41
論文OImginalPapeNP
HandSignalRecognitionAimingatRobotControl
YukihiroMIYosHI
Abstrnct:Amethodtorecognizehandshape,i、e、,numbersoffingersstuckout;usingacolorimageispresent-
ed・Ameananddispersionofhuevaluesandameanofsaturationvalues,applyingtheHSLcolormodel,ina smalIsquareareaofanimagearerevealedtobeusefulfeaturestodistinguishskin-areaandcontourofthehand,
ThecontourofthehandistracedbytheborderfollowingmethodusingthesmaI1squarearea・Changesoftan-
gentialdirectionsofthecontourofthehandareusedtoidentifythenumberoffingersstuckout・Thewavelet shrinkageisusedunclosethefeaturepattemsofthetangentialdirectionsandmakeanautomatichand-recogni‐
tionpossible・Althoughonlyafewimagesevaluate。,numbersofthelingerswererecognizedautomatically.
squarearea・Cha gthesmaI1s
ngersstuckout.
entifythe
Keywords:colorimagerecognition,meanofhue,dispersionofhue,skin-area,handrecognition,wavelet shrinkage
惣「-1行
酢
1.IIltroductionn
1jpE
Recognitionofhandsignalsusingdigitalimageprocess- ingisexpectedasamethodforhuman-computer interface1)andamethodtoinstructarobotonawork2〕.’、
BBBmL6~ロロ■
ロロ
■■
thispaperamethodtorecognizehandposture,especiaUy anumberoflingersstuckout,isdevelopedColorinfOr‐
mation,e、9.,RGBvaluesandHSLvaluesofpixeIsinclude vari2tionandthedirectuseofthosevaluestendstoresult
paperamethodtoreco
mberoffingersstuck A こ■
一一
’inanunstablerecognition・Ameanandsomekindofdis‐
ゴ
persionareadequateasthecuetodetecthandareainan image・Inordertorealizeanaturalsignaling,neitherusage ofthegIovenorsimplecolorbackgroundsaresupposed forthemethodinthisDaDer・AndaiminRatthehuman‐paper・And imingatthe Tobotcommunication,time-consumingandsophisticated
caIcuIationlikeanartificialneuralnetworkisnotsup- Fig.1lmageshowingahandandabackgroun。
posed.
2.Skilnareacognitioningmimage 250 skin
Fig.1isa640x48024-bitRGBcolorimageshowinga handandabackground・Thisimageiscapturedusinga colorvideocameraunderHuorescentceilinglightsina roomFig・ZshowsR,0,Bvaluesofpixelsintherectan‐
2
1 DE
1
gli kin
gularareaA,i、e,theskinarea,andintherectangular areaB,i・a,thenon-skinarea,oftheimageshowninFig.
1.IntegernumbersbetweenOand255representsred,
greenandblueintensitiesofthepixels・Thepixelsinthe skinareaAandthoseinthenon-skinareaBoccupydiiTer- entRGBspace,asfarasfortheimageinFigl,soclassifi‐
cationintotwoclasses,i,e、,askinareaandanon-skin xels・The
area,ispossibleusingsomesortofclassifier,e,9.,aneural network3).Itisdesirable,however,thatfeaturestorecog‐
nizetheskinareaareintuitivelyunderstandable. HueisFig,ZRGBintensityofpixelsintheskinareaAandthenon-
skinareaB oneofthosefeatures.
bemappedtoh
R,0,Bvaluesofpixelscan ue,satura-
tionandIightnesscomponentsreferringtothedoublehex- coneHSLcolormodeI4).Hue,saturationandlightness
TransactionsoftheKokushikanUniv・FacultyofEngineering・No.40(2007)
42
作|蝋琴二
$|蝉二二二Ⅱ
small;(3)saturationvaluesforaskinareaarenotso smalLHuealone,however,isinsufTicienttorecognizethe skinbecauseoftheoverlapintherangebetweenl25and
l55・
Threefeatures(ameanofhuevalues,Hh,,dispersionof huevalues,Ha,andameanofsaturationvalues,Sm,ina squarearea)canbeusedtorealizeabove-mentioned strategy・TheHh1andtheSmareestimatedas:
Hh,=LZH『j,(4)
〃x,ys,"=-LzsxI,(5)
〃x,yDispersionofhuevaluesiscalculatedapplyingtheMan- hattandistanceasfollows:
Ha=上z'H)"-Hkj,’(6)
〃x,yHXyandSⅢ。,arethehuevalueandthesaturationvalueof
140 +
oSkinarea
+Non-skinarea
+
加川0000
118642雫への①。|⑩ンのゴー」。□○両」のQの一□
+++++++
+++++
+tFiニト+サキ+
++++++
++鳶+ 什土++ 農
讃+ 革鍼
++
0
2060100140180 Meansofhuevaluesノウルフ
Fig.4Meansanddispersionofthehuevaluesintheskinarea andthenon-skinarea
500 500
400 400
の
E300o j凸四
ji2OO
コの E300o
温巴
ji2OO
。100 100
0 1002003000100200
HueHHueH
(a)Skinarea (b)Non-skinarea
Fig、3HueandsaturationvaluesfortheskinareaAandthenon-skinareaB 300
HandSignalRecognitionAimingatRobotControl 43
thepixelofthecoordmatex,Jan。〃isthenumberofpix- elsintherelevantsquarearea・Fig.4showsthemean,Hh,,
S,,,.Iftheareaatthecenterisaskinarea,then,a20x20 squareareaneighboringtoanon-skinbackgroundissear- chedwhilethesquareareaismovedfromthecentertoleft direction・Ifnot,arasterscanoperationisusedtoIindthe anddispersion,Lb,ofthehuevaluescalculatedusingpix‐
eIsinlO×lOsquareareaswhicharedefinedbysectioning
cIosetheskinareaAandthenon-skinareaBFiE.4rev宮aAandthenon-skinareaB、Fig.4re boundarysquareskinarea・Fig.5illustratesthefirstcase:
acemerareaAisjudgedasaskinareaandaboundary squareareaBisfoundwhiletheareaismovedIeftfromA,
Acontourofthehandistracedbyusingsquareareasof lOx10pixelsstartingatthesquareareaBofFi9.5.Eight eaIsthataskinareacanbedistinguishedfromanon-skin
areabyusingbothHh,andHlノ:therelevantareaofaimage canberegardcdasaskinareawhenHmisinanarrow rangeand秘isverysmalLInwhatfolIows,anareahav-
ingaH,Ivaluebetweenl20andl60,aHhvaluelessthan lOandaS,,JvaIuegreaterthan50istakenasaskinarea,
ie.,anareaonahand.
lOxlOneighborsquareareas(horizontaLverticalanddi‐
agonaldirectneighborareasofthecurrentlOx10boun- darysquarearea)areinvestigatedbasedontheHh,,Hband S,,ltofindnextboundarysquarearea,Thenthecurrent boundaryareaisupdatcd、Theseoperationsareiterated
naIdirectn ,value
3.IIzmdcontourreCognitim
Ahandcontourisrecognizedfromacapturedimagein twosteps:(1)findasquareareaonahand;(2)tracethe contourmovingsquareskinareasthatareboundedby non-skinbackgroundarea・
Atthefirst,asquareareaof20x20pixelsatthecenter
untilthecurrentboundaryareacomesbacktothesquare areaBofFi9.5.Fig.6illustratestheresultantboundary lOx10squareareasforthesameimageasFi9.5.Almost thecorrectboundaryisdetected.
4.HzmdsignrecognitioM oftheimageisinspectedbyusingtheaforementioned
criteriabasedonthemeanofhuevaIues,H>,1,thedisper- sionofhuevalues,HJ,andthemeanofsaturationvalues,
lnordertoidemifyshapeofthehand,especialIyanum- beroffingersstuckout,twoparametersareinvestigated・
Oneisatangentialdirection,T,ofthecontourofthe
[ 一認
.■印を…bL
.厩1
,6-可]■
’
- -迅少squarearea magecenter
■
F山19
_=-=毎
 ̄ ̄
、〈
background
lOx10squareareas fOrmingacontourofthehand
Fig.5 SquaTeareaatcenterrecognized areaneighboringabackground
asskinandasquaTc
Fig.6BoundarylOx10squareareasofthehand
(b)
Fig.7Examplesofhandimages (a)
頁
…聖 :=二三鑿 蕊 showinE
ontour
篝
 ̄?
二
』~黒
白リヒ型富
山--、ご-印 解司一鱒
■■TransactionsoftheKokushikanUniv、FacultyofEngineering・No.40(2007)
44
300 300
0 200 ト
100
。』 200
100
2040 60 Numbersofboundarysquares
(a)TvaIuesofFig6
2040 0
Numbersofboundarysquares
(a)TvaluesofFig6
60
300 200 200 100
◎
ト100
O ト
0 20406080
Numbersofboundarysquares
(b)TvaluesofFig7(a)
100120 20406080100
Numbersofboundarysquares
(b)TvaluesofFig7(a)
300 300
200
○
ト 100
200
、
ト100
0 4080120160
Numbersofboundamsquares
(c)TvaluesofFig7(b)
Fig.9Tangentialdirection7softhehandcontoursafterthe waveletshrinkage
04080120160 Numbersofboundarysquares
(c)TvaIuesofFig7(b)
Fig.8Tangentialdirectionsofthehandcontours
TablelResultsoftheautomaticpatternrecognitionoftheimages NumbersoftheconvexesM
inthetangentialdirections Ratior:thecontouragainst theboundingbox
O78
Recognizednumberofthefingers Images
Fig.6 Fig.7(a)
Fig.7(b)
1 Oorl
2 5
2 1.03
5 1.33
handThetangentialdirection,nisoneoftheeight(two horizontal,twoverticalandfourdiagonaDdirectionsof thecurrentboundarysquare:T=0.,90.,180.and270o meanx,-y,-xandydirectionofthecoordinateshown inFi9.6,respectively・Tisdecidedbythepositionalrela- tionamongtheformerboundarysquare,nextboundary squareandthenon-skinsquareintheeightneighbor squaresaroundthecurrentboundarysquare.Anotherisa ratiordehnedasfollows:
〃=-〃c (7)
〃B
where〃cand〃BarenumbersofthelO×10squaresonthe contourofthehandandontheboundingboxrespectively・
Theratiorislessthanlwhenthecontourisanellipse・
Fig.7showstwoexamplesofhandimages・Thecon- toursshown,onlyforillustrations,inFig.7bydotsare determinedbasedonthepositionalrelationsbetweenthe boundarysquareareasandthenon-skinbackground
areas・
Fig.8showsthetangentialdirections,7s,ofthecon‐
toursofthesehandsinFig6,Fig.7(a)andFig.7(b).A changeoftangentialdirection,Tbover90゜meansaconvex oraconcave・Thechangesofthetangentialdirectionsvary bytheoutwardformofthehands,sousingthetangential directionsofthecontourscanrecognizetheshapesofthe hands:thenumberofthefingersstuckout,Thechanges showninFig8,however,exhibitvariationstoughto recognizeautomatically,Thewaveletshrinkageusingthe Haarwaveletisusedtounclosethefeaturepatterns・Fig9 showsthetangentialdirectionsafterthewaveletshrinkage byusingthescalingsequenceoflevel3・Tablelshowsthe resultsoftheautomaticpatternrecognition;thatis,the numberoftheconvexes,M,ofFig9andtheratio,r・One convexinFig9meanstwochangesofthetangentialdirec- tions:therecognizedhandisregardedashavingashape likeellipseincludingashapewithonefingerstuckout.
HandSignalRecognitionAimingatRobotControl 45
TwoconvexesinFig、9meansashapewithtwofingers、
5.Conclusion
ically.
References Amethodtorecognizehandshape,i、e、,numbersoffin-
gers,usingacolorimageispresentedinthispaper・Skin areasinanimagearedistinguishedfromanon-skinarea byvaluatingthreefeatures:ameananddispersionofhue valuesandameanofsaturationvalues,applyingtheHSL colormodel,inasmallsquareareaofanimagearecalcu- latedforskin-recognition、Movingthesmallsquarearea onacapturedimageandvaluatingthethreefeaturescoor- dinatesandtangentialdirectionsofacontourofahand areestimated、Changesofthetangentialdirectionsofa contourreHectashapeofthehand,thatis,anumberof fingersstuckout・Althoughonlyafewimagesevaluate。,
numbersofthefingersstuckoutwererecognizedautomat‐
1)J・IsaacsandSFoo,OptimizedWaveletHandPoseEstima‐
tionforAmericanSignLanguageReCognition,Proceedings ofthe2004CongressonEvolutionaryComputationVol、1,
797-802.
2)N・KawarazakietaL,CooperativeWorkSystembetweenHu‐
manandManipulatorusingHandGestureInstructions(in Japanese),JournaloftheRoboticsSocietyofJapan,VOL23, No.6(2005),121-126.
3)M、AAkbariandM、Nakajima,ANewSkinDetectorfor DifTerentColorSpaces,TechnicalreportoflEICE,IE2004-
143,111-115.
4)M・TAKAGLHSHIMODA(Eds.),Handbookoflmage Analysis(inJapanese),UniversityofTokyoPress,(2004),
1191-1193.