ObjectRecognitionbyaMobileRobotusing
Omni-directionalVision
HenrikAndreasson,TomDuckett
AASS,Dept.ofTechnology,currency1OrebroUniversity,SE-70182currency1Orebro
Abstract.Thispaperproposesanewmethodforrecognizingtypicalobjectsfoundin
indoorofceenvironments(tables,chairs,etc.,)byamobilerobotequippedwithan
omni-directionalvisionsensor,withoutrequiringanypre-installedgeometricmodels
ofobjects.Theapproachutilizesthemotionoftherobottoacquireaninternalrepre-
sentationofagivenobjectusingstructurefrommotionoropticow.First,asetof
low-levelpointfeaturesareselectedfromthesegmentedareaoftheimagecontaining
theobject.Thelow-levelfeaturesaretrackedbyasetofindependentKalmanltersas
therobotmovesthroughtheenvironment,inordertoextractthe3Dpositionsofthese
points.Asetofhigh-levelfeaturesisthenextractedforinputtoapatternrecognition
system,basedonthespatialdistributionofthelow-levelpointfeatures.Thesamefea-
tureextractionmethodisthenappliedforrecognitionofthelearnedobjects.Results
arepresentedforsomerstexperimentsonarealrobotinalaboratoryenvironment.
1Introduction
Theproblemofobjectrecognitionisacentraltopicofinterestforresearchersincomputer
vision,articialintelligence,thecognitivesciencesandrobotics.Withoutthisability,thepos-
sibilitiesforrobotstocarryoutusefultasksremainlimited.However,afteroverftyyearsof
research,therestillexistsnogeneralpurposealgorithmforobjectrecognitionbyautonomous
robots.Instead,theproblemisusuallysolvedonasystem-by-systembasis,usingrecognition
techniquesthatarehand-craftedforasmallclassofobjectsinoneparticularapplication.Any
improvementinobjectrecognitiontechnologywouldbeuseful,andasignicantadvance
couldrevolutionizethestudyofembodiedintelligentsystemssuchasrobots.
Thereareanumberofusefulpropertiesthatanyobjectrecognitionmethodforanintel-
ligentrobotshouldhaveinordertobeusefulforapplicationsinreal-worldenvironments.It
shouldbeabletotakecareofobjectswithallkindsofshapes,andtosenseobjectsinaclut-
teredandoccludedworld.Itshouldbeinvarianttovariationsinscale,translation,lighting
conditions,etc.,androbusttotheextranoisecausedbythemotionoftherobot.Furthermore,
thereshouldbenoneedforpre-installedobjectmodels,andthealgorithmshouldbeableto
runinreal-time.Anidealapproachwouldbetolettherobotdrivearoundintheenvironment
andlearnitsowninternalmodelsofdetectedobjectsfromitsownsensoryperceptions,and
thentobeabletorecognizethesametypesofobjectsusingthesameperceptualapparatus.
Thispaperpresentsarsteffortatbuildingacompleteobjectrecognitionsystemfora
mobilerobot,usingomni-directionalvisionasthemainsensoryinput.Ourapproachexploits
themobilityoftherobot,usingthisinformationtoextractstructurefrommotionfroma
sequenceofimages.Inthispaper,weconsidertherecognitionoftypicalobjectsfoundinin-
doorofceenvironments(wherestate-of-the-artmobilerobotsarecurrentlyabletonavigate),
PATTERNRECOGNITION
objectclassification
features/histogram
EXTRACTION
HIGHLEVELFEATURE
(x,y,z)
pointsvelocites
setofpointsand
LOWLEVELFEATURE
EXTRACTION
birdviewimage
IMAGETRANSFORM
segmentedomniimage
SEGMENTATION
omnicamimagex,y,th
robotsodometry
TRACKING
Figure1:Left:overviewoftherecognitionalgorithm.Theboxesindicatethemainsteps,withintermediatedata
structures.Dashedboxesindicatestepsthatareperformedmanuallyinthecurrentimplementation.Right:robot
platform(ActivmediaPeoplebot).Theomni-directionalcameraismountedontopoftherobot.
includingtables,chairsandtrash-cans.Itisassumedthattheseobjectsareorientedconsis-
tentlyintheverticalaxis,i.e.,chairsandtablesremainuprightanddonotfallover,butthe
recognitionalgorithmshouldbeinvarianttorotationsaroundtheverticalaxis.Inthecurrent
implementation,theareaoftheimagecontainingtheobjectisrstsegmentedbyhand:future
workwillinvestigateafullyautomaticsystem.Anoverviewoftherecognitionalgorithm,to-
getherwithadetaileddescriptionofitscomponentfunctions,canbefoundinSection2(see
alsoFig.1).Thisisfollowedbyexperimentalresults(Section3),togetherwithconclusions
andsuggestionsforfuturework(Section4).
1.1RelatedWork
Themostcommonapproachin3Dobjectrecognitionistocollectasetof2Drepresentations
oftheobjectfromdifferentviews,withoutrequiringadeepunderstandingoftheunderly-
ing3Dstructureoftheobject.Theviewscanberepresentedbyanaspectgraph[2],where
Figure2:Left:Originalimagefromtheomni-cam.Right:transformedbirdviewimage,size:400x400pixels,
resolution:40pixels/meter.
eachviewisconnectedtoitsclosestneighbours.Thefeaturesusedforrecognitioncanbe
globalshapemodels[6],HOTcurves[13],ormorelocalrepresentationssuchasedges[14].
Otherrecognitionmethodsthatdonotrequireexplicitgeometricinformationoftenusecolour
andluminanceinformation,e.g.,withcolourcooccurencehistograms[5],phase-basedlocal
features[4],orprincipalcomponentsanalysis[7].
Ifa3Dsurfacemodeloftheobjectisavailable,thenrepresentationbyspinimagescan
beused[12].Spinimagesarerepresentationsofsurfacesthatareconstructedfromadense
collectionofpointsandaresuitableforregistrationormatchingofsurfaces.Thistechnique
hasbeenshowntoworkwellinclutteredenvironments.Otherkindsofsensors,suchasrange-
ndersensors,canalsobeusedtoextract3Dsurfacemodelsforobjectrecognition[15].
Theuseofmotiontodetect3DstructureisoftencalledStructurefromMotion(SfM).For
aintroductiontothistopicsee[11].Thefocusisusuallyonhowtodetectthemotionofthe
cameraortheobjectwithoutanypriorinformationaboutthecorrespondenceinformation.
MuchoftheresearchinSfMconcernsndingouttheCorrespondedStructurefromMotion,
whichassumesthatthecameraparametersandthe3Dmotionbetweenthecameraandobject
isunknown,whichisnotthecaseinourmethod.
2Method
Ourapproachistouseamobilerobotwithasinglecamera,withoutrequiringanypre-
installedmodelsoftheobjects.Thecameraisomni-directional,i.e.,theviewingangleis
(almost)360degrees.Itisplacedontopoftherobot,lookingdownwardsabovetheoor.
Thecameraisxedanditisonlytherobotthatmoves.Inthecurrentexperiments,therobot
onlytravelsforwardswithoutrotation.Thesamesensors(omni-camplusodometryforesti-
matingself-motion)areusedforbothtrainingoftheclassiersandrecognitionoftheobjects.
Bymovingtherobotaroundinaknownmannerandmeasuringthepixeldisplacementina
sequenceofimages,the3Dstructureoftheobjectisestimated.
AnoverviewoftherecognitionalgorithmisgiveninFig.1.Theareaoftheimagecon-
tainingtheobjectisrstsegmentedmanually.Aftertransformationoftheimagetoabird’s
eyevieworbirdview(Fig.2),asetoflow-levelpointfeaturesareextracted(Fig.3).The
Figure3:Fromlefttoright:segmentedomni-image,birdview,low-levelpointfeaturesthataretracked,and
high-levelfeatureextractionforpatternrecognition.
pointfeaturesaretrackedbyasetofindependentKalmanltersinordertoestimatetheir
3Dcoordinates,byreferencetothegroundvelocityoftherobot.Thetrackedpointsarethen
groupedbyahistogramaccordingtotheirrelativeheightintheworld(Fig.3,right)inorder
toobtainasetofhigh-levelfeaturesforinputtoapatternrecognitionsystem.Theindividual
stepsofthealgorithmaredescribedindetailasfollows.
2.1Segmentation
Intheexperimentspresentedinthispaper,objectsweremanuallysegmentedintheorigi-
nalomni-camimages.Imageswerecollectedwiththerobotdrivingpastastationaryobject
standinginfrontofawhitebackgroundintheroboticslaboratoryatourinstitute.Thebor-
deroftheobjectwasmanuallyselected(usingtheGNUimagemanipulationprogram‘The
GIMP’)andtherestoftheimagewaslledwithwhite.
2.2ImageTransformation
Duethecurvatureofthemirrorintheomni-cam,itisdifculttoextractgeometricalfeatures
directlyfromtherawimages(e.g.,horizontalsurfacesappeartwisted).Instead,atransforma-
tiontoa‘birdview’isused,whichcanbedenedasanimagetakenfromaviewlocatedhigh
abovethesurface.Thebirdviewtransformstheimageinordertokeepthephysicalshape
intactinthegroundplane.Forexample,achessboardlyinghorizontalatanyheightwillgive
anon-twistedchessboardinthebirdviewtransformation.
Thesizeofanobjectinthebirdviewwillincreasewithheight.Thismeansthathorizontal
areaswillnotchangeusingthebirdviewtransformation.Linesintherealworldwilltrans-
formintolinesinthebirdview,comparedtoarcsintheoriginalomni-camimage.Areasthat
arehigherwillseemtobebiggerandfurtherawayfromtheimagecenter.Theresolutionof
thebirdviewisgivenbypixels/meteratthegroundlevel,andthereforethepixelcoordinates
canbemappeddirectlytoaworldcoordinatesystem.Foramoredetaileddescriptionabout
transformationsonomni-camimagessee[10].
Totransformanimageintoabirdview,thetransformationfunctionthatconvertsfroma
a0a2a1a4a3a6a5a8a7coordinateintherealworldtoapixelintheomni-camimagehastobeknown.Since
meterR
rpixel
meterR
rpixel
a9a10a9
a9a10a9
a9a10a9
a9a10a9
a11a10a11
a11a10a11
a11a10a11
a11a10a11
a12a10a12a13
a14a10a14a15a10a15a16
a16
a16
a17
a17
a17
a18a19
Figure4:Calibrationoftheomni-camtondthebirdviewtransformationfunction.
thedistanceisinvarianttotheorientationofthecamera,thetransformationfunctioncanbe
writtenasa20a22a21a24a23a26a25a27a23a26a28a30a29a32a31
a0a34a33a6a35a37a36a39a38
a23a26a40
a7a41a3(1)
where
a20a22a21a42a23a34a25a27a23a26a28
isthedistancefromthecenterofthecameratothepointonthegroundlevel
intherealworld,anda33a37a35a37a36a39a38
a23a43a40
isthedistancecalculatedinpixelsintheomni-camimagefrom
thecentertothepixelcorrespondingtothatpointintherealworld.Thisfunctioncanbe
calculatedanalyticallyiftheparametersforthemirrorandthecameraareknown,whichis
rarelythecase.Thefunctionusedinourexperimentswasapolynomialofdegree3interpo-
latedwithastandardleastsquarettingalgorithm,byusingimageswherethedistanceand
thecorrespondingpixelswereknown.Tospeedupthetransformation,alook-uptablewith
memorypointerstothepixelsintheomni-camimagewascreated.
2.3Low-levelFeatureExtraction
Thepointstotrackareselectedintherstimageinthesequence.Pointsthatarelocated
oncornersofobjectsarethemosteasytotrack.Toselectpointswithstrongmatching
capabilities,aneighbourhooda44of3x3pixelsisselectedaroundeachpixelintheim-
age.Thederivativesa45a38anda45a47a46arecalculatedwithaSobeloperatorforallpixelsinthe
blocka44.Foreachpixe,ltheminimumeigenvaluea48iscalculatedformatrixa49wherea49
a29
a50a51
a45a53a52
a38a55a54a57a56a58
a51
a45
a38a55a54a57a56a58
a45a47a46
a54a27a56a58
a51
a45
a38a55a54a27a56a58
a45a47a46
a54a57a56a58a51
a45a53a52
a46
a54a27a56a58a59and
a51isperformedovertheneighbourhoodof
a44.Thepix-
elswiththehighestvaluesofa48arethenselectedbythresholding.Forfurtherdetailssee[16],
orthefunctioncvGoodFeaturesToTrackintheOpenCVlibrary[3].
2.4TrackingoftheLow-levelFeatures
Thenextstageistotrackthepointsastherobotdrivespasttheobjectataconstantspeed,
usingthesequenceofimages.Bytrackingthepointswiththebirdviewprojection,itis
possibletoestimatetheheight(a60-coordinate)ofapointdirectlyfromitsrelativevelocityin
theimagesequence,andthentousethisinformationtoestimatethehorizontalposition(a1-
anda5-coordinates).Inourapproach,thepointsaretrackedwiththeiterativeLucas-Kanade
P0(tk)
P0)(tk+1
Pbv
P)(tk+1
P(tk)h
h
BirdViewOrigin
Figure5:Estimationofthea61a27a62a64a63a26a65a67a66positionfromthetrackedpoint,giventhecorrespondinga68-coordinate.
methodusingpyramids[1].Theideaistondthedisplacementa69
a38and
a69a70a46thatminimizesa71
as
a71
a0
a69
a38a70a3
a69a72a46
a7
a29a73a75a74a77a76a79a78a67a74
a80
a38a41a81
a73a74a75a82a78a74
a73a41a83a84a76a79a78a67a83
a80
a46
a81
a73a83a55a82a78a83
a0a77a0a86a85a87a0a86a1a4a3a6a5a88a7a24a89a91a90a92a0a86a1a94a93
a69
a38a70a3a6a5a95a93
a69a72a46
a7a84a7a41a3(2)
wherea85istheimageattimea96,a90istheimageattimea96a93a69a72a96anda0a34a97a87a38a98a3a6a97a46a7isthepointto
track.a99a38anda99a100a46refertothesizeoftheareathatisminimized.Intheseexperiments,values
ofa99a38
a29a102a101
anda99a38
a29a103a101
wereused.
Tohandlelargepixeldisplacementswithoutrequiringtoomuchcomputation,theimages
a85anda90aredividedinto3-4moreimagesthataresub-sampledbyafactorof2.Therstmin-
imizationofa71andtherstestimationofa69a38anda69a72a46areperformedonthelowestsub-sampled
image.Thenminimizationisperformediterativelyonthenextlevelusingthepreviousesti-
matesofa69a38anda69a72a46,andsoon.Thismakesitpossibletotrackpointswithlargedisplacements
withhighprecision.Forafulldescriptionofthisalgorithmsee[1].Toremovenoiseandto
estimatethevelocityofthepoints,anindependentKalmanlterisappliedtoeachofthe
pointstracked[9].Inthisworkweassumethattherobottravelsforwardwithoutrotation,so
odometryisusedonlytoestimatetheheighta60ofthetrackedpoint.
2.5High-levelFeatureExtraction
Theheighta60ofapointin3Dspaceisafunctionofitsapparentvelocityintheimagese-
quence.Apointwithhighervelocityshouldbelocatedhigherthanapointwithlowerve-
locity,assumingthatthecorrespondingobjectisstationary.Sincepointsthatarehigherare
alsolocatedfurtherawayfromtheorigininthebirdview,theapparentvelocitycanalsobe
usedtoestimatethea1-anda5-coordinatesusingtheprojectionshowninFig.5.Theground
levelpixelvelocitya104a41a105a24a106a0a96a108a107a7a42a89a105a24a106a0a96a108a107
a76a87a109
a7
a104isrstestimatedusingtheodometryoftherobot.
Thepixeldisplacementa104a41a105a42a110a0a96a108a107a7a42a89a105a24a110a0a96a108a107
a76a87a109
a7
a104atheighta111andthedistanceofthetrackedpoint
fromthebirdviewcentera104a41a105a42a112a114a113a89a105a24a110a0a96a108a107
a76a87a109
a7
a104isthencalculated.Thedistancethatthepoint
shouldbemovedtowardstheoriginofthebirdviewinordertogivethecorrespondinga1-
anda5-positionatgroundlevelcanthenbecalculatedas
a104a41a105a24a110
a0
a96a108a107
a76a87a109
a7a42a89
a105a24a106
a0
a96a108a107
a76a87a109
a7
a104
a29
a104a41a105a24a112a114a113
a89
a105a100a110
a0
a96a108a107
a76a87a109
a7
a104a116a115a118a117
a89
a104a41a105a24a106
a0
a96a108a107
a7a42a89
a105a100a106
a0
a96a108a107
a76a87a109
a7
a104
a104a41a105a24a110
a0
a96a108a107
a7a42a89
a105a100a110
a0
a96a108a107
a76a87a109
a7
a104a120a119a122a121
(3)
Toobtainthehigh-levelfeaturevaluesrequiredforinputtothepatternrecognitionsys-
tem,ahistogramisconstructedbasedonthevelocitydistributionofthetrackedpoints(see
Fig.6).Intheseexperiments,histogramswereusedwithsevenbinscorrespondingtoseven
speedintervalsbetween3.0pixelsperframeto9.0pixelsperframe.Theapproachissimple
2
6
4
8
2
6
4
8
nrofpoints
93
pixelvelocity
(pixels/frame)
nrofpoints
93
pixelvelocity
(pixels/frame)
Figure6:Histogramshowingnumberofpointswithdifferentpixelvelocities,left-cone,right-chair3.
Figure7:Fromlefttoright:chair1,chair2,chair3,table1,table2,drawers,bottle,cone,trashcan.
buteffective,providedthatobjectsareorientedconsistenlyintheverticalaxis.Itshouldbe
invarianttorotationsaroundtheverticalaxis,thoughitwouldfailifanobjectisknocked
over,turnedupside-down,etc.Moresophisticatedmethodsforfeatureextractionwillbesub-
jecttofutureresearch:forexample,itshouldbepossibletorecoverinformationaboutthe
orientationoftheobjectsrecognisedifanappropriaterepresentationisused.
2.6PatternRecognition
Thepatternrecognitionmethodusedinthispaperisaverysimpleandintuitiveclassier
knownasaminimumdistanceclassier(mdc)[8].Inthismethod,meanvectorscalculated
fromthetrainingdataforeachclassareassumedtobeidealprototypesfortheobjects.To
classifyanewinputvector,theEuclideandistancetoeachoftheprototypesiscalculated,
andthevectorisassignedtotheclasswiththeshortestdistance.Equivalently,thedecision
functionforaminimumdistanceclassiercanbewrittenas
a69a67a123
a0a86a124a100a7
a29
a124a87a125a87a126
a123
a89a117
a127
a126a128a125
a123
a126
a123
a3(4)
wherea124isthepatternvectortobeclassied,anda126a123isthemeanvectorofeachclassa99a129a123.
Classicationofagivenobjectisthendeterminedbytheclassthatproducesthehighest
decisionvalue.
3ExperimentalResults
Themethodwastestedusingimagedatarecordedwiththemobilerobotusing9different
objects(seeFig7).Alsoshownarethenumberofviewsfromwhichdatawascollected(e.g.,
drivingtherobotpastthe‘North’,‘South’,‘East’or‘West’sideoftheobject),andthetotal
NameDescriptionNo.ofViewsNo.ofImagesCorrectclassication
chair1Ofcechair416088%
chair2Regularchair416063%
chair3Ofcechair4160100%
table1Squaretable28075%
table2Roundtable14088%
drawersChestofdrawers4160100%
bottle1.2mcylinder14085%
conePlasticcone140100%
trashcanGreentrashcan140100%
Table1:Objectsusedandclassicationresults.
numberofimagesrecorded.Thesequenceofimagesforeachobjectwereseparatedinto
20smallersequencescontaining10imageseach.Foreachsequencethepixelsspeedswere
estimatedandthehistogramcreated.Classicationwasrepeated100times,usingarandomly
selectedsetof70%ofthedatafortrainingandtheremainingdatafortesting.Theaverageof
theseresultsisgiveninTable1.Chair2hadthelowestrateofcorrectclassications,dueto
thefactthattheinitialdistributionoflow-levelfeaturesvariedalotbetweendifferentviews.
Furtherworkisneededtoaddressthisproblem.Thetotalcomputationtimeforoneiteration
ofthealgorithm(excludinghandsegmentation)wasa130a131a117a133a132a72a132msona2GHzPentium4,which
indicatesthatrecognitioninreal-timeispossible.
4ConclusionandFutureWork
Inthispaper,wehavepresentedarstattemptatrecognitionoftypicalobjectsfoundinin-
doorofceenvironmentsbyamobilerobot.Therobotlearnsitsowninternalrepresentation
ofagivenobjectfromitsownsensoryperceptionsasittravelspastthatobject,bycombining
exteroceptivesensoryinformationfromanomni-directionalcamerawithproprioceptivesen-
soryinformation(self-motion)fromodometry.Themethodconstrainstheobjectrecognition
problembyexploitingthephysicalpropertiesofthetherobotanditsinteractionwiththe
environment.Atpresent,wehaveonlyconsideredrecognitionofsomeselectedobjectsina
laboratoryenvironmentusinghandsegmentationoftheimages,buttheexperimentsdemon-
stratetheconceptofrecognisingobjectstructurefrommotioninanembodiedintelligent
system.Thepatternrecognitionsystemonlyusesinformationconcerningtheheighta60ofthe
trackedpointfeatures,andthepointdistributioninthehorizontalplanesisnotconsidered.
Futureworkwillincludeimprovementsatalllevelsoftherecognitionalgorithm,forexam-
ple,automaticsegmentationofobjectsbyclusteringoflow-levelpointfeatureswithsimilar
attributes;furtherexploitationoftheembodimentoftherobot,e.g.,byattemptingtopushob-
jects,learningaffordancesofobjects,etc.;betterhigh-levelfeaturestoallowagreaterlevel
ofdiscriminationbetweendifferentobjecttypes;moresophisticatedpatternrecognitiontech-
niques;integrationofdifferentsensormodalities,e.g.,fovealvision,thermalvision,laserand
ultrasonicrange-ndersensors,etc.;discriminationofmovingobjectssuchashumansfrom
non-movingobjects;andexperimentsinclutteredenvironments.
References
[1]Jean-YvesBouguet.PyramidalimplementationoftheLucasKanadefeaturetracker,descriptionofthe
algorithm.Technicalreport,IntelCorporation,MicroprocessorResearchLabs,1999.
[2]K.W.BowyerandC.R.Dyer.Aspectgraphs:anintroductionandsurveyofrecentresults.International
JournalofImageingSystemsandTechnology,2:315328,1990.
[3]GaryRBradski.OpenSourceComputerVisionLibrary.IntelCorporation,2001.
[4]GustavoCarneiroandAllanD.Jepson.Multi-scalephase-basedlocalfeatures.volume1,pages736743.
IEEEComputerVisionandPatternRecognition,2003.
[5]P.ChangandJ.Krumm.Objectrecognitionwihcolorcooccurrencehistograms.IEEEComputerVision
andPatternRecognition,1999.
[6]C.M.CynandB.B.Kimia.3Dobjectrecognitionusingshapesimilarity-basedaspectgraph.International
ConferenceonComputerVision,2001.
[7]R.Dillman,M.Ehrenmann,andM.Ambela.Acomparisonoffourfastvisionbasedobjectrecognition
methods.InProceedingsoftheInternationalConferenceonRoboticsandAutomation,pages18621867.
IEEE,2000.
[8]RichardO.Duda,PeterE.Hart,andDavidG.Stork.PatternClassication.Wiley-Interscience,2000.
[9]ArthurGelb,editor.AppliedOptimalEstimation.TheMITPress,1974.
[10]ChristopherGeyerandKostasDaniilidis.Paracatadioptriccameracalibration.IEEETransactionson
PatternAnalysisandMachineIntelligence,24(5):687695,2002.
[11]TonyJebara,AliAzarbayejani,andAlexPentland.3Dstructurefrom2Dmotion,1999.
[12]AndrewJohnsonandMartialHebert.Usingspinimagesforefcientobjectrecognitionincluttered3D
scenes.IEEETransactionsonPatternAnalysisandMachineIntelligence,21(5):433449,May1999.
[13]D.J.Kriegman,B.Vijayakumar,andJ.Ponce.ReconstructionofHOTcurvesfromimagesequences.In
CVPR93,pages2026,1993.
[14]ArthurR.PopeandDavidG.Lowe.Probabilisticmodelsofappearancefor3Dobjectrecognition.Inter-
nationalJournalofComputerVision,40(2):149167,2000.
[15]RajeshP.N.RaoandDanaH.Ballard.Objectindexingusinganiconicsparsedistributedmemory.
TechnicalReportTR559,1995.
[16]JianboShiandCarloTomasi.Goodfeaturestotrack.InIEEEConferenceonComputerVisionandPattern
Recognition(CVPR’94),Seattle,June1994.
|
|