe surfaceofthesphere,orofthelocationofacornerwhichiscomputedfromtheintersectionofthree planes.Onewaytodetermineaccuracyistocomparederived-pointtoderived-pointmeasurements. Anotherwaytoassessaccuracyistomeasuretheshapeofasurfacewithknowngeometry,andthen comparetheresultsagainstamodel.Examplesincludetheradiusofasphereorcylinder,ortheflatness ofaplane,ortherelativeorientationsoftwoplanes,oraCADmodel.However,creatinglargehigh accuracyshapesofknowngeometryisverycostly. CopyrightKaarta2016 4 KAARTA Pointcloudscanbecomparedwithcarefullysurveyedpointsacquiredusingexpensiveandsophisticated equipment.Oftentheoutputfromthesedevicesisacceptedasareferencestandard,butthese measurementsalsohaveerrorsandnoiseassociatedwiththem.However,thesemeasurementscanbe usedasgroundtruthiftheiraccuracyisatleastanorderofmagnitudebetterthanthesystemunder test. Accuracyisoftenconfusedwithprecision,repeatabilityandotherterms.Itisimportanttodefinethe termswhenhavingaconversation.Measurementaccuracy,byinternationallyaccepteddefinition,is theclosenessofagreementbetweenameasuredquantityvalueandatruequantityvalueofa measurand.1Thetrickyparthereistheterm‘truequantity’becauseitassumesgroundtruth.Precision istherepeatabilityorreproducibilityofthemeasurement. Inthefollowingdiscussion,groundtruthisdefinedbyaFAROFocus3DX330,showninFigure1,whichis alsoapointcloudsystem.Thislidarsystemispositionedatfixedlocationstoscanitsenvironment.Asa result,itoperatesinitscalibratedmodeateachpositionitisplaced.Thehigh-densitypointcloudit acquiresatonepositionisthenregisteredandmergedwiththepointcloudacquiredataneighboring position.Registrationofthepointcloudscanspecificallytakeadvantageofhand-placedfiducial featuresintheenvironmenttoensurehighaccuracy. A Comparison of Two Lidar-based Systems Toassessaccuracy,wemeasuredanaturalenvironmentthatincludedman-madestructures.Wemade measurementsusingtwodifferentLidar-basedsystemsandcomparedtheresults.TheRealEarth StencilproductincorporatesaVelodyneVLP-16Lidarsystemthatsimultaneouslyscans16rotatinglines oriented1.875oapartwhiletheusermanuallycarriestheunitwhilemovingtoacquiredensedatain3D. TheStencilisamobilemappingsystemthatproduces3Dmapsinreal-timewithoutpost-processing.The RealEarthsystemneedsnoGPS,noinfrastructure,oranypriormapstobuilda3Dmodel. Thesecondsystem,aFAROFocus3DX330,wasusedasthereferencetodefinegroundtruth.Itacquires datawhilescanningitsenvironmentfromastaticpositionusingarotatingmechanism.Thissystemis phase-basedand,alongwithsimilarproductsfromReiglandLeica,producesadensepointcloudwhose accuracyisconsideredtobeof‘surveyquality’,whichisgenerallydefinedtobe5mmorless. RangenoiseoftheFAROinstrumentusingalowreflectivitytargetis2.2mmat25meters. MeasurementaccuracyofthescannerusedintheStencilproductis3cmupto100meters.Thisisan orderofmagnitudelessthantheFAROscannerwhosemeasurementsarethereforeconsideredthe reference. Acquiring the Point Clouds PointclouddatawascapturedusingtheFAROdeviceandRealEarth’sStencilsystem.TheFAROsystem wastripod-mountedatfivefixedpositionstoobtainfiveindependentpointclouds.Thescanswere 1 JCGM 200:2012. International vocabulary of metrology: Basic and general concepts and associated terms (VIM). France: BIPM; 2012. CopyrightKaarta2016 5 KAARTA madefromrotating360oaroundaverticalaxis.Thepointcloudswereregisteredtoeachotherina commoncoordinatesystemusingFARO’sproprietarysoftwarewithsphericalmarkersplacedinthe sceneforfurtheranalysis. Whenusingthehand-heldStencilsystem,theoperatorsimplywalkedaroundthenaturalscenewhile collectingthepointcloud.Asweepconsistsofasetofrangedatainsixteenplanespassingthroughthe originoftheinstrument’scoordinatesystem.Thecomplicationisthattheuseriscontinuously “painting”thescenebothbymovinghorizontallyandtiltingthescannervertically,toacquireafullscene of3Ddata.Thealgorithmsusedforthistakethemotionintoaccountthroughrapidfeaturetrackingand areassistedbyanintegratedIMUinthismatter.TheIMUdataisusedtounwarpeachlidarscanandto provideanotionofrelativemotiontothescanmatchingalgorithmsandthefinalaccuracyisnotbased ontheaccuracyoftheIMUitself. Thescannedareaisofacabinofapproximately60squaremeters(10.5mX5.5m),witharoofandwalls. Figure2showsthescenegeneratedfromthepointclouddata.Thecabinwasonrelativelylevelground surroundedbytreesandbrush.Asphericalmarkerwaslocatedseveralmetersawayfromeachcorner ofthecabin.Theappearanceofthesphericalmarkersprovidedawaytocomparederived-pointto derivedpointdistances.Inaddition,sincethemarkersexistedinbothdatasets,theyprovidedawayto alignandcomparetheRealEarthandFAROpointclouds.Rapidlyandaccuratelyaligningtwopoint cloudswasusefulformanuallyeditingbothpointcloudsatthesametimetoselectfeaturesofinterest. Inourevaluation,wecomparedthepointcloudsgeneratedbythehand-heldStencil,whichwas continuouslymoving,andatripod-mountedFAROsystemwhichneededtoberelocatedtoseveral positions.Thescantimetoacquirethepointcloudswasabout10minutesfortheStenciland1hourfor theFAROdevice. TheFAROpointcloudisusedasareference.Tocomparepointclouds,wemeasureddistancesfromthe StencilpointcloudtotheFAROpointcloud.Thisisanalogoustomeasuringthedistancefromapointto asurface.SmallsurfacepatchesarecomputedforlocalgroupsofpointswithintheFAROpointcloud, anddistancesarecomputedtothesepatchesfromindividualpointsintheStencilpointcloud.See Figure5.SincethelocationofStencilrelativetothesurfaceisfound,theerrorsare“signed”.Errors outsidethesurfaceoftheobjectarepositive,whileerrorsinsidearenegative. Derived-Point to Derived-Point Measurements Onewaytoassessaccuracyistocomparethedistancesbetweenderivedpointssuchasthecentersof spheresortheintersectionsofthreeplanes.Wedidthisforthespherespresentinthescene,andthe cornersofthecabinatafixedheight.Figure2showsthelocationsofthespheresinthescene. CopyrightKaarta2016 6 KAARTA Figure2:AviewofscansfromtheStencilandFAROsystems. Wethenalignedthepointcloudstomakethecomparisoneasier.ThepointcloudsfromtheStenciland FAROinstrumentswereroughlyalignedinCloudCompareusinganIterativeClosestPoint(ICP) algorithm.ICPminimizesthemeansquareerror.Then,man-madeobjectsinthescenewerekeptwhile allofthefoliageandtheflagontheflagpolewerecroppedfromthescene.Wethenalignedthe remainingpointcloudagain.Thealignmenttookplacewithoutrescalingsinceweassumethelidar sensorshavebeencalibratedandrangemeasurementsareaccurate.Ourresultsandthealignment showthiswasagoodassumption. Figure3showstheremainingman-madeobjectsinthescene.Wekeptseveralobjectsincludingthe cabin,theflagpole,bulletinboard,sign,flowerpots,spheres,andpointsfromthesideofacarinthe parkinglot.ThetrajectoryofStencilduringscanningisshownastheredtraceinthefigure.Sinceonly themeasuredsurfacecoordinatedatawasprovided,weusedtheclosestStencildatapoint(nearest neighbor)toeachCloudCompareFARObased"core"pointalongwiththeStencilpositiontodetermine thesignofthesurfacenormal.Theanglebetweenthesurfacenormalandvectorfromthesurfaceto thesensorhadtobelessthan90degrees;otherwiseweflippedthedirectionofthesurfacenormal. CopyrightKaarta2016 7 KAARTA Figure3:Man-madeobjectsinthescene.ThetrajectoryoftheStencilwhenmeasurementsweretakenisshowninred. ThedirectionoftheoutwardsurfacenormalinFigure3isinthegeneraldirectionofthesensor.Asa result,weknowthesignoftheerrorwhencomparingthepointclouds.Themergedpointcloudsfor theseobjectswerealignedandtheresultantmeansquaredistancebetweenthecloudswasdetermined tobelessthan1cm. Afteraligningthepointclouds,weidentifiedthelocationofeachsphericalmarker.Wefitspheresto thelocalcloudofdatatomeasuretheircenterandradius.Thespheresarelampglobeswitharadiusof 15.24mm(6inches).Theyaretranslucent,sothereissomeuncertaintyintheirmeasuredcentroidsand radii.Lowpassfilteringofthepointcloudalsoaffectsthesevalues.Ameasureofconfidencein determiningthelocationsofthespherecenterswastheRMS(rootmeansquare)errorofthefit.Once thecentersofthesphereswerelocatedwecomputedallofthemutualdistancesbetweenthespheres. Figure4showsthepointclouddatainthevicinityofeachsphereforbothsetsofdata.Extraneous pointsfromthegroundortripodalongwithinstrumentnoiseweresuppressedbyiterativelyfittinga spheretothedatawhileremovingoutliers.Thatis,wefitaspheretothedata,removedoutliers,andfit aspheretotheremainingpoints.Theoutliersarethebluepointsshowninthefigures.Theredpoints wereincludedinthefit.Apossiblealternativeistousearandomsampleconsensus(RANSAC)based algorithm.Theresultisthelocationofthecenterofeachsphere,alongwithitsradius.Wecomputed thedistancesbetweenthespheresusingthecomputedspherecenters. CopyrightKaarta2016 8 KAARTA Stencilpointcloudspheredata FAROpointcloudspheredata Figure4:Pointclouddatafromthespheres. Table1showstheresults.ThedifferencesbetweentheStencilandFAROmeasurementsoftheinter spheredistanceisseveralmillimeters. Side1-2 Side2-3 Side3-4 Side4-1 Stencil FARO DistanceError (meters) (meters) (millimeters) 10.323 10.318 5 15.349 15.351 -2 14.291 14.277 14 18.608 18.615 -7 Table1:DistancesbetweenthespheresareshowninmetersandthedifferencebetweentheStencilandFAROmeasurements areshowninmillimeters. Thedimensionsofthecabinwerealsoderivedfromthepointclouddata.Thecabinisaclassicstructure withagabledroofandfourwalls.Thefourcornersofthecabincanbeconsideredderived-points.The wallsofthecabinarenotperfectlyflatandtheyincluderecessedfeaturessuchaswindowsanddoors. Butthesedeviationsaresmallrelativetothesizeofthecabinandaremeasuredbybothsystems.Wefit aplanetoeachofthefourwallsofthecabinandthenestimatedthelocationsofthecorners approximatelyonemeterabovethegroundfromboththeStencilandFAROpointclouds.Figure5 showsarenderingofthewallswiththeStencilandFAROpointcloudsoverlaid. Figure5:Arenderingofjustthewallsofthecabin.Wefitplanestothefourwallsandthencomputedahorizontalcrosssection toidentifythefourcornerpointsofthecabin. Figure6showsthecoordinatesofthefourcornersofthecabinalongwiththecoordinatesofthefour spheres. CopyrightKaarta2016 9 KAARTA Figure6:Thelocationsofthefourcornersofthecabinareshownintherendering.ThewalllengthsmeasuredbytheFAROare showninTable2. Table2showstheestimatedwalllengthsmeasuredonemeterabovetheground.TheFAROestimates differby0.3%fromtheStencilestimates. Stencil FARO (meters) (meters) 10.414 10.458 5.538 5.554 10.418 10.463 5.553 5.563 Wall1 Wall2 Wall3 Wall4 LengthError (cm) -4.4 -1.6 -4.5 -1.0 Table2:Shownaretheestimateddifferencesinthewalllengthsofthecabinbasedontheintersectionoftwoadjacentwalls. Thethirdplaneneededtodefinethefourcornerpointsislocatedapproximatelyonemeterabovetheground. Comparing Point Clouds WethencomputedtheaveragedistancebetweenpointcloudsusingtheFAROpointcloudasthe reference.Todothis,asmallsurfacepatchisfittolocalgroupsofpointsfromtheFAROpointcloud. ThedistancefromnearbypointsintheStencilpointcloudtoeachsurfacepatchisthencomputed.The distancesaresignedsincethesurfacenormalpointsinthegeneraldirectionofwherethesensorwas locatedwhenthedatawasacquired.SeeFigure7. CopyrightKaarta2016 10 KAARTA Figure 7: The graphic shows how distance is measured between the Stencil and FARO point clouds. Source: Cloud Compare M3C2 documentation. A noise filter was applied to the point clouds from the scene with man-made structures and the point clouds were then aligned. The filter suppresses high frequencies associated with the sensor noise. It also rounds out sharp edges such as corners. Figure 8 shows a histogram of the errors. The shape of the histogram indicates there is a slight distortion of the Stencil point cloud relative to the FARO. The point clouds have been aligned to minimize the mean square error. Ideally, the errors seen in the histogram should be symmetric with an average of zero. When the point cloud shapes are in agreement the width of the histogram is proportional to the noise. The magnitude of the average error is approximately 7mm. CopyrightKaarta2016 11 KAARTA Figure8:Histogramofpointclouderrorsfortheman-madeobjectsinthearea.Thecaptionshowstheaverageerror,the standarddeviation,andthefullwidthathalfmaximumofthehistogram. Thepointcloudsforthecabinalonewerethenextractedfromthedata.Thehistogramoftheerrors wasfoundtobebimodalasseeninFigure9. Figure9:Histogramofpointclouderrorsinthedataforthecabinalone.Thecaptionshowstheaverageerror,thestandard deviation,andthefullwidthathalfmaximumofthehistogram. TounderstandtheshapeofthehistogramweusedafeatureofCloudComparetoadjustthescaleof datatoimprovethefitwiththeFAROpointcloud.Thisscalechangeisanon-rigidtransformation. AllowingtheStencilpointcloudtoberesizedby0.5%(afactorof1.005)resultedintheuni-modal histogramshowninFigure10.ThisscalechangeisequivalenttoincreasingeveryStencilrange measurementby0.5%,suggestingasmallerrorincalibrationorpointcloudalignment.Thechangeof scaleneededtoimprovethefitbetweenthetwopointcloudsisconsistentwiththedifferencesinthe lengthsofthewallswecomputedusingthederivedcornerpointsofthecabin.Thisisworth investigatingfurther.Notethatthescalingwasusedonlyforcomparingthehistograms.Itwasnotused toimprovetheerrordifferencesbetweentheFAROandStencilcitedearlier. CopyrightKaarta2016 12 KAARTA Figure10:HistogramofpointclouderrorsinthecabinwalldatausingFAROasthereferenceafteralignmentofthecabinitself. Thecaptionshowstheaverageerror,thestandarddeviation,andthefullwidthathalfmaximumofthehistogram. Thefullwidthathalfmax(FWHM)ofthehistogramis2.5cm,or+/-1.25cm.Thisisconsistentwiththe noisespecificationofthelaserscannerusedintheStencilafterfiltering.Ifthereisanerrorinthe measuredshapeofthecabin,notjustscale,itshouldshowupasabimodaldistributioninthis histogram. CopyrightKaarta2016 13 KAARTA Summary RealEarth’smappingsystemacquiresandprocessespointclouddatainrealtimetoassembleapoint cloudmapwithouttheuseofaGPS,maps,orinfrastructurewhileitisinmotion.Thesystemwas comparedagainstaFAROFocus3DX330system,whichacquirespointclouddatawhilestationary.The resultsshowdifferencesinthepointcloudsgeneratedbythetwosystemsontheorderof1to2cm overanareaof20squaremeters. Differencesinthepointcloudmeasurementsweredeterminedintwoways.Thefirstmethodinvolved computingderivedpoints.Thesceneincludedspheresandacabinwithcorners.Localpointclouds associatedwiththesphereswereextractedfromtheglobaldataandfittingroutineswereusedto estimatethecentersofthespheres.Thedistancesbetweenthespherecentersasmeasuredbythetwo deviceswerethencompared.Cornersofthecabinweredeterminedfromtheintersectionofplanar surfacescomputedfromthecabinpointclouds.Cornertocornerdistanceswerethencompared. Thesecondmethodusedtocomparepointcloudsfirstalignsthepointclouds.Thepointcloudacquired bytheFAROinstrumentisconsideredthegroundtruth,sinceitsprecisioniswithinoneortwo millimeters.WefitasurfacepatchtosmallgroupsofpointsacquiredbytheFAROandthencompute distancestothesepatchesfromthepointcloud.Ideally,theerrorhistogramshouldbesymmetricand theaveragedistanceclosetozero.Thewidthofthehistogram,asmeasuredbyitsfullwidthathalf maximum,isameasureoftherandomnoiseofthepointcloud.Anybreakinthesymmetryofthe histogramindicatesaproblemmergingofthepointcloudsthatareusedincreatingtheglobalpoint cloud. Intheexample,pointcloudswereacquiredinanaturalenvironmentwithgrass,treesandaflagmoving inthewind,inadditiontoman-madeobjects.Theresultsshowthatamobilemappingsolutioncan provideaccurateinformationforarchitecturalandengineeringuse.Insummary: • • • • Wecollectedpointclouddataofanaturalenvironmentthatincludedman-madestructures.We usedaFAROFocus3DX330fromfixedlocationsandRealEarth’sStencilsystem,whichacquires pointclouddatawhilemoving. Pointcloudsacquiredbythetwosystemswerealignedusingtheopensourceprogram CloudCompareforcomparison.Man-madeobjectsincludinglampglobes,acabin,andavehicle wereextractedfromthescenetocomparetheirpointclouds. Weprocessedthepointcloudstocomparedistancesbetweenthespheresandthedimensions ofthecabin. Analysesofpointclouderrorsindicateasmallamountofdistortionintheglobalpointcloud. Afterlocalre-alignmenterrorhistogramsareuni-modalandsymmetricaroundzeroerror. Differenceswerewithin0.2-1.4cmincomparisonwithastationarylidarforderivedpointsat spherecenters.Differenceswerewithin1-4.5cmfordistancesbetweenderivedpointsusing wallintersections. CopyrightKaarta2016 14 KAARTA Theoverallaccuracyoftheresultantdatawasexcellentandthemobilesystemcanprovideexcellent informationthatcanbeusedtomodelstructuresinCADandprovidedirectmeasurementsofbuildings andstructures. References [1]VelodyneVL0-16ScanningLaserRangeFinderSpecification.www.velodynelidar.com. [2]“Single-planeversusThree-planeMethodsforRelativeRangeErrorEvaluationofMedium-range3D ImagingSystems”,DavidK.MacKinnon,NationalResearchCouncilCanada,SPIEProceedings,June2015. [3]“TowardstheDevelopmentofaDocumentaryStandardforDerived-PointtoDerived-PointDistance PerformanceEvaluationofSphericalCoordinate3DImagingSystems”,BalaMuralikrishnanetal.,NIST. ProcediaManufacturing,V.XXX,2015,pp1-12. [4]“SLAMforDummies”,SørenRiisgaardandMortenRufusBlas.MITocw.mit.edu16-412online course. [5]“LOAM:LidarOdometryandMappinginReal-time”,ZhangandSingh.Robotics:ScienceandSystems Conference,July2014. [6]CloudComparev2.6.1-Usermanual.http://www.danielgm.net/cc/documentation.html [7]“Accurate3Dcomparisonofcomplextopographywithterrestriallaserscanner:applicationtothe Rangitikeicanyon(N-Z)”,DimitriLague,NicolasBrodu,JérômeLeroux [8]FAROTechnologyWhitePaper,Large-Volume3DLaserScanningTechnology [9]“AccuracyassessmentoftheFAROFocus3DandLeicaHDS6100panoramictypeterrestriallaser scannersthroughpoint-basedandplane-baseduserself-calibration”,Chow,Lichti,&Teskey,Canada. LaserScannersII,5865.RomeMay6-10,2012. [10]“Assessingtheperformanceof3D-imagingsystems”,DavidK.MacKinnon.SPIENewsroom.DOI: 10.1117/2.1201102.003393 CopyrightKaarta2016 15