Three-Dimensional-Object-Recognition-and-6-DoF-Pos

PointCloudLibraryThree-DimensionalObjectRecognitionand6DoFPoseEstimationByAitorAldoma,Zoltan-CsabaMarton,FedericoTombari,WalterWohlkinger,ChristianPotthast,BernhardZeisl,RaduBogdanRusu,SuatGedikli,andMarkusVinczeWiththeadventofnew-generationdepthsen-sors,theuseofthree-dimensional(3-D)dataisbecomingincreasinglypopular.Asthesesensorsarecommodityhardwareandsoldatlowcost,arapidlygrowinggroupofpeoplecanacquire3-Ddatacheaplyandinrealtime.Withthemassivelyincreasedusageof3-Ddataforper-ceptiontasks,itisdesirablethatpowerfulprocessingtoolsandalgorithmsaswellasstandardsareavailableforthegrowingcommunity.ThePointCloudLibrary(PCL)[1]aimsatprovidingexactlythese.Itisacollectionofstate-of-the-artalgorithmsandtoolstoprocess3-Ddata.ThelibraryisopensourceandlicensedunderBerkeleySoft-wareDistribution(BSD)termsand,therefore,freetouseforeveryone.ThePCLprojectbringstogetherresearchers,universities,companies,andindividualsfromallaroundtheworld,anditisrapidlybecomingareferenceforany-oneinterestedin3-Dprocessing,computervision,androboticperception.ThePCLcoreisstructuredinsmallerlibrariesofferingalgorithmsandtoolsforspecificareasof3-Dprocessing,whichcanbecombinedtoefficientlysolvecommonproblemssuchasobjectrecognition,registrationofpointclouds,segmentation,andsurfacereconstruction,withouttheneedofreimplementingallpartsofasystemneededtosolvethesesubtasks.Inotherwords,thetoolsandalgorithmsprovidedbyPCLallowresearchersandcompaniestobetterconcentrateontheirspecificareasofexpertise.Inthisarticle,wefocuson3-Dobjectrecognitionandposeestimation.Specifically,ourgoalistorecognizerigidobjectsfromasingleviewpointandestimatetheirpositionandorientationintherealworld.Theobjectsusedtotrainthesystemarerepresentedas3-Dmeshes,andtherealobjectsaresensedusingadepthsensorsuchastheKinect.Inparticular,wereviewseveralstate-of-the-art3-Dshapedescriptorsaimedatobjectrecognition,whichareincludedinPCLandarepubliclyavailable.Asagooddescriptorintheendisasmallpartofanobjectrecognitionsystem,wepresenttwoentirepipelines80�IEEEROBOTICS&AUTOMATIONMAGAZINE�SEPTEMBER20121070-9932/12/$31.00ª2012IEEEDigitalObjectIdentifier10.1109/MRA.2012.2206675Dateofpublication:10September2012©DIGITALVISIONforourrecognitionsystembuiltusingbitsandpiecesavail-ableinPCL:onereliesonglobaldescriptorsthatrequirethenotionofobjectsandhencedeployaspecificprepro-cessingstepbasedonsegmentationwhiletheotheroneuseslocaldescriptorsthatarecomputedlocallyaroundkeypointsand,thus,donotrequireapresegmentationstep.Throughoutthearticle,theadvantagesanddisadvantagesofthedifferentmethodsusedarepresentedtoprovidethereaderwiththenecessaryinsightsfordesigningthefeaturesofarecognitionsystemtobeusedinaspecificfront-endapplication.ThetutorialalsoprovidesspecificguidelinesonhowtousePCLprimitivesforthevariousmethodspresented.Finally,anexperimentalevaluationisperformedtodemonstratetheapplicabilityandtheeffectivenessofthepresentedmethodsforthetaskathand.Moreover,bothtrainingandgroundtruthtestdatasetsarepubliclyavail-abletogetherwithasetoftoolstoperformevaluationsonthesedatasets.LocalDescriptorsGenerally,3-Dlocaldescriptorsaredevelopedforspecificapplicationssuchasregistration,objectrecognition,andlocalsurfacecategorization.Forsuchapplications,eachpointisassociatedwithadescriptordescribingthelocalgeometryofapoint.SignatureofHistogramsofOrientationsThesignatureofhistogramsoforientation(SHOT)descriptor[2]encodesasignatureofhistogramsrepre-sentingtopologicaltraits,makingitinvarianttorotationandtranslationandrobusttonoiseandclutter.Thedescriptorforagivenkeypointisformedbycomputinglocalhistogramsincorporatinggeometricinformationofpointlocationswithinasphericalsupportstructure.Foreachsphericalgridsector,aone-dimensionalhistogramisconstructedbyaccumulatingpointcountsoftheanglebetweenthenormalofthekeypointandthenormalofeachpointbelongingtothesphericalsupportstruc-ture.Thefinaldescriptorisformedbyorderlyjuxta-posingallhistogramstogetheraccordingtothelocalreferenceframe.Discretequantizationofthesphereintroducesaboundaryaffectwhenusedincombinationwithhisto-grams,resultinginabruptchangesfromonehistogrambintoanother.Therefore,quadrilinearinterpolationisappliedtoeachaccumulatedelement,resultinginanevenlydistributionintoadjacenthistogrambins.Finally,forbetterrobustnesstowardpoint-densityvar-iations,thedescriptorisL1-normalized.Thedimen-sionalityoftheusedsignatureis352.Algorithm1givessomeguidelinesonhowtocomputeSHOTdescriptorsusingPCL.Inthisalgorithm,first,anobjectoftheclasspcl::SHOTEstimationiscreatedusingthetemplateparametersofourinputdata(cloudandnormals).Akd-treeisprovidedtotheestimatortoperformNNsearches.IndicesrepresentthekeypointsfromtheinputcloudwhereSHOTdescriptorsshouldbecomputed.Theradiussearchrepresentsthesizearoundeachkeypointthatwillbedescribed.Finally,callingthecomputemethodreturnsapointcloudwithasmanydescriptorsaskeypoints.FastPointFeatureHistogramDescriptorssuchaspointfeaturehistogram(PFH)[3],fastPFH(FPFH)[4],andviewpointfeaturehistogram(VFH)[5]canbecategorizedasgeometry-baseddescriptors.Thesedescriptorsrepresenttherelati