




版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
GamesPlane:AugmentedRealityGamesman
MillerHollinger
ElectricalEngineeringandComputerSciencesUniversityofCalifornia,Berkeley
TechnicalReportNo.UCB/EECS-2025-136
/Pubs/TechRpts/2025/EECS-2025-136.html
June4,2025
Copyright©2025,bytheauthor(s).
Allrightsreserved.
Permissiontomakedigitalorhardcopiesofallorpartofthisworkfor
personalorclassroomuseisgrantedwithoutfeeprovidedthatcopiesare
notmadeordistributedforprofitorcommercialadvantageandthatcopiesbearthisnoticeandthefullcitationonthefirstpage.Tocopyotherwise,torepublish,topostonserversortoredistributetolists,requirespriorspecificpermission.
Acknowledgement
ToDanGarcia,forbeinganexcellentresearchadvisor.
ToEricPaulosforkindlyreviewingthispaper.
ToJoshZhangforhisassistanceincameracalibrationwhichgreatlyimprovedthesystem’s
accuracy.
ToAlvaroEstrellaforextendingtheGamesmanUniAPI.
ToSriyaKantipudiforaddingnewgamestoGamesPlane.
ToAbrahamHsuforconsultingwithmeaboutvectormath.
ToGamesCraftersasawhole.
ToBarakStout,my9th-gradecomputerscienceteacherwhosepatienceandknowledgeledmeintothewonderfulworldofcomputation.
Aboveall,tomyparents,JohnandKarenHollinger,forsupportingme
throughoutmyacademiccareerandbeinganendlesssourceofkindnessandadvice.
1
GamesPlane:AugmentedRealityGamesman
byMillerHollinger
ResearchProject
SubmittedtotheDepartmentofElectricalEngineeringandComputerSciences,UniversityofCalifornia,Berkeley,inpartialsatisfactionoftherequirementsforthedegreeofMasterofScience,PlanII.
ApprovalfortheReportandComprehensiveExamination:
Committee:
TeachingProfessorDanGarciaResearchAdvisor
Date
*******
ProfessorEricPaulosSecondReader
Date
2
Abstract
Inthisreport,IintroduceGamesPlane,asystemthatusesaugmentedrealitytooverlaythepre-calculatedcoloredvalueofmovesforaplayerplayingaphysicaltwo-person,
complete-informationboardgame.Thevaluescomefromapplicationprogramminginterface(API)callstoourGamesmansystemthathasstronglysolvedthegame.Itprovides20framespersecondupdatesforboardsofallshapesandsizeswithnear-100%accuracy.Finally,itishighlyconfigurable,withtoolsanddocumentationprovidedforeasyextensibility.
3
Acknowledgements
ToDanGarcia,forbeinganexcellentresearchadvisor.Hisguidance,encouragement,and
positivitymadeGamesPlanepossible,andmyloveforcomputerscienceresearchisthankstohim.
ToEricPaulosforkindlyreviewingthispaperandhisfeedbackontheproject’sdevelopment.
ToJoshZhangforworkingalongsidemetodevelopandtesttheGamesPlaneproject,andespeciallyforhisassistanceincameracalibrationwhichgreatlyimprovedthesystem’s
accuracy.
ToAlvaroEstrellaforextendingtheGamesmanUniAPItointegratewithGamesPlane.
ToSriyaKantipudiforaddingnewgamestoGamesPlaneandworkingwithmetoimproveitsaccuracy.
ToAbrahamHsuforconsultingwithmeaboutvectormath.
ToGamesCraftersasawhole;fortheMonday-Wednesday-Fridaylunchmeetingsandthelate-nighthangouts.Mycollegeexperiencewouldnothavebeenthesamewithoutthem.
ToBarakStout,my9th-gradecomputerscienceteacherwhosepatienceandknowledgeledmeintothewonderfulworldofcomputation.
Aboveall,tomyparents,JohnandKarenHollinger,forsupportingmethroughoutmyacademic
careerandbeinganendlesssourceofkindnessandadvice.
4
Contents
Abstract 2
Acknowledgements 3
Contents 4
Chapter1:Introduction 6
Chapter2:Background 8
2.1:GamesCrafers,Gamesman,GamesmanUni,anGamesPlane td8
2.2:BringingGamesmanoPhysicalBoarGames td8
Chapter3:RelatedWork 10
3.1:Makingrealgamesvirual:Trackingoargamepieces1 1tbd[]0
3.2:AugmeneRealiyChessAnalyzerARChessAnalyzer2 1tdt()[]0
3.:AnInelligenChessPieceDeecionTool 13tttt[3]0
3.4:Chessoaranchesspiecerecogniionwihhesupporofneuralneworks4 11bddttttt[]
3.:ChessPieceDeecion 115tt[5]
3.:ComparisonoGamesPlane 116t
3.7:SysemComparisonChar 1tt3
Chapter4:DevelopmentandChallenges 14
4.1:ConcepualizaionofGamesPlane 14tt
4.2:FirsIeraion 14ttt
4.:TransiionoArUco-OnlyApproach 13tt5
4.4:GamesPlaneArchiecure 1tt6
4.:PrinaleGamesPlanes 175tb
4.:SofwareDevelopmen 16tt8
4.7:HomePage 20
4.:Crafsman 218t
4.:Documenaion 29tt3
4.1:LaunchGame 240
4.11:SarerGuie 2ttd6
Chapter5:Results 27
5.1:BoarSaeDeecion 27dtttt
5.2:OpimalAccuracy 27t
5.:HeighTolerance 23t8
5.4:PieceRoaionTolerance 2tt8
5.:PoorlyAlignePiece 25d9
5.:CameraRoaions 26tt9
5.7:OherShorcomings 1tt3
5.:Spee 18d3
Chapter6:FurtherWork 32
Chapter7:Conclusion 33
5
Bibliography 34
Appendix:User’sGuide 35
Appenix1:SaringGamesPlaneLocally dtt35
Appenix2:PlayingaGamewihGamesPlane dt35
Appenix:AingaNewGame d3dd35
Sep1:CreaeaPhysicalBoaranPieces ttdd36
Sep2:Creaehe.sonFileinCrafsman tttjt36
Sep:WrieaUWAPIConverer 7t3tt3
Sep4:RegiserhegameinAppLaunchGame.py ttt/38
Sep:PlayYourNewGame! t538
Appenix4:Trouleshooing dbt38
6
Chapter1:Introduction
GamesCraftersisaresearchgroupdedicatedtosolving,analyzing,andplayingtwo-playerturn-basedcomplete-informationgamessuchasTic-Tac-Toe,NineMen’sMorris,orOthello.GamesCrafters’systemGamesman[8]accomplishesthisgoal,usingacodedescriptionofagametocreateasolution.Overtime,itsscopeexpandedanditaddedpuzzlesandmore
complexgames.Apublicwebsite,GamesmanUni[9],wascreatedtotransitionGamesmanawayfromadownload-and-recompileX11-based-GUImodelandtowardsthemodernageofwebapps.Figure1.1showsaviewofthesite.
Figure1.1.GamesmanUnisportsnearly100playablegamesandpuzzles,allstronglysolved.
Despitethisgrowth,Gamesmanhadyettocrossacriticalboundary…intothereal
world.Althoughmostofthegamesitsolvedwerenearlyalwaysplayedface-to-face,findingthevaluesofmovesthroughGamesmanrequiredopeningawebsite(orrunningalocalscript)toinputineachmoveoneatatime,thencross-referencetheresultswiththereal-worldboard.
WithGamesPlane,weentertherealworldusingArUcotracking-poweredboardsandpieces.GamesPlaneisasystemforcreating,readingpiecedatafrom,anddisplaying
informationontospeciallydesignedgameboards.UsingGamesPlane,youcandefinethe
informationforagameboard,useacameratolocatewherethepiecesare,andthenseethebestmovesinaugmentedreality.Indoingso,GamesPlaneservesasareal-lifeinterfacetoGamesman,bringingtheperfectplayhintsoncerelegatedtocomputerscreenstophysicalgameboards.
7
Figure1.2.TheoverlaydisplayedbyGamesPlaneonBlack’sturn.Yellowmovesare
“draw”moves,andredmovesare“losing”moves.Greenmovesare“winning”moves,thoughno
suchmovesexistinthisparticularposition.
AuserofGamesPlanecanprintoutoneofmanygameboardandpiecesets,thenrunGamesPlaneontheirdevice.GamesPlanefindsthepiecesandsendsthatinformationofftoGamesmanUni,whichreturnsanimageshowingvalidmovesintheposition,withgreen
“winning”,yellow“tie”and“draw”moves,andred“losing”moves.Thisimageisdisplayedin
augmentedreality(i.e.,projectedoverthecamerapicture)usinginformationabouttheboard,creatingthefinalresult:areal-lifeimageoverlaidwithanARgraphicshowingmovesbasedonthepieces’positions(Figure1.2).
GamesPlaneworksusingArUcomarkers,whicharesquareblack-and-whitetags
featuringspecificpatternseasilyrecognizablebycameras.DetectionofArUcomarkersis
implementedbyOpenCV,makingthemaneasytoolforcomputervisionprojectssuchas
GamesPlane.Auserplacesanchormarkersontoagameboardandattachesthemtopieces.Thesemarkers’relativepositionsinanimageareusedtodeterminewherethepiecesareinspace,andthenonthegameboard.
GamesPlanehastheuniquequalityofbeingextensiblebydesign:withitsintegrated
documentation,starterguide,andCraftsmantool,futureuserswillhaveaneasytimelearningtousethesystemandsetupboards.
Thesystemrunsatahighframerate—atleast20framespersecond,evenonlow-endsystems.Therearecurrently5gamesthatfunctiononGamesPlane.Templatesareprovidedtoprint8.5x11”boardsforavarietyofgames,buthypotheticallyanysizeorshapecanfunctionaslongascameraqualityishighenough.ItcanfailwhenArUcomarkersareobscuredorthe
cameraisattoolowofanangle.butitisalsofairlytolerant:itfunctionsevenwhenpiecesare
imperfectlyaligned,unevenlyrotated,orofdifferentsizes.
8
Chapter2:Background
2.1:GamesCrafters,Gamesman,GamesmanUni,andGamesPlane
GamesCrafters’centralgoalhasalwaysbeentostronglysolvedeterministictwo-player
turn-basedgameswithcompleteinformationlikeTic-Tac-Toe,orConnectFour.Weusea
systemcalledGamesman[8]tosolvethesegames;itexhaustivelysearchesallpossibleboardstatesforagame,eventuallycreatingadatabasestoringeverypositionandifitisawin,lose,tie,ordrawinperfectplay(andhowmanymovesitisfromtheendofthegame).UsersmostoftenaccessthesedatabasesusingGamesmanUni[9],apubliclyavailablewebsitethatlets
usersplayboardgamesonlinethroughgraphicaluserinterfaces(GUIs).Duringtheirgames,playersareshownvaluemoves:themovestheycanmakecoloredred,yellow,orgreenforiftheyarelosing,drawing/tieing,orwinningrespectively.ThisflowofinformationisshowninFigure2.1.
Figure2.1.InformationaboutthegameisgeneratedbyGamesmanandthenservedbyUWAPIto
GamesmanUni.TheuserinterfacecreatedbyGamesmanUniissenttoGamesPlane.
2.2:BringingGamesmantoPhysicalBoardGames
Foralongtimenow,ithasbeenagoalofthelabtobringthepowerofGamesmantothephysicaldomain.Thereasonforthisisfairlystraightforward:boardgamesaregenerallyplayedface-to-face,notonline.BeforeGamesPlane,touseGamesmanwithareal-lifegamewould
requiretheusertoenterintoGamesmanUnieverymovetheymake,whichisacumbersomeapproachthatslowsdownthepaceofgameplay.
Attemptsatcrossingthebarrierbetweendigitalandphysicalhavebeenmadeinafewinstances.Forexample,inFall2024,aprojecttodisplayvaluemovesforConnectFourwasattemptedusingcolordetection.Althoughthismethodwassuccessful,itwasnotgeneralizabletoothergames,asitworkedbydetectingtheaveragecolor(redoryellow)ineachConnect
Fourboardspace.Additionally,thissystemonlyworkedonvideorecordings,notonlivevideo.
9
Theobjectiveofcreatingasystemthatwouldallowmanygamestobeplayedinrealliferemained.
Asidefromthegoalofcrossingintothephysicaldomainandgeneralitytomanygames,athirdgoalwasaccessibility.AnotherkeyideaofGamesCraftersisthatgamesshouldbe
enjoyedbyeveryone.GamesPlane,then,aimstobeaseasytouseaspossiblefornewusers.Thissuggeststheinclusionofone-clickinterfaces,easysetup,andexhaustivedocumentation.Together,usabilityfeaturessuchasthesecontributetotheongoinguseanddevelopmentofthe
software,bothbycasualandtechnicalusers.
10
Chapter3:RelatedWork
Theconceptofasystemthatconvertsvideoofboardgamesintodigitalboardstatesisinitselfnotnovel.Multiplepapershavebeenauthoredonthetopic,generallyfocusingon
implementationsforpopularboardgameslikeChessorGo.GamesPlane’simplementation
provesitselfuniquethroughitsabilitytobegeneralizedtomanyboardgames,rapidresponsetime,andvaluemovedisplays.
3.1:Makingrealgamesvirtual:Trackingboardgamepieces[1]
ThisstudentprojectfromUSCSmakesuseofRANSACandHiddenMarkovModelstofindthegridofaGoboard,andthendeterminethecolorsofpiecesonthejunctionsbetween
spaces.ThesystemisfunctionalonstandardunmarkedGostones,evenatoff-angles,andis
abletodetecttheentire19x19Goboard.Ithasgooddetectionaccuracyatabout91%.Its
accuracyisboostedusinganA*algorithmthatusespreviouslyknownboardstatestopredict
thelikelihoodofdetectedboardstates:ifastateisdetectedbutismanymovesawayfromtheinitialstate,itisdeemedlesslikelyandasimplerexplanationisused.Themaindrawbackof
thissystemisitstimetooperate,taking40secondstorun20iterationsofanA*algorithmperpicture.Italsodoesnothaveanyreal-timeboardoverlay,insteadonlyfocusingonrecordingthegameovertime.
3.2:AugmentedRealityChessAnalyzer(ARChessAnalyzer)[2]
ARChessAnalyzermakesuseofaConvolutionalNeuralNetwork(CNN)approach
alongsideanARoverlaytoshowrecommendedmovesoverlaidonimagesofchessboards.
TheirCNNapproachenablesthemtouseexisting,unmodifiedchessboardsandpieces.Theytouta93.45%accuracyinstaterecognition,whichisexcellentconsideringthatupto32piecesmustberecognizedandpositionedcorrectlyforaboardstatetobecorrect.Whencomparedwithsomeoftheotherrelatedworks,thisaccuracyisespeciallyimpressive.However,the
complexcalculationsinvolvedinrunningaCNNresultsina3-4.5secondwaitingperiod
betweentakingapictureofthegameandseeinganARoverlay.Consideringthataugmentedrealityreliesonrealisticallysuperimposingthedigitalworldovertherealworld,thisdelay
becomesanotabledrawbackasitbreakstheappearanceofthedigitalobjectsappearing“inreallife.”
3.3:AnIntelligentChessPieceDetectionTool[3]
ThispaperusesCNNstolocateandcategorizechesspieces,withthegoalofcreatingaboardstate.Thepaperfocusesprimarilyonthemethoditusestocategorizepieces:aYOLO
objectdetectionalgorithm.YOLOmeans“YouOnlyLookOnce,”andreferstoanalgorithmthatrunstheimagethroughitsnetworkasingletime.BeforeYOLO,approachessuchasR-CNN
wereusedwhichwouldoftenneedtopropagateasingleimagethroughanetworkthousandsoftimes.WithYOLO,fasterorreal-timeobjectdetectionbecomespossible.Foundchesspiecesaredisplayedonaseparatechessboardtothesideoftheimageofthegameboard.Ithas
acceptableaccuracyper-pieceat84.29percentcorrectinthebestcase,however,fullboardstatedetectioncanbeerroneousconsideringupto32piecescanbeonachessboardatonce.ItrequirestheuseofacustomCNN,whichtakesbetween2and12hourstotrainand3to21secondstorun.Consideringthatitalsorequiresamassivedataset(about140,000picturesinthecaseofchess)totrain,itbecomesveryinefficienttoconvertthisCNN-basedapproachtoothergames.
11
3.4:Chessboardandchesspiecerecognitionwiththesupportofneuralnetworks[4]
ThispaperfromtheInstituteofComputingScienceatPoznanUniversityofTechnologyusesanovellatticedetectortofindachessboardinanimage,andthenasupportvector
machineandaconvolutionalneuralnetworktolocatethepieces.Theymakeuseofachess
enginetodeterminewhatboardpositionsaremostlikelytoincreaseaccuracy(e.g.havingthreewhitebishopsisveryuncommon).Withthisapproach,theyachieveanastounding99.57%
accuracyinchessboarddetectionand95%accuracyinpiecedetection.Drawbacksoftheirapproachincludethefactthatitdoesnottransferwelltoothergames,requirealattice-shapedboard,anditsexecutiontime,whichsometimesreaches4.5secondsforasingleframe.
3.5:ChessPieceDetection[5]
ThisapproachusesaYOLOCNNtodetectthespacesofachessboardaswellasthepiecesonit.Generally,itsapproachissplitintothreesteps:a“boarddetection”stepwheretheedgesoftheboardarelocated,a“griddetection”stepwherethespacesoftheboardare
delineated,anda“chesspiecedetection”stepwhereindividualpiecesineachspaceare
recognized.Itleveragesthefactthatchessboardshavealternatinglightanddarkcolorsto
detectavarietyofboardsusingOpenCV.AswiththeotherCNNapproaches,themajor
drawbackinthiscaseisthatitreliesonaspecificallytrainedCNNdesignedforchesspieces,
andsoconversiontoothergamesisacostlyaffair.Additionally,itsuseofcontourtracingmeanstheapproachcanonlypracticallyfunctiononchessboardsorothersquareboards/
3.6:ComparisontoGamesPlane
Consideringpreviousworkinthefieldofgameboardrecognitionandmovedisplay,GamesPlanedifferentiatesitselfinafewcriticalways.
Generality:GamesPlanecanfunctionwithpracticallyanygame,notjustChessand/orGo.EvengamesthatarenotcompatiblewithGamesmancanstillhavetheirboardstates
extracted.Critically,thisappliesevenforgameswithirregularboards:anyboardshape(e.g.hexagonal,triangular,orcompletelyirregular)canfunctionwithGamesPlane.Pieceswithflattops(likecheckers,disks,ortiles)workbestwithGamesPlane,butothershapescanworkaswellaslongasArUcosareflatwhenattached.
Accuracy:ThankstoitsArUco-basedapproach,GamesPlanehasexcellentaccuracyatavarietyofangles,whileothersystemstendtofocusonfunctioningatonespecificangle.Withthecamerapositionedfromthetopdown,itsrecognitionaccuracyisessentiallyperfect.
Accuracyisextremelyimportantinthecaseofstrategygames,asevenasinglemisplacedpiececancompletelyalterwhatthebestmoveis.
Speed:GamesPlanerunsatahighframerate,around25framespersecond.
Additionally,thereisnovisibledelaybetweenreallifeandthevideodisplay,evenonthe
low-endlaptopusedfortestingtheproject.Ahigher-endcomputercouldhypotheticallyachieveevenhigherframespersecond.ItonlyhastoqueryGamesmanUnioncetodisplayARvaluemoves,however,sothisisstillgenerallyfasterthantheCNN-basedapproaches.Additionally,GamesPlanecancacheitsoverlays,leadingtoinstantresponsetimes.
AROverlay:ByusingGamesmanUni’soverlays,GamesPlanegainsaccesstothe
interfaceofeverygameimplementedinGamesmanUniwithoutneedingtowriteadditional
interfacecode.Thatis,wedon’thavetore-drawtheinterface’sarrowsforslidingmovesanddotsforplacementmoves.Weinsteadusethegraphicdrawnbytheexistingsystemand
overlaythat,leveragingabstractionanda“don’trepeatyourself”softwarephilosophy.
Additionally,becauseGamesmansolvesgamesfully,ouroverlayshowsperfectvaluemoves,notalgorithmicorAIsuggestions.
12
AdownsideofoursystemisthattheusermustprepareaGamesPlaneboardwithArUcomarkers.However,withapre-madePDFreadytoprint,eventhisprocesstakesnomorethanafewminutespergame.ExistinggameboardscanalsobeconvertedtoGamesPlane
compatibilitybyaddingArUcoanchormarkers.GamesPlanethereforemakesaworthwhiletradeoffthatresultsinavarietyofbenefitsvaluabletoboardgameplayers.
13
3.7:SystemComparisonChart
System
ApplicableGames
State
RecognitionAccuracy
OperationSpeed
ValueMovesDisplay
Underlying
Technology
Makingreal
gamesvirtual:
Trackingboardgamepieces
Go
90.57%
40seconds
perframe
None
RANSAC,MarkovModels,A*
Augmented
RealityChess
Analyzer
(ARChessAnalyzer)
Chess
93.45%
4.5secondsperframe
Displayschessenginemoves
ConvolutionalNeuralNetwork
AnIntelligent
ChessPiece
DetectionTool
Chess
84.29%perpiece
3-4secondsperframe
Separateboarddisplay
ConvolutionalNeuralNetwork
Chessboardandchesspiece
recognitionwiththesupportofneuralnetworks
Chess
95%
4.57secondsperframe
Nodisplay;onlylocatespieces.
SupportVector
Machine,
ConvolutionalNeuralNetwork
Chesspiecedetection
Chess
81%
Real-Time
Nodisplay;onlylocatespieces.
ConvolutionalNeuralNetwork
GamesPlane
Any
Gamesman
Game
97-100%
Real-Time
Displaysvaluemovesin
augmentedreality
ArUcowithOpenCV,Gamesman
14
Chapter4:DevelopmentandChallenges
4.1:ConceptualizationofGamesPlane
Whencreatingtheinitialconceptforthisproject,IknewIwouldneedaboardthatwouldprovidesomekindofvisualanchorthatacameracoulddetect.I’dseenAprilTags[11]fromrobotics
applicationsbefore,andsodecidedtouseArUcotagsasthey’reintegratedintoOpenCV[12].PythonisalanguagealreadywidelyusedforvariousapplicationsinGamesCrafters,sousingPythonwithOpenCVwasanaturalapproachasitwouldensurefutureGamesCrafters
memberswouldbeabletoreadthecode.
4.2:FirstIteration
GamesPlaneitselfbeganasasingleobject,called“TheGamesPlane.”Thiswasawooden
board,aboutafoottoaside,featuringa5-by-5gridwith2-inchwidesquares(seeFigure4.1).ItiscompatiblewithgamessuchasTic-Tac-Toe,4x4Othello,Chung-Toi,orabout10others
supportedbyGamesman.Theboardhadspecialslotsinwhich5ArUcomarkerscouldbeplaced.IassumedthattheArUcotrackingwouldbeveryerroneous,andthathavingmoremarkerswouldessentiallyallowformultiplesamples.
Figure4.1.TheoriginalGamesPlane:Awoodenboard,intendedtobetheonlyboard
compatiblewiththesystem.
Forthepieces,I3D-printedcustompieces.AsseeninFigure4.2,thesewerestyledintheshapeofarook,butshorter.MythoughtprocesswasthathavingtallpieceswouldobscuretheArUcomarkerspastedontotheboardwhenviewedfromanangle,soshorterpieceswerepreferable.Atthetime,IwasunsureifIwouldbeplacingArUcomarkersontothepiecesorifIwouldbeusingsomekindofimagedetection,andsoIkeptthetopsflattoallowspaceforanArUcomarkertobeattached.Iadditionallyaddedridgesontotheedgeinthehopethatitwouldmakethepiece’sshapemoredefined,makingiteasiertodetectwithanyimagedetection
methodImightuse.
15
Figure4.2TheoriginalGamesPlanepiece,meantforusewithHaarCascades[6]asexplained
below.Itwasmeanttoevoketheshapeofachessrook.
ThefirstpipelineIenvisionedfortheGamesPlanewasasfollows:theArUcomarkers
wouldhavehardcodedreal-lifepositions.Iwouldthenuseanobjectdetectionmethodcalled
HaarCascades[6]tolocatethepiecesintheframe,andusetheirpositionsintheframe
comparedtothatoftheArUcostoestimatetheirreal-worldposition.ThereasonforthisdecisionwasthatIwantedthegamepiecestolooklikenormalgamepieces—Ithoughtthatadding
ArUcomarkersontopofthemwouldmakethemlooktoodifferentfromtraditionalboardgamepieces.
4.3:TransitiontoArUco-OnlyApproach
Unfortunately,thisinitialapproachworkedverypoorly.ThemainissuewaswiththeHaar
Cascades.Inordertodetectanewobject,it’snecessarytotrainanewcascadeusingimagescontainingtheobject(positiveexamples)andimagesnotcontainingtheobject(negative
examples).Ithereforetookabout250picturesofthepiecearoundmyapartment,andused
imagesfromtheinternetasnegativeexamples.Thisfailedtoadequatelytrainthecascade:itendedupwithafalse-positiverateofabout70%onimagesnotcontainingthepiece.TheissuewasdowntothevarietyofimagesIfedin.Bytakingmanysimilarpictures,Iinadvertently
trainedthecascadetodetectafewmuchsimplerfeatures:specifically,ashadowthatappearedontheedgeofmytableandanothershadowthatappearedalongsidethewall(seeFigure4.3).Withhowmanysituationscreateshadowssimilartothese,itwouldregularlyfindthepiecein
locationsitwasnot.Atthispoint,thetwowaysforwardwereeithertotakemuchmorevariedpicturesofthepieceinmanycontextsortosimplychangetoanall-ArUcoapproach.
Consideringthescopeoftheproject,itseemedmoresensibletoswitchtorelyingonArUcosentirely.
16
Figure4.3.TheHaar-basedapproachhadverylowaccuracy,andwouldusuallydetecta
randomcrackinthetable.ThegreensquareiswheretheHaarclassifierthinksthepieceis—
it’sactuallyinmyhand.
IbeganbyattachingArUcomarkerstothetopofthe3Dprintedpieces.However,therewasanissuehereaswell—ArUcomarkersneedwhitespacearoundtheiredges.When
detectingamarker,contrastbetweenwhiteandblackisused,andsoifthemarkerdoesnot
haveawhiteborderitfailstodetectit.Thetopsofthepiecesweretoosmall,soplacinga
markerwithaborderontopofthemmadethemarkersinvisibletothecamerafromevenashortdistance.Torectifythis,IstoppedusingthepiecesandstartedusingtheArUcomarkersontheirownwithnopiecesupportingthem—thisway,atleastwhiletesting,theywouldbelargerandmorevisible.
Atthispoint,withasolelyArUco-basedapproach,IwentaboutimplementingOpenCV’sArUcodetection.Thechallengeitpresentedwasoneofcoordinatespaces.MyfinalgoalwastoextractwhatIcall“boardcoordinates”:whereagamepieceisonaboard(e.g.f4inchess).Togetthesecoordinates,I’dneedtofirstobtainboard-centeredworldcoordinates,whichare(x,y,z)coordinatesforwhereagamepieceisinspacealignedtothegameboard’sframeof
reference.Thesewouldthencomefromcamera-centeredworldcoordinates,whichcanbe
obtainedbyestimatingtherelativepositionandposeofanArUcomarkertoacamerabyusinganimagethecamerahastaken.Convertingbetweenthesefourcoordinatespaces,and
displayinginformationfromeachinasensibleway,madeupmuchoftheworkoftheproject.Smallmathematicalerrorsweredifficulttodetectbuthadamassiveeffectontheaccuracyofpiecelocation.
Afterresolvingtheseconversions,Ifinallywasabletolocateapiece,butinaccurately.
Thedetectionwouldregularlyplaceapieceseveralspacesawayfromitstruelocation.In
strategygameswhereasinglespacecanmakeamassivedifferenceinthebestmove,thiswasexpectedlyunacceptable.Thebreakthroughinraisingaccuracycamewhendiscussingthe
camera’scalibrationfile.Acalibrationfilecontainstheinformationaboutacamera’sintrinsic
properties,specificallyitsoutputimagesizeandfocallength.Analyzingthefilerevealedthatthecamerahadbeencalibratedverypoorly,possiblyduetoanissueintheGitHubrepositoryIusedto
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 2025企业签订货物买卖委托合同书
- 2025二手汽车买卖合同
- 2025工业污染控制合同
- 2025企业合同管理表格:试用与履行
- 甘肃省武威市2023−2024学年高一下册期末考试数学试卷附解析
- 2025届浙江省台州市黄岩区中考二模数学试卷
- 身份验证漏洞管理基础知识点归纳
- 社区社区发展规划管理基础知识点归纳
- 石大学前儿童保育学课外必读:我的第一章练习题幼儿生理的特点
- 2025年计算机网络技术试题
- 广东华泰纸业有限公司纱管原纸工艺设备技改项目环境影响报告书
- 红皮病护理的课件
- SH-T-3503-2017-附录A-交工技术文件通用表
- 完整CECAGC3-2010建设项目工程结算编审规程完整
- 招标文件技术规范书
- 初中音乐-第五单元环球之旅(二)欧洲与大洋洲教学设计学情分析教材分析课后反思
- 2023年四川省内江市中考数学试卷【含答案】
- 2023-2024学年四川省绵阳市小学语文六年级期末高分通关题附参考答案和详细解析
- 电机学知到章节答案智慧树2023年东北电力大学
- 气候变化科学概论试题及答案
- 湖南省郴州市2016年中考数学试卷(解析版)
评论
0/150
提交评论