混乱场景中的物体识别
2270IEEETRANSACTIONSONPATTERNANALYSISANDMACHINEINTELLIGENCE,VOL.36,NO.11,NOVEMBER2014
3DObjectRecognitioninClutteredSceneswithLocalSurfaceFeatures:ASurvey
YulanGuo,MohammedBennamoun,FerdousSohel,MinLu,andJianweiWan
Abstract—3Dobjectrecognitioninclutteredscenesisarapidlygrowingresearcharea.Basedontheusedtypesoffeatures,3Dobjectrecognitionmethodscanbroadlybedividedintotwocategories—globalorlocalfeaturebasedmethods.Intensiveresearchhasbeendoneonlocalsurfacefeaturebasedmethodsastheyaremorerobusttoocclusionandclutterwhicharefrequentlypresentina
real-worldscene.Thispaperpresentsacomprehensivesurveyofexistinglocalsurfacefeaturebased3Dobjectrecognitionmethods.Thesemethodsgenerallycomprisethreephases:3Dkeypointdetection,localsurfacefeaturedescription,andsurfacematching.Thispapercoversanextensiveliteraturesurveyofeachphaseoftheprocess.Italsoenlistsanumberofpopularandcontemporarydatabasestogetherwiththeirrelevantattributes.
IndexTerms—3Dobjectrecognition,keypointdetection,featuredescription,rangeimage,localfeature
Ç
1
INTRODUCTION
BJECT
recognitioninclutteredscenesisafundamentalresearchareaincomputervision.Ithasnumerousapplications,suchasintelligentsurveillance,automaticassembly,remotesensing,mobilemanipulation,robotics,biometricanalysisandmedicaltreatment[1],[2],[3],[4],[5].Inthelastfewdecades,2Dobjectrecognitionhasbeenextensivelyinvestigatedandiscurrentlyarelativelymatureresearcharea[6].Comparedto2Dimages,rangeimageshaveshowntoexhibitseveraladvantagesforobjectrecogni-tion.Forinstance,(i)rangeimagesprovidemoregeometri-cal(depth)informationcomparedto2Dimages.Rangeimagesalsoencodesurfacemetricdimensionsunambigu-ously.(ii)Featuresextractedfromrangeimagesarecom-monlynotaffectedbyscale,rotationandillumination[7].(iii)Anestimated3Dposeofanobjectfromrangeimagesismoreaccuratecomparedtoanestimatedposefrom2Dimages.Accordingly,rangeimageshavethepotentialtoovercomemanyofthedif cultiesfacedby2Dimagesinthecontextofobjectrecognition[8].Theseadvantagesmake3Dobjectrecognitionanactiveresearchtopic[9].Moreover,therapidtechnologicaldevelopmentoflowcost3Dacquisitionsystems(e.g.,MicrosoftKinect)makerangeimagesmoreaccessible[10],[11],[12].Furthermore,advancesincomput-ingdevicesenabletheprocessingofanycomputationally
Y.GuoiswiththeCollegeofElectronicScienceandEngineering,NationalUniversityofDefenseTechnology,Changsha,Hunan410073,China,andwiththeSchoolofComputerScienceandSoftwareEngineering,TheUni-versityofWesternAustralia,Perth,WA6009,Australia.E-mail:yulan.guo@nudt.edu.cn.
M.BennamounandF.SohelarewiththeSchoolofComputerScienceandSoftwareEngineering,TheUniversityofWesternAustralia,Perth,WA6009,Australia.
M.LuandJ.WanarewiththeCollegeofElectronicScienceandEngineer-ing,NationalUniversityofDefenseTechnology,Changsha,Hunan410073,China.
O
Manuscriptreceived6May2013;revised24Nov.2013;accepted3Apr.2014.Dateofpublication10Apr.2014;dateofcurrentversion9Oct.2014.RecommendedforacceptancebyA.Fitzgibbon.
Forinformationonobtainingreprintsofthisarticle,pleasesende-mailto:reprints@ieee.org,andreferencetheDigitalObjectIdenti erbelow.DigitalObjectIdenti erno.10.1109/TPAMI.2014.2316828
intensive3Dobjectrecognitionalgorithmtoruninafairlyacceptablemanner.Allthesecombinedfactorshavecontrib-utedtothe ourishmentofresearchtowardsthedevelop-mentof3Dobjectrecognitionsystems.
Existing3Dobjectrecognitionmethodscanbedividedintotwobroadcategories:globalfeaturebasedmethodsandlocalfeaturebasedmethods[13],[14].Theglobalfea-turebasedmethodsprocesstheobjectasawholeforrecog-nition.Theyde neasetofglobalfeatureswhicheffectivelyandconciselydescribetheentire3Dobject(ormodel)[15].Thesemethodshavebeenwidelyusedinthecontextof3Dshaperetrievalandclassi cation[16],[17].Examplesinthiscategoryincludegeometric3Dmoments[18],shapedistri-bution[19],viewpointfeaturehistogram[20],andpotentialwellspaceembedding[21].They,however,ignoretheshapedetailsandrequireapriorisegmentationoftheobjectfromthescene[11].Theyarethereforenotsuitablefortherecognitionofapartiallyvisibleobjectfromclutteredscenes[14].Ontheotherhand,thelocalfeaturebasedmethodsextractonlylocalsurfacesaroundspeci ckeypoints.Theygenerallyhandleocclusionandclutterbettercomparedtotheglobalfeaturebasedmethods[14].Thistypehasalsoproventoperformnotablybetterintheareaof2Dobjectrecognition[22].Thisconclusionhasalsobeenextendedtotheareaof3Dobjectrecognition[23].Onthatbasis,thefocusofthispaperison3Dobjectrecognitioninclutteredsceneswithlocalsurfacefeatures.
Severalsurveypaperswerepublishedon3Dobjectrecog-nitionanditsrelated elds,suchas3Dshapematchingandmodeling.Amongthese,severalsurveypaperson3Dobjectrecognitionarealsoavailable.Forinstance,thearticlesin[24],[25],[26],[27],[28]and[29].Tworeviewson3Dmodel-ingandrangeimageregistrationby[30]and[31]arealsoworthmentioning.However,noneofthesepapersspeci -callyfocusesonlocalfeaturebased3Dobjectrecognition.Bronsteinetal.[32]presentedareviewonkeypointdetectionandlocalsurfacefeaturedescriptionmethods.Theyhowever,coveredonlyfourkeypointdetectorsandsevenfeaturedescriptors.Anarticleontheevaluationof
0162-8828ß2014IEEE.Personaluseispermitted,butrepublication/redistributionrequiresIEEEpermission.
Seedoc.guandang.net.
3D pose in single images of cluttered scenes. ... object recognition has always and localization; ...of objects with sharp surface intersections [7]....
object recognition in cluttered natural scenes. ... a specific configuration of local features define... a few objects and their surface characteristics ...
in cluttered video data with no prior learning. ...local internal “correlation surface” (Fig. 2.a...Puzicha. Shape matching and object recognition ...
Visual Event Recognition in Videos by Learning ...features may fall on cluttered backgrounds because ...complex scenes with spatial and temporal ambiguities...
enable automated learning and recognition of objects in cluttered scenes [21]...But they do not consider the local features of the background, that are ...
cluttered backgrounds, occlusions, as well as ...extracted volumetric features from regions with ...cation and general object recognition [23]. In ...
cluttered backgrounds, occlusions, as well as ...extracted volumetric features from regions with ...cation and general object recognition [23]. In ...
热门文档