• Aucun résultat trouvé

A MATLAB-based eye tracking control system using non-invasive helmet head restraint in the macaque

N/A
N/A
Protected

Academic year: 2021

Partager "A MATLAB-based eye tracking control system using non-invasive helmet head restraint in the macaque"

Copied!
10
0
0

Texte intégral

(1)

A

MATLAB-based

eye

tracking

control

system

using

non-invasive

helmet

head

restraint

in

the

macaque

Paolo

De

Luna

1

,

Mohamed

Faiz

Bin

Mohamed

Mustafar

1

,

Gregor

Rainer

VisualCognitionLaboratory,DepartmentofMedicine,UniversityofFribourg,CheminduMusée5,FribourgCH-1700,Switzerland

h

i

g

h

l

i

g

h

t

s

•WehavedevelopedaMATLABbasedeyetrackingcontrolsystemwithexcellenttimingperformance.

•Theexperimentwasperformedusingnon-invasiveheadrestraint.

•Thesystemperformancereportedadequatestabilityinbothshortandlongtimescales.

Background:Trackingeyepositionisvitalforbehavioralandneurophysiologicalinvestigationsinsystems andcognitiveneuroscience.Infraredcamerasystemswhicharenowavailablecanbeusedforeyetracking withouttheneedtosurgicallyimplantmagneticsearchcoils.Thesesystemsaregenerallyemployedusing rigidheadfixationinmonkeys,whichmaintainstheeyeinaconstantpositionandfacilitateseyetracking. Newmethod:Weinvestigatetheuseofnon-rigidheadfixationusingahelmetthatconstrainsonly gen-eralheadorientationandallowssomefreedomofmovement.WepresentaMATLABsoftwaresolution togatherandprocesseyepositiondata,presentvisualstimuli,interactwithvariousdevices,provide experimenterfeedbackandstoredataforofflineanalysis.

Comparisonwithexistingmethod:Oursoftwaresolutionachievesexcellenttimingperformancedueto theuseofdatastreaming,insteadofthetraditionallyemployeddatastoragemodeforprocessinganalog eyepositiondata.

Results:Wepresentbehavioraldatafromtwomonkeys,demonstratingthatadequateperformancelevels canbeachievedonasimplefixationparadigmandshowhowperformancedependsonparameterssuch asfixationwindowsize.Ourfindingssuggestthatnon-rigidheadrestraintcanbeemployedforbehavioral trainingandtestingonavarietyofgaze-dependentvisualparadigms,reducingtheneedforrigidhead restraintsystemsforsomeapplications.

Conclusion:Whiledevelopedformacaque monkey,oursystemofcoursecanworkequallywellfor applicationsinhumaneyetrackingwhereheadconstraintisundesirable.

1. Introduction

Thescleralsearchcoil(Judgeetal.,1980;Robinson,1963)has

beenusedextensivelysinceitsintroductioninthe1960sfor

mon-itoringofeyepositioninthecontextofstudiesofvisually-based

behaviorinmacaquemonkeys.Itsuserequiresa surgical

inter-vention,whereaninsulatedwirecoilisattachedtothescleraby

glueorsutures,andthewireislinkedtoaconnectorimplanted

∗ Correspondingauthor.Tel.:+41263008689;fax:+41263009734. E-mailaddress:gregor.rainer@unifr.ch(G.Rainer).

1 Theseauthorscontributedequallytothiswork.

ontheanimal’shead.Suchimplantshavealimitedlifespanand

oftenrequire repairor replacement,potentially involving

addi-tionalinterventions.Inthelate1990sopticaleyetrackingsystems

becameavailable;theytrackthepositionofthepupililluminated

byanappropriatelightsourceusinganinfraredsensitivecamera.

Infraredeyetrackingsystemsarenon-invasiveanddonotrequire

anysurgical procedures,and are therefore increasinglyused in

behavioralandneurophysiologicalstudies.Itisimportanttonote

thatwhile theeyetrackingitself isnon-invasive,it isgenerally

usedin conjunction witha head-postthat is implantedon the

skullof themonkey andattachedtotheexperimentalsetupby

arigidconnection.Thehead-postensuresthatthepositionofthe

headandparticularlytheeyeremainsfixedduringthebehavioral

Published in -RXUQDORI1HXURVFLHQFH0HWKRGV±

which should be cited to refer to this work.

(2)

session.Analternativetothismightbeanon-rigidhead

restrain-ingsystem,whichmaintainstheheadinanapproximatelyconstant

positionwithoutarigidconnectiontothesetup,thuspermitting

somemovements.Indeed,eyetrackinginhumansubjectsoften

employsnon-rigidheadrestraintusinga chin-restandbitebar.

Eye-trackingusinghead-mountedeyetrackers,asitisalsoused

inhumansubjects,arenot easilytoleratedbymonkeysubjects.

Non-rigidheadrestraintwouldallowthetrainingofmonkeyson

gazedependentbehavioralparadigmswithouttheneedof

surgi-calinterventionssuchthatthesecouldbepostponedtoafterthe

trainingiscompleted.Arecentpublicationhasdescribeda

thermo-deformableplastichelmetthatcanbeeasilyadaptedtotheshape

ofindividualmonkeys’head(MachadoandNelson,2011).Inthis

study,wedemonstratethateyetrackingofsufficientqualityfor

manyapplicationsispossibleusingsuchahelmet-basedrestraint

system,anddescribeperformancecharacteristicsandlimitations

ofeyetrackingundertheseconditions.

Forcontrollingthesequenceofeventsduringavisualtask,

com-putersoftwareisnecessarytogathereyepositiondatafromtheeye

tracker,presentvisualstimuli,anddeliverrewardsatappropriate

timeswhileatthesametimeprovideonlinefeedbacktothe

exper-imenter.Therearenumerousproprietaryaswellasopen-source

softwarepackagesavailableforthispurpose.Foreaseof

program-mingandadaptationofcode,solutionswritteninawidelyused

high-levelinterpretedprogramminglanguagesuchasMATLABare

preferable,particularlysincePsychophysicstoolbox,awidelyused

packageforvisualstimulusdisplay isalsoavailablein this

pro-grammingenvironment.UsingMATLABtoperformonlinecontrol

isinherentlysomewhatproblematic, becausethis programming

languagewasnotdesignedforreal-timecontrol.Nevertheless,we

areawareoftwosoftwaresolutionsthatemployMATLABtogether

withPsychophysics Toolbox for the control of gaze-dependent

behavior(AsaadandEskandar,2008;Asaadetal.,2013;Meyerand

Constantinidis,2005).Ourownsoftwareimplementationisbased

onthesetwosystems.Wehaveintroducedseveralnewelements

thatenhanceperformanceandreliability,makingaparticulareffort

toproducecompactcodethatcanberelativelyaccessiblefor

begin-ners.

2. Materialsandmethods

2.1. Subjects

Two11year-oldmalemacaquemonkeys(Macacafascicularis),

referredtoasMonkeyC(6.8±0.2kg,mean±standarddeviation)

andMonkeyD(8.2±0.1kg),participatedintheexperiments.

Ani-mals were maintained on a diet of fresh fruit, vegetablesand

monkeychowwithwateravailableadlibitum.Duringthe

exper-iments,foreachsuccessfulfixationthemonkeyswererewarded

witha45mgpurifieddustlesspellet(Bananaflavor5TUQtab,

Test-Diet,1050ProgressDrive,RichmondIN48384,USA),deliveredvia

ametaltube.Monkeysperformedanaverageof183±11(S.E.M.)

trialsperexperimentalsession, andhadnohead-postsorother

implants.Headrestraintwasachievednon-invasivelyusing

ther-moplastichelmets(Uni-frameThermoplasticsMTAPUI2232,Civco

MedicalSolutions)adjustedtoeachindividualmonkey(Machado

andNelson,2011).Thethermoplasticmaterialwasmadeflexible

bysubmergingthesheetina80◦Cwaterbathfor2min.Duringthe

next2min,beforethematerialregaineditshardness,thehelmet

waspre-shapedonatoydollhead,finelyadjustedbyhand,andtwo

openingsforeyesandmouthwerecutwithscissors.Final

adjust-mentofthehelmettoeachmonkeydidnotrequiresedationofthe

animals,andwasperformediterativelysuchthatagoodfitthat

permittedchewingmovementswasobtainedafteraboutfour

iter-ations.Importantly,thehelmetfitwasadjustedtogivemonkeys

acertainamountoffreedominheadposition,bothlaterallyaswell

asvertically,toallowforchewingofrewardsaswellas

minimiz-ingdiscomfort.Allanimalprocedureswereinfullcompliancewith

SwissandEuropeanUnionanimalexperimentalregulationsand

approvedbytheFribourgcantonalveterinaryauthorities.

2.2. Setup

Themonkeywasseatedinaprimatechair(custom-madeat

Uni-versityofFribourg),onawheeledtrolleyatthebasewhichcouldbe

lockedinafixedpositioninsideawoodenrecordingboxasshown

inFig.1A. Aninfra-red camera(ISCANETL-200system,x2 lens

attached)wasmountedatanelevationangleof27◦atadistanceof

65cmfromtheanimal.Theinfraredlightsourcewasattachedto

theleftsideofthecameraobjectivelens,andtiltedforoptimal

illu-minationofthelefteye(Fig.1B).Theinsideoftheboxwascovered

withmatteblackplasticpanels.Visualtargetswerepresentedona

cathoderaytube(CompaqP90019-inchCRTmonitoratresolution

of1280×1024pixels,withtherefreshrateof75Hz)monitor

pos-itioned65cmdirectlyinfrontoftheanimal.Weemployedblack

squaresof0.5◦diameter(20×20pixels),arrangedalongacircleof

radiusof6◦regularlyspaced45◦ aroundtheninthsquareinthe

centerofthecircle(Fig.1C).Targetswerepresentedona

back-groundofintermediateintensity(50%grayscale),correspondingto

aluminanceof13cd/m2,measuredusingaMinoltaTVCA-IIcolor

analyzer.Alternatively,sixred0.8◦diameterlightemittingdiodes

(LED)onablackpanelmountedatadistanceof30cmfromthe

animalcouldalsoserveasvisualtargets(Fig.1D).Duringthe

behav-ioralsessions,auditorywhitenoise wasdeliveredviaa speaker

insidetheboxtomaskexternaldistractions.Thespeakerwasalso

usedtodelivera1kHzpuretonefor1sasauditoryfeedbackfor

eachcorrecttrial.

2.3. Hardwareconfiguration

We used an infrared eye tracking system (ISCAN ETL-200).

Indetail, thecamerawasconnectedviaaPCI acquisitionboard

(ISCANRK8X6PCI-0)toacomputerrunningtheeye-track

acquisi-tionsoftware(DQWversion1.20N).Thecomputerwasequipped

witha 2.4GHzIntel Core2Duoprocessor, 4GBofRAM anda

3-channel analog output card. The Eye-Tracking System’s analog

output was interfaced via a single PCI-6221 acquisition board

(NationalInstruments)connectedtoaBNC-2110connectorblock

(NationalInstruments)toasecondPCthathostedtheExperiment

ControlSystem(Fig.2).Weusedtwoanalogchannelstorelay

hor-izontalandverticaleyeposition.TheISCANsystemwasconfigured

totrackeyepositionusingthedifferencebetweenthepupiland

cornealreflection,asiscommonlydoneineyetrackingsystems

withoutrigidheadrestraint.Itiscrucialtousethisdifference

sig-nalfornon-invasive eyetracking,asit is lesssensitivetohead

translationsthantherawpupilsignal.

TheExperimentControlSystem(HPCompaq8200Elite,

Win-dows 7) wasequipped with a 3.3GHz Intel i5-2500 processor

and 4GB of memory. An NVidia GeForce 7300GT video card

wasinstalled for managing two displays at a resolution up to

1280×1024pixels.Thesubjectdisplayviewedbytheanimalwas

duplicatedusingaVGAsplitter tobevisiblealsotothe

experi-menter.Inaddition,anexperimenterdisplaywasusedtoprovide

onlinefeedbackabouttheanimal’sbehaviortotheexperimenter.

The“Extend these displays”option of Windows 7 wasusedto

simultaneouslymanage experimenterand subject displays.The

PCI-6221boardwasalsousedtoilluminatevisualtargetLEDsand

todeliverrewardpellets(ENV-203-45pelletdispenser,Med

Asso-ciates,St.Albans,VT05478,USA).Anoverviewoftheconnections

betweenthesetupcomponentsisshowninFig.2.Tooptimize

per-formanceoftheISCANandexperimentalcontrolsystemPCs,all

(3)

Fig.1. Behavioralexperimentsetup.(A)Sideviewoftherecordingboxwherebehavioraldatawasrecorded.Monkey’sheadpositionismarkedbyaneyesymbol.Onthe lefttheCRTscreenand,ontopofit,theinfra-redcameraaredrawn;thewhiterectangleinthemiddleistheblackLEDpanel.(B)Topviewofrecordingbox.LEDpaneland CRTscreenhorizontalsizewere22◦and16respectively.TheCRTmonitor,theLEDpanelandtheinfra-redcamerawerecenteredonmonkey’sheadposition(eyesymbol).

(C)PositionsofvisualstimulionCRTmonitor.Eightblacksquaretargetsalonga6◦circlesurroundaninthblacksquaredcentralfixationpoint.Eachstimulusisabout0.5

(20×20pixels).(D)PositionsofvisualstimulionLEDpanelarrangedontworowsofthreeelements.EachstimuluswasaredLEDofapproximately0.8◦diameter(3mm).

non-criticalprocessesincludinginterface-,security-,network-and

devices-relatedoneswereterminatedandneithercomputerwas

connectedtoanynetwork.

2.4. MATLABconfiguration

AllthedatawereacquiredbyandstoredintotheExperiment

ControlSystemusingacustom-madescriptruninMATLAB

(ver-sionR2010a,TheMathworksInc.).Thescriptusedfunctionsfrom

theMATLABDataAcquisitionToolbox(DAQ,version2.16)and

Psy-chophysicsToolbox(PTB)version3.0.11(Brainard,1997;Kleiner

etal.,2007;Pelli,1997).Thescriptpresentsaseriesofrandomly

selectedsinglevisualtargetsandprovidesrewardtotheanimalif

thattargetisfixatedforapredeterminedperiodoftime.Thistype

Fig.2. Overviewofconnectionsofhardwarecomponents.Eyepositionsignalswere computedbyISCANsoftwareandsentviaaNIBNC2110connectorblocktoaNI BNC6221acquisitionboardtotheExperimentControlsystem.Thegazeposition wasmonitoredwithMATLABduringpresentationofvisualstimulieitheronaCRT monitor(andtoamirrordisplayvisibletotheexperimenter),oronanLEDpanelvia theNIPCI6221digitaloutputchannels.MATLABalsopresentedthelastacquired gazepositiontotheexperimenter.FinallyonedigitaloutputchanneloftheNIPCI 6221wasusedforcontrollinganelectricvalveforrewarddelivery.

ofscriptisroutinelyemployedforthecalibrationofeyeposition.

Threefragmentsofthesourcecodeareprovidedbelow,illustrating

importantelementsofDAQconfiguration,databufferingandonline

behavioralcontrol.Thefullscriptcanbeobtainedfromour

web-site( http://www.unifr.ch/inph/vclab/home/internal/eye-tracking-software).

2.4.1. DAQconfiguration

Ideally,onewantstostreamthemostrecentsampleacquiredby

theboardassoonaspossibleforonlinecontrolwhileatthesame

timestoringsamplesontheboardtobepickedupattheendofeach

behavioraltrial.However,simultaneousstoringandstreamingof

datathroughtheDAQtoolboxisassociatedwithsevere

perfor-manceimpairments(seeSection3).Forthisreason,oursolution

focusesonlyonstreaminganddoesnotrelyonstoringdataonthe

board.Thefollowingcodesnippetillustratesoursolutionforanalog

dataacquisition. ai=analoginput(’nidaq’,’Dev1’); addchannel(ai,[01]); ai.SampleRate=1000; ai.SamplesPerTrigger=2; ai.TriggerRepeat=0; start(ai); startLoop=GetSecs;

while(GetSecs−startLoop<time4fix), thisSample=getsample(ai); ...

end

Theboardisconfiguredtostoreonlytwosamples,the

mini-mumpermittedbyMATLAB.Afterstartingdatalogging,onlythe

firsttwosamplesarestoredontheboard,whereasallthe

subse-quentsamplesarenotstoredandthuscannotberecoveredlateron

byagetdatafunctioncall.However,subsequentsamplescanbe

pre-viewedusingthegetsamplefunction.Inthiscase,themostrecently

(4)

acquiredsampleismadeavailabletoMATLABwithout

compro-misingperformance.Ourapproachthusoptimizesdatastreaming

toMATLAB.Asaresult,onlythedatathatareactuallystreamed

toMATLABusinggetsampleforonlinecontrolcanberetainedfor

subsequentdataanalysis.

2.4.2. Circularbuffer

Since infrared eyetrackingdata canbenoisy,it is usefulto

includesometemporalaveragingtoreducethevariabilityinthe

acquireddata.ForthispurposeweuseaFirst-InFirst-Out(FIFO)

circularbuffer,withtheparameterbuffer lengththatdetermines

overhowmanysubsequentsamples,acquiredusinggetsample,the

averagingprocesstakesplace.

bufferlength=25;

Buffer=zeros(bufferlength,2);

function[Buffer,eyeX,eyeY,sampleInfo]=bufferizeSamples(ai,Buffer,Gains,Offsets) %Puttogethertheinformationaboutthelastsample

timestamp=GetSecs; thisSample=getsample(ai); sampleInfo=[thisSampletimestamp];

%Updatethebufferwiththelastdatapointandtakeanewaveragepoint Buffer=[Buffer(2:end,:);thisSample];

avgBuff=mean(Buffer,1); %Convertvoltstodegrees

eyeX=(avgBuff(1)−Offsets(1))*Gains(1); eyeY=(avgBuff(2)−Offsets(2))*Gains(2); end

In ourMATLABscript,thisfunctioniscalledseveralhundred

timeseachsecond,andtimestampedusingthePTBGetSecs

func-tion.Thelastsamplereplacesthefirstoneinthebufferandanew

gazepositioniscomputedbyconvertingtheaverageinputsignal

fromvoltagetodegrees ofvisualangle,usingadhocgainsand

offsetsforboththeaxes.

2.4.3. Mainloop

Thefollowing codefragmentillustratesthemainloopofthe

behavioralcontrolsystem,wherethesystemchecksforwhether

theeyepositioniswithinaradiusofthetargetposition.Itusesthe

circularbufferanddatastreamingelementsdescribedabove.

startLoop=GetSecs;

whileGetSecs−startLoop<time4fix %1)Getandstoreanewsample

[Buffer,eyeX,eyeY,sampleInfo]=bufferizeSamples(ai,Buffer,Gains,Offsets); Data(trial).EyeDataX(index)=sampleInfo(1);

Data(trial).EyeDataY(index)=sampleInfo(2); Data(trial).Time(index)=sampleInfo(3);

index=index+1;

%2)Updateeyepositiononexperimenter’sdisplay set(eyePosH,’XData’,eyeX,’YData’,eyeY); drawnow;

%3)Checkforfixation

ifsqrt((TargetX−eyeX)∧2+(TargetYeyeY)2)<=FixWinRadius

Data(trial).FixStart=GetSecs; break;

end end

Theloopconsistsofthreecomponents:(1)usingthecircular

buffer,anewgazepositionisevaluatedaspreviouslydescribed,

andthelastsampleandcorrespondingtimestamparestoredina

trial-wisemannerinpre-allocatedmatrices;ithastobenotedthat

wealsostoredthelengthofthebuffer,thegainsandtheoffsetsin

ordertoreconstructallthegazepositionsfromtherawdata;(2)

theexperimenter’sdisplayisupdatedwiththenewgazeposition;

(3)ifthegazepositionisinsidethefixationwindowtheloopis

interrupted;inthatcaseasimilarloopwillcheckifthesubjectis

keepingthefixation.Werefertothiscodefragmentwithallofits

threepartsasthe“Mainloop”.

2.5. Eyetracking

Weusedasimplefixationtasktocalibratetheeyeposition

sig-nal,andtotestthesystemperformanceasafunctionofseveral

parameters.Asdescribedabove,weemployedasetofnine

tar-getsarrangedinacircularfashiononaCRTmonitor,aswellassix

LEDtargetspositionedonapanelmountedinfrontoftheanimal.

Monkeysneededtofixateasingletargetforabriefperiodtoobtain

arewardpellet.Wevariedboththelengthofthecircularbuffer

(values:1,10,15,20and25samples)andtheradiusofthe

fixa-tionwindow(values:1.5◦and2◦ fortheCRTscreenand2◦,2.5◦

and3.5◦fortheLEDpanel).LargervalueswerechosenfortheLED

panelfixationwindowbecausetargetsweremoreeccentricinthat

case.Duringblocksof50trials,arandomcombinationoffixation

windowradiusandbufferlengthwaschosen.

Eachsessionstartedwiththemonkeyoutsidetherecordingbox

withoutthehelmet.Thenthehelmetwasfixedontheprimatechair

andthemonkeywasmovedinthebox.Atthebeginningofeach

sessionthecamerawascenteredonthemonkey’slefteyeandabout

20trialswereusedtoadjustthegainsandtheoffsetsforthe

hor-izontalandverticaldirectionsusinga25-samplebuffer.Onlyone

setofvisualstimuli(CRTscreenorLEDpanel)waspresentedduring

thesamesession.

Duringeachtrial,aftertheappearanceofthetargetonthepanel

(LED)orCRTscreen(squaredot),themonkeyhadtostartthe

fix-ationwithin2sandholditfor200mstoobtainarewardpellet

withsimultaneousauditoryfeedback.Theinter-trialintervalwas

7s.Attheendofeachblock,themonkeywasmovedagainoutside

theboxandthehelmetwasremoved;theinter-blockintervalwas

about2min.

3. Results

Weranextensivetestsonoursystem,andobservedthatthe

choiceof MATLABDAQacquisitionparameters hadavery large

impact on system performance, with non-optimal parameters

introducing substantial lags between acquisitions of individual

samplesaswellastiminginaccuracies.

3.1. Timingconsiderations

We first compared the performance of DAQ storing mode

(parametersettingai.SamplesPerTrigger=Inf;)thathasbeen

previ-ouslyproposedinseveralreportstoDAQstreamingmodewithout

storing(parameterai.SamplesPerTrigger=2;seeSection2).In

stor-ingmodetheboardiscollectingdataandmakingthemavailable

toMATLABatthesametime,whereasinstreamingmodeallthe

samplesarediscardedbytheboardand canonly bepreviewed

withgetsamplefunctioncalls.Forthistest,weusedtheMainloop

describedabove,andswitchedfromdatastoringtodatastreaming

attimet=0ms.Asynthesizedsinewaveat50Hzproducedbya

signalgenerator(GWInstekSFG2004)wasusedasaninputinthis

caseinsteadofaneyepositionsignal.Importantly,thegetsample

functionwasusedinbothmodestoobtainadatasampleforonline

eyemovementcontrol.TheresultsareshowninFig.3A.Clearly,

callinggetsampleindatastoringmodeisassociatedwith

unpre-dictablelagsbetweentheinputsignalandtheacquiredsample.

Theselagsarefurthermorevariable,andwefoundthemtobein

rangebetween10and40ms.Inaddition,previewingdatausing

getsampleindatastoringmodeproducedwidespreadsample

rep-etitions.Individualsamplescouldberepeateduptoonethousand

timesormore,becauseatthetimeofthegetsamplecalltheboard

isbusystoringdatatomemoryandthusreturnsthesame

previ-ouslyacquiredsample.Usingdatastreamingmodetheseproblems

disappear, and samples are now lying very close to the input

(5)

Fig.3.(A)Twocyclesofasynthesized1Hzsinusoidalwaveareshownintheleftpanel.Attimet=0theacquisitionboardstoppedsimultaneousdatastoringandstreaming toMATLABworkspace,anddiscardedallnewcollectedsamplesfromitsmemory.AllpresenteddatapointswerecollectedinMATLABcallingthegetsamplefunction.In therightpanel,azoomedinviewoftimet=0isshown.Duringcombinedstoring/streamingmode,samplescanberepeatedupto1000timesandtheyareacquiredafter avariabledelayofabout50ms.Instead,duringsteamingonlymode,allsamplesliecloselytotheinputsignalwithnoredundancyortimelag.(B)Stackedbarsshowtime lagsbetweentwoconsecutivesamples.Notethatthesignalsamplingratewas1kHzandtheboard’smemorysizewassetto2samples.RunningtheMainloopwithdisplay updateateachcycle(darkgraybars)generatedalagofabout4msofwhichabout3.5mswereneededtothedisplayupdateitself(lightgraybars).Updatingthescreenevery 100ms(blackbars)loweredtheaveragelagtoabout1.3ms,ofwhich1mswasauser-definedpausetobesuretoacquireanewsample.Notethatwhentheexperimenter displayupdateoccurred,anaveragelagofabout4.3mswasobserved.

signal,certainlywithoutunpredictablelagsorrepetitions.Forthese

reasons,weabandoneddatastoragemodeandreliedonlyondata

streaming.Itfollowsthatattheendofeachtrial,getdatacannot

beusedtorecoverallofthesamples,andonlythesamples

gath-eredusinggetsampleareactuallyavailabletobestoredforoffline

analysis.

Toacquireand displayone eyepositionsample,oursystem

requiredadelayofabout4ms,asshowninFig.3Bforthehistogram

correspondingto“Mainloop(frequentupdate)”.Inthiscase,the

experimenterdisplayisupdatedforeveryacquiredsample.The

slowestcomponent oftheloop is accountedfor by the

experi-menter’sdisplayupdate,whichrequiresanaveragedelayofabout

3.5msforeachcalltoMATLAB’sfunctiondrawnow.Runningthe

Mainloopwithinfrequentscreenupdate(forexampleonceevery

100ms),itispossibletoachieveaperformanceofunder1.5msfor

mostsamples.Inthiscase,onlythefewsamplesthatareactually

drawnonthescreenexhibitadelayofabout4.5ms.

Usingfrequentscreenupdatemode,whichdisplaysallacquired

samplesontheexperimenterscreen,oursoftwaresolutionexhibits

adelayof under5ms,whichwe considersuitablefor

monitor-ing eyeposition signals acquired withan infrared camera. We

thereforeusedthisconfigurationinthesubsequentexperiments

detailedbelow.

3.2. Circularbuffering

Itisusefultoemploydatasmoothingtoreducespatialvariability

ofgazepositionsandattenuatetechnicalnoisecomingfromtheeye

tracker.Thiscanbeachievedbyaveragingoverthelastnacquired

samplesstoredina(FIFO)circularbuffer.Aconsequenceofsignal

averagingisthatanalogsignalsarealsotemporallydelayed,and

thisdrawbackbecomesrelevantforstudyingtransientandrapid

phenomena,likeforexamplesaccadiceyemovements.Weuseda

synthesized0.1Hzsquarewave(GWInstekSFG2004)to

approxi-mateasaccadiceyemovement,andexaminedthetemporaldelays

introducedbycircularbuffering.Forabufferlengthof10samples,

anabruptchangeintheanalogsignalinputattime0generatesa

linearramplastingabout40msinoursystem,andthusintroduces

alaginthedetectionofthesignaltransitionofatleastt=20ms

(Fig.4A)Dependingonthefixationwindowsize,tcanbelarger

butcannotexceed40msforabufferlengthof10.Examiningthe

dependenceoftheminimumlagtonbufferlength,weobserved

(6)

Fig.4. (A)Theeffectofusingacircularbufferinonlineapplicationsisshown.The inputsignalwasasynthesized0.1Hzsquarewave.Eachsquaresymbolrepresents themeanvalueofthelast10samplesstoredinacircularbuffer(seeSection2). Thetimelineiscenteredbetweenthelastsampleofastationaryperiodandthefirst sampleofthetransitionperiod.Werefertominimumtimedelayfordetectinga saccadeast.(B)Increasingbufferlengthlinearlyincreasest.Notethataveraging acrossabufferwith1sampleislikeusingrawdata.(C)Powerdecayrelationship (R2=0.99)betweencoefficientofvariation(C

v)ofGaussiannoisewaveandbuffer

length.

alinearrelationship,asshowninFig.4B.Thisillustratesthat

tem-porallaggrowssystematicallywithbufferlength.Atthesametime,

thevariabilityofanalogsignalsisreducedwithincreasingbuffer

length.Weillustratethisbycomputingthecoefficientofvariation

(/)ofdata(100,000data-points)sampledfromanormal

distri-bution(=1,=1)thataresmoothedwithvariousbufferlengths

(Fig.4C).Itcanbeseenthatvariabilityquicklydecaysforbuffer

lengthsuptoabout15,andthencontinuestodeclinemore

grad-ually.Thissuggeststhatacircularbufferofatleast15samplesis

advisablewhenthereissubstantialnoisepresentintheinputdata.

Takentogether,averagingacrosslargecircularbuffersisoptimal

forreducingtechnicalvariabilityininfraredeyetrackinganalog

inputs. However,it introducestemporal lags in detectability of

transientsignalchanges,suchasthosethatoccurduringsaccadic

eyemovements.Theparameterchoiceforcircularbufferlength

dependsonatrade-offbetweenthesetwoeffects.

3.3. Behavioraldata

Thegazepositionsoftwomonkeys,CandD,weremonitored

duringafixationtask,whilethesubjectsfixatedatargeteitheron

aLEDpaneloronaCRTscreenfor200ms(monkeyC:50blocks

LEDpanel,26blocksCRTscreen;monkeyD:51blocksLEDpanel,

28blocksCRTscreen).Eachblockoftrialsconsistedof50trials

withrandomizedfixationlocationandafixedcombinationofbuffer

lengthandfixationwindowradius.Datapresentedherearebased

onabout4blocks oftrialsforeach parametercombination.The

monkeysweresubjectedtonon-invasiveheadrestraintusingan

individualizedthermoplastichelmet(seeSection2),thatallowed

uptoabout1cm of lateralandverticalhead movement.These

smalltranslationalmovementswerecompensatedusingthe

sig-naldifferencebetweenpupilcenterandcornealreflectionanddid

notaffecttheperformanceofthesystemsincethepositionofthe

eyesremainedwithinawindowsuitableforeyetracking.Notethat

thereisatrade-offbetweenthemagnitudeofpermissible

trans-lationalmovementsandthespatialaccuracyoftheeyetracker,

becausetheeye trackerfield of viewneedstobe expandedto

allowlargertranslationalmovements,reducingspatialresolution.

Ourchosenparametersrepresenta compromisebetween

resid-ualpermissiblemovementsandeyetrackingresolution.Monkeys

werenotpre-trainedonfixationorothertasks,sothattheycan

beconsiderednaïve. The resultsof thebehavioral experiments

areshowninFig.5.Clearly,systemperformanceispoorwithout

circularbuffering(bufferlength=1sample),wheremanybroken

fixationerrorsoccur.Asexpected,largerfixationwindows,aswell

aslongerbufferlengths,ledtohigherbehavioralperformancefor

both thesets of visualstimuli. A two-way analysis of variance

(ANOVA)revealedasignificanteffectoffixationwindowsize(FWS)

andbufferlength(BL)onbehavioralperformance(p<10−9)with

nointeractionbetweenthesefactors(p>0.1).Toobtainhigh

spa-tialresolution, temporal accuracy hastobe sacrificed and vice

versa.Forexample,monkeyDperformedsimilarlyontheLEDpanel

taskforparametercombinationsofFWS=3.5◦/BL=10samplesand

FWS=2◦/BL=25samples.ParametercombinationsforFWSandBL

willlikelydependonparticularexperimentalrequirementsandcan

beoptimizedaccordingly.

Toillustratetheperformanceofoursystem,weshowhorizontal

andverticaleyepositiontracesaroundtimeoffixationacquisition

(Fig.6AandB)forsixteencorrecttrials(FWS=2◦,BL=15samples)

takenfromtheLEDpaneltaskforthecentraltargetposition.The

tracesshowsaccadiceyemovementsmadetothetargetposition

ataroundtimet=0msfromavariableinitiallocationmostlyinthe

upperrightvisualfieldquadrant.Afteracquisitionoffixation,the

eyepositionremainsinsidethefixationwindowof±2◦.Toassess

sample-to-samplevariabilityofeyepositionestimatesduring

fixa-tion,wecomputedthedistanceofdatasamplesfromthemeaneye

positionduringthefixationtimeperiodbetween0msand200ms

foreachtrial.Thissample-to-samplevariabilityquantifieseye

posi-tionsignalstabilityatshorttimescales.Weestimateamedianvalue

of0.09±0.001◦ for both horizontaland verticaldimensions.To

assessthestabilityofoursystematalongertimescale

correspond-ingtothecourseofanentirebehavioralsession,weexaminedthe

(7)

Fig.5. PerformanceoftargetfixationformonkeysC(AandC)andD(BandD)duringblocksof50trials.Differentlinesrepresentdifferentfixationwindowradiiasshown infigureslegend.TargetswerepresentedonanLEDpanel(AandB)oraCRTmonitor(CandD)forbothanimals.

meangazepositionduringthefixationperiodsofsubsequent

tri-alsforthecentraltargetpositionsonlyinCRTandLEDpaneltasks.

Thesedataareshownforanexamplebehavioralsession lasting

aboutonehourinFig.6CandD.Alinearregressionrevealedthat

bothhorizontalandverticaleyepositiondriftedonlylittle

dur-ingthecourseofthis session,asshownbytheslopeparameter

estimates(mhorizontal=0.08◦/handmvertical=0.20◦/h).Ananalysisof

eyepositiondriftacrossallbehavioralsessionswiththreeormore

blocksisshowninFig.6E.For73%ofthesessions,thedriftwasless

than1◦/hbothhorizontallyandvertically.Theremainingsessions

exhibitedlargerdrift,whichmightnecessitatemanual

readjust-mentofgainand/oroffsetcalibrationbetweenblocksoftrials.Of

interestisalsothegenerallylargervariabilityingazepositionin

theverticalcomparedtothehorizontaldimension,whichisquite

noticeableinFig.6CandD.Toquantifythiseffect,weestimated

thedistanceofthewithin-trialmeanfixationlocationsfromthe

within-sessionmean.Thetrial-to-trialvariabilitywascomputed

asthemedianvalueofthesedistances.Consideringthe

popula-tionbehavioraldata(Fig.6F),theaveragehorizontaltrial-to-trial

variability(0.32±0.05◦)wasindeedsmallerthanthevertical

vari-ability(0.47±0.08◦)asassessedwithapairedt-test(p<0.001).We

attributethis differencetolickingmovementsinanticipationof

rewards,thatcausepreferentiallyverticalratherthanhorizontal

headdisplacementsandthereforeintroducenoisemostlyinthe

verticaldimension.

4. Discussion

Thecurrentstandardmethodforeyetrackingin non-human

primatesresearchrequires invasiveprocedureslikethe

implan-tationsof a fixedhead-postand a metal wire inthe sclera. As

(8)

Fig.6. Sixteentrialsareshown(AandB).Attimet=0targetfixationwasacquired.Horizontaldashedlinesaretheboundariesofthefixationwindowforcentraltargetonthe LEDpanel.(CandD)EachsquaresymbolrepresentstheaveragegazepositionduringthefixationtimeforthecentralpositionontheLEDpanelrecordedduringanexample session,respectivelyforthehorizontalandverticaldimension.Timelinestartsatthebeginningofthefirstblock;eachclusterofpointsispartofa50-trialsblock.Foreach sessionwithmorethanthreeblocks,driftandvariabilityofaveragegazepositionsduringfixationperiodsweresystematicallymeasuredcalculating,respectively,theslope ofthelinearregressionline(grayline)andstandarddeviationofthemeanpositions.(E)Distributionoftheabsolutevaluesofregressionlinesslopesforbothdirections.The driftwaslessthanonedegreeperhourformorethan73%ofthesessionsinbothdirections.(F)Distributionofstandarddeviationsofmeangazepositionsduringfixation periodsisshownforbothdirections.Valuesarelessthan0.5◦moreofteninthehorizontaldirection(52%ofthesessions)thantheverticalone(21%ofthesessions).The

distributionsaresignificantlydifferent(p<0.001).

(9)

recentlyshown, non-invasiveheadrestraintcanbeachievedby

employingaplastichelmet,adjustedtoeachindividualanimalthat

allowslimitedheadmovementswhileconstrainingtheoverall

ori-entationofthehead(MachadoandNelson,2011),andaninfrared

camera-basedsystemfortrackingeyeposition.Thisapproachis

howeveralsouseful forneurophysiological work,sincesurgical

procedurescanbecarriedoutafterappropriatebehavioraltraining,

reducingthedurationthatdevicesneedtoremainimplantedonthe

animaltoachievetheexperimentalobjectives.Forthebehavioral

datareportedhere,monkeyswererewardedwitha45mgpelletfor

correcttrials,andmadechewingmovementsduringthe

consump-tionoftheserewards,aswellasfrequentlyattimesduringthe

taskandpriortorewarddelivery.Nevertheless,thissystemcan

functiondespitethesechewingmovements,mainlybecausethe

infraredeye-trackingdeviceemploysthedifferencesignalbetween

thepupilcenterandthecornealreflection,whichisinsensitiveto

chewingaswellastranslationalheadmovements.Wenoticed

how-ever,thatspatialaccuracyofoursystemisgreaterinthehorizontal

thanintheverticaldimension;aneffectthatwepartiallyattribute

tothemorefrequentoccurrenceofverticalcomparedtohorizontal

headmovementsbothinthecontextofchewingaswellas

gen-eralbehavior.Inadditionhowever,theplacementofthecamera

isnon-optimalfortrackingverticaleyemovements.Wechosethis

positionbecausethespacebelowthemonitor,whichisgenerally

recommendedforpositioningthecamera,isoccupiedby

experi-mentaldevicesinoursetup.Onemightconsideranabove-mounted

camerainconjunction witha45◦ slantedmirrorasan

alterna-tiveconfiguration(Kimmeletal.,2012)toimproveverticalgaze

tracking.

Our behavioral control system is implemented in MATLAB

(Mathworks,NatickMA,USA)andusestheMATLABdata

acqui-sitiontoolboxaswellasthefreePsychophysicstoolbox.Forthe

behavioraldatareportedhere,weuseda singlem-filescript of

about400linesofcode,whichcollectsanalogeyepositiondata

fromtheeyetrackerforonlinecontrolandstorage,displaysvisual

targets and monitors their fixation, delivers rewards, provides

onlinefeedbacktotheexperimenteraboutcurrenteyeposition

andtrialspecificdata.Ourimplementationissimilartothe“WaVE”

system(MeyerandConstantinidis,2005),inthatthecodeis

basi-callycontainedinasinglescript,whichhasadvantagesforeaseof

modificationsandadaptationsforuserswhoareskilledin

MAT-LAB.OtherMATLABsolutionssuchas“MonkeyLogic”(Asaadand

Eskandar, 2008; Asaad et al., 2013)provide a generalpurpose

environment that utilizesa multitude of scripts,and provide a

graphicaluserinterfaceformodifyingexperimentalparameters.

Animportantdifferencebetweenthesetwosystemsandour

imple-mentationconcernsthewaythatdataarecollectedfromthedata

acquisitionboard.Whereasboth“WaVE”and“MonkeyLogic”use

thecombinedstorageandstreamingmodeoftheboard,our

imple-mentationreliesuniquelyondatastreamingmode.Exclusiveuse

ofdatastreamingmodeavoidstiminginaccuraciesandextensive

datarepetitionthatcanoccurduringcombinedstorage/streaming

mode(Fig.3A).Itisamajoradvantageofoursystemthateye

posi-tiondatacansampledwithhightemporalaccuracyandwithout

repetition.Providingexperimenterfeedbackwheneveraneweye

positionsampleisacquired,oursystemachievesatiming

perfor-manceofabout4msperupdate(Fig.3B).Over3msofthisupdate

timeisneededtorefreshtheexperimenterdisplay,andifthis

dis-playrefreshiscarriedoutinfrequently,insteadofaftereachnewly

acquireddatapoint,thetimingperformancecanbeenhanced

sub-stantially.Inourcase,wecanachieveabout1.3msperupdateon

mosttrials,andabout4msforupdatesthatincludedisplayrefresh.

Notethatthe1.3msupdatetimeincludesawaitingperiodof1ms,

thatweintroducedbecausetheDAQsamplingratewassetto1kHz

anditthereforemakesnosensetocollectdataattimeintervals

below1ms.Withoutthewaitingperiod,oursystemalsoreaches

updateintervalsunder0.5msasreportedinotherstudies(Asaad

andEskandar,2008;MeyerandConstantinidis,2005),butmostof

theseupdatesarepickingupredundantdatafromtheDAQboard.

Wenotethattimingperformanceofthebehavioralcontrolsystem

generallydoesnotneedtoexceedthetimingperformanceofthe

infraredtrackingdevice.Inourcase,weusedacamerawith120Hz

refreshrate,correspondingtoabout8msbetweenframes,such

thata configurationwithfrequentexperimenterdisplayupdate

wassuitable forthishardware.Forcameraswithhigherrefresh

rate,onemayconsiderinfrequentexperimenterdisplayupdateto

improvesystemtemporalprecision.

Anotablefeatureofoursystemistheuseofacircularbuffer,

whichreducesnoiseintheeyepositionestimatesandenhances

spatialaccuracy.Becausethebufferingmustbecausalinonline

applications,itnecessarilyintroducesatemporaldelaythatis

pro-portionaltothebufferlengthover whicheyepositions samples

areaveraged.Thechoiceofbufferlengthwilldependonparticular

experimentalconsiderations,butwehavefoundthatvaluesaround

15samples,correspondingtoabout60ms,tendtoworkwellfor

thesimpleeyepositioncalibrationtaskinvestigatedhere.For a

givenbufferlength,thedetectiondelayforfixationonsetdepends

onthedistancebetweeninitialandtargeteyepositions,andwill

haveaminimumvalueoftasshowninFig.4B.Itisimportantto

rememberthatformanyapplications,thetemporaldelaycanbe

essentiallyequatedtoaprolongationoffixationtime.We

there-foredonotviewthetemporallagintroducedbycircularbuffering

asproblematicformostapplications.Anexceptiontothisaretasks

involvingsmoothpursuiteyemovements,whicharethusoutside

thescopeoftheproceduresdescribedhere.Notethatsince the

actualacquiredeyepositionsamplesandnotthecircularbuffered

valuesaresavedforlateranalysis,asimple(non-causal)moving

averagefiltercanbeappliedofflinetodeliveraveridicalestimate

ofeyepositionwithoutanytemporaldelays.Inapplicationswhere

fasterresponsetimesarenecessary,theinfrequentupdateoption

describedabovecanbeusedtoreducethetemporallagbyafactor

ofuptoabout3,orthecameracanbeplacedinamoreoptimal

positionsuchthatcircularbufferingmaynotbenecessary.

The necessity of circular buffering is underscored by the

behavioraldata.Withoutcircularbuffering,monkeysachieved

per-formancelevelsofonlyabout60%correctwithafixationwindow

size(FWS)of3.5◦onasimplefixationtask.Withcircularbuffering,

theFWScouldbereducedtovaluesaslowas2◦,andstillpermit

behavioralperformanceofabout80%correctinmostcases.Note

thatmonkeyswerenotextensivelypre-trainedonfixationtasks

forthesetests,sothattheirperformancemightimprovewith

addi-tionaltraining.Thereportedvaluesarecertainlysuitabletoserve

asabasis oftrainingmonkeysonawide varietyofoculomotor

paradigms.

We quantifiedthe stability ofour systemboth atshort and

longtimescales,andwereabletodemonstrategenerallyadequate

stability that permitsgatheringdata over periodsof aboutone

hourwithouttheneedforrecalibrationofanyeye-tracker-related

parameters.Foreyepositionsignalstabilityatshorttimescales,we

obtaineda valueof0.09±0.001◦ forsample-to-sample

variabil-ity,whichissimilartopreviousreportsforinfraredeyetrackingon

rigidlyhead-postedmonkeys(Kimmeletal.,2012).Foreyeposition

stabilityatlongertimescales,weusedalinearregressionto

deter-minethetrial-to-trialdriftoverthecourseofbehavioralsession

lastingapproximatelyonehour.Wefoundthatdrifttendedtobe

generallysmall,withaboutthreequartersofthesessionshowing

driftoflessthan1◦/handmanysessionsshowingdriftunder0.5◦/h

(Fig.6E).Forhead-postedmonkeys,drift valuesof upto0.3◦/h

havebeenreported(Kimmeletal.,2012),whichsuggeststhatlong

timescalestabilityofeyetrackinghelmet-restrainedmonkeyscan

reachvaluesobservedintherigidlyhead-postedcondition.There

arehoweveroccasionalsessionswithlargerobserveddrift,that

(10)

willrequiremanualreadjustment ofeyecalibrationparameters

duringthesessiontoachieveoptimalperformance.The

trial-to-trialvariabilityof eye positiondatacorrespondingto thesame

fixationlocationoverthetime-courseofaboutonehourwasabout

afactoroftwolargerthanvaluesreportedforhead-postedmonkeys

(Kimmeletal.,2012),suggestingthatoursystemperformsonly

moderatelyinferiorforthisparameter.

Takentogether,wehavedescribedacompactMATLAB

imple-mentationforbehavioralcontrolandeyetrackingthatisoptimized

fornon-rigidheadrestrainedmacaquemonkeyapplications.Itis

alowcostsolutionusinga120Hzframeratetrackingsystemand

thefreelyavailablePsychophysicstoolbox.Itreliesexclusivelyon

thestreamingmodeofthedataacquisitionboard,whichconfers

substantialbenefits,particularlyintermsoftemporalfidelity,

com-paredtothewidelyusedcombinedstreamingandstoragemode.

Althoughoursystemdoesnotreachperformancecharacteristics

ofinvasiveeyecoilbasedeyetrackersorhighframerateinfrared

trackingsystemswithrigidhead-posting,itneverthelessachieves

aperformancethatissuitableformanyapplications.Wenotethat

thissystemcanofcoursealsobeusedforhumanpsychophysics

studies,particularlyinsituationswhereheadimmobilizationusing

chin-restorbite-barisnotpossibleorundesirable.

References

AsaadWF,EskandarEN.Achievingbehavioralcontrolwithmillisecond resolu-tioninahigh-levelprogrammingenvironment.JNeurosciMethods2008;173: 235–40.

AsaadWF,SanthanamN,McClellanS,FreedmanDJ.High-performanceexecution ofpsychophysicaltaskswithcomplexvisualstimuliinMATLAB.JNeurophysiol 2013;109:249–60.

BrainardDH.Thepsychophysicstoolbox.SpatVis1997;10:433–6.

Judge SJ, Richmond BJ, Chun FC. Implantation of magnetic search coil for measurement of eye position: an improved method. Vis Res 1980;20: 535–8.

KimmelDL,MammoD,NewsomeWT.Trackingtheeyenon-invasively: simulta-neouscomparisonofthescleralsearchcoilandopticaltrackingtechniquesin themacaquemonkey.FrontBehavNeurosci2012;6:49.

Kleiner M, Brainard D, Pelli D. What’s new in Psychtoolbox-3? Perception 2007;36.

MachadoCJ,NelsonEE.Eye-trackingwithnonhumanprimatesisnowmore acces-siblethaneverbefore.AmJPrimatol2011;73:562–9.

MeyerT,ConstantinidisC.Asoftwaresolutionforthecontrolofvisualbehavioral experimentation.JNeurosciMethods2005;142:27–34.

PelliDG.TheVideoToolboxsoftwareforvisualpsychophysics:transforming num-bersintomovies.SpatVis1997;10:437–42.

RobinsonDA.Amethodofmeasuringeyemovementusingscleralsearchcoilina magneticfield.IEEETransBiomedEng1963;10:137–45.

Figure

Fig. 1. Behavioral experiment setup. (A) Side view of the recording box where behavioral data was recorded
Fig. 3. (A) Two cycles of a synthesized 1 Hz sinusoidal wave are shown in the left panel
Fig. 5. Performance of target fixation for monkeys C (A and C) and D (B and D) during blocks of 50 trials
Fig. 6. Sixteen trials are shown (A and B). At time t = 0 target fixation was acquired

Références

Documents relatifs

We propose a simple event identification approach, which uses a sliding window model to extract events and the context of events in real-time from the live public data stream

L’archive ouverte pluridisciplinaire HAL, est destinée au dépôt et à la diffusion de documents scientifiques de niveau recherche, publiés ou non, émanant des

As a previous study in the field of music RS showed that MS influences the way users look to a music RS interface and previous stud- ies in the field of information retrieval

To make a final decision on the state of the data transmission system from the aircraft and the presence or absence of intruder interventions into the data channel, a decision

The Number object of the MWNumericArray class is used to convert formats between the application and the corresponding methods of the CLR – Model class (the

The event finding efficiency and fraction of correctly found events have been studied as a function of time period used for event finding and number of digis required for

Kerstin Schneider received her diploma degree in computer science from the Technical University of Kaiserslautern 1994 and her doctorate (Dr. nat) in computer science from

An expert system analyzes the synchronized data received from two different devices, searching for selected features of the signal.. An additional advantage of this approach is