(Abstract—Thispaperpresentsareal-timecolourimageclassificationalgorithmformobilerobotnavigationbasedontheadvancedtechnologyofFieldProgrammableGateArrays(FPGA).Inordertocalibratetheobjectcolourindifferentlightingconditions,akindofstatisticellipsoidalmodelisadopted.Webuilda3-DColourLook-upTable(CLUT)inwhichonly18bitsareusedtorepresentakindofcolourinsteadofconventional24bits.Thismethodresolvestheproblemofobjectoverlappinginacolourspace,andhasthemeritsofhigherclassificationaccuracyandlowermemorycostthantraditionalones.ItisimplementedonFPGA,whichcanhighlyreducetheCPUcomputationburdenandremarkablyimprovetheperformanceofrobotvisionsystems.Thismethodisvalidatedbytheapplicationsonamobilerobotintheresearch.I.INTRODUCTIONEAL-timeimageprocessingisoneofkeychallengesinmobilerobotnavigation.Althoughparallelprocessingcomputersoffersareasonablesolution,theyareexpensiveandmaynoteasytoembedonarobot.ProgrammablehardwareintheformofFPGAhasbeenproposedasaconventionalplatformtoexploitimageprocessinginordertoachievehighperformance[1][2][3].BasedonFPGA,wedesignanintelligentframegrabber(IFG),whichcannotonlygrabbutalsoprocessimagesinrealtime.RInthevisionsystemofintelligentmobilerobots,classifyingthepixelsinanimageintodifferentcolourclassesnotonlyprovidessolidfoundationfornavigation,decision-makingandself-localizationofmobilerobots,butalsoimprovestheperformanceofavisionsystemgreatly.Theapproachtoaccomplishingthistaskincludeslinearcolourthresholding,nearestneighbourclassification,colourspacethresholdingandprobabilisticmethods.Colourspacethresholdinghasreal-timeperformance,andcannotadaptthevarietyoflightingconditions,whichrestrictsitsuseinmanyreal-worldapplications.Therefore,manyresearchersdevotedtoimprovetheapproachandmadesomeprogresses[5][6].Butthemethodsdevelopedsofardidnotchangetheoverlapphenomenonoftargetcolourclassessincetheydefinedthecolourclassesin1Dor2Dcolourspace.A3DColourLook-UpTable(CLUT)isseldomusedinavisionsystembecauseofhighmemorycost.Inthispaper,a3DCLUTisbuiltonlyusing18-bitintegers,whichnotonly(QingruiZhou,KuiYuan,andWeiZouarewiththeInstituteofAutomation,CAS,Beijing,China.E-mail:kui.yuan@mail.ia.ac.cnH.HuiswiththeDepartmentComputerScience,UniversityofEssex,ColchesterCO43SQ,UnitedKingdom.E-mail:hhu@essex.ac.uk.keepsthegoodclassificationresultbutreducesthememorycostgreatly.Toensuretheadaptationoftheclassificationalgorithmtolightingconditions,targetcoloursarecalibratedbasedonamodelofstatisticellipsoid.Then,weimplementthealgorithmusingFPGAonanIFG,whichimprovestheperformanceofthevisionsystemandhasbeenusedinRoboCupmiddlesizerobots.Therestofthispaperisorganizedasfollows.SectionIIdescribesthealgorithmofimagetargetclassificationbycolour.SectionIIIintroducestheintelligentframegrabberandthealgorithmimplementationonFPGA.InsectionIV,experimentalresultsarepresentedtoshowtheperformanceoftheproposedsystem.Finally,abriefconclusionandfutureworkaresummarizedinSectionV.II.ALGORITHMFORCLASSIFICATIONA.ColourSpaceUptonow,severalcolourspaceshavebeenwidelyusedinmobilerobotnavigation,includingHSV,YUVandRGB,eachofwhichhaspeculiarcharacteristicsandutilities.Thechoiceofcolourspaceforclassificationdependsonseveralfactorssuchasthetypeofthedigitizinghardwareandtheutilityforaparticularapplication.RGBisthemostcommonlyusedmodelforaTVsystemandpicturesacquiredbydigitalcameras.Videomonitorsdisplaycolourimagesbymodulatin...