• Tidak ada hasil yang ditemukan

Yawn analysis with mouth occlusion detection.

N/A
N/A
Protected

Academic year: 2017

Membagikan "Yawn analysis with mouth occlusion detection."

Copied!
10
0
0

Teks penuh

(1)

ContentslistsavailableatScienceDirect

Biomedical

Signal

Processing

and

Control

j ou rn a l h o m e pa g e :w w w . e l s e v i e r . c o m / l o c a t e / b s p c

Yawn

analysis

with

mouth

occlusion

detection

Masrullizam

Mat

Ibrahim,

John

J.

Soraghan,

Lykourgos

Petropoulakis,

Gaetano

Di

Caterina

DepartmentofElectronicandElectricalEngineering,UniversityofStrathclyde,204GeorgeStreet,G11XWGlasgow,UnitedKingdom

a

r

t

i

c

l

e

i

n

f

o

Articlehistory:

Received11August2014 Receivedinrevisedform 10December2014 Accepted9February2015

a

b

s

t

r

a

c

t

Oneofthemostcommonsignsoftirednessorfatigueisyawning.Naturally,identificationoffatigued individualswouldbehelpedifyawningisdetected.Existingtechniquesforyawndetectionarecentredon measuringthemouthopening.Thisapproach,however,mayfailifthemouthisoccludedbythehand,asit isfrequentlythecase.Theworkpresentedinthispaperfocusesonatechniquetodetectyawningwhilst alsoallowingforcasesofocclusion.Formeasuringthemouthopening,anewtechniquewhichapplies adaptivecolourregionisintroduced.Fordetectingyawningwhilstthemouthisoccluded,localbinary pattern(LBP)featuresareusedtoalsoidentifyfacialdistortionsduringyawning.Inthisresearch,the StrathclydeFacialFatigue(SFF)databasewhichcontainsgenuinevideofootageoffatiguedindividualsis usedfortraining,testingandevaluationofthesystem.

©2015ElsevierLtd.Allrightsreserved.

1. Introduction

Fatigueisusuallyperceivedasthefeelingoftirednessor drowsi-ness.Inaworkrelatedcontext,fatigueisastateofmentaland/or physicalexhaustionthatreducestheabilitytoperformworksafely andeffectively.Fatiguecanresultduetoanumberoffactors includ-inginadequaterest,excessivephysicalor mentalactivity,sleep patterndisturbances,orexcessivestress.Therearemany condi-tionswhichcanaffectindividualsandwhichareconsidereddirectly relatedtofatiguesuchasvisualimpairment, reducedhand–eye coordination,lowmotivation,poorconcentration,slowreflexes, sluggishresponse orinabilitytoconcentrate. Fatigueis consid-eredasthelargestcontributor toroadaccidents,leadingtoloss of lives. Data for 2010 from the Department of Transport, UK

[1],points to 1850people killed,and 22,660seriouslyinjured, withfatiguecontributingto20%tothetotalnumberofaccidents

[2]. US National Highway Traffic Safety estimates suggest that approximately100,000accidentseachyeararecausedbyfatigue

[3].

Fatigueisidentifiablefromhumanphysiologysuchaseyeand mouth observations, brainactivity, and by using electrocardio-grams(ECG) measuring,for example,heartratevariability. The physical activities and human behaviour may also be used to identifyfatigue[4–6].Inthis paperyawning, indicatingfatigue, is detectedand analysed from video sequences. Yawning is an

Correspondingauthor.Tel.:+44(0)1415484458.

E-mailaddress:gaetano.di-caterina@strath.ac.uk (G.DiCaterina).

involuntary action where the mouth opens wide and, for this reason,yawningdetectionresearchfocusesonmeasuringand clas-sifyingthis mouth opening.Frequently, however,this approach is thwartedbythecommonhumanreaction tohand-coverthe mouthduringyawning.Inthispaper,weintroduceanewapproach

todetectyawning bycombiningmouth openingmeasurements

witha facialdistortion (wrinkles)detectionmethod.For mouth openingmeasurementsanewadaptivethresholdforsegmenting themouthregionisintroduced.Foryawningdetectionwiththe mouthcovered,localbinarypatterns(LBP)featuresandalearning machineclassifierareemployed.Inordertodetectthewrinkles, theedgedetectorSobeloperator[30]isused.Differentlyfrom[32], wherethemouthcovereddetectiontechniqueisappliedtostatic imagesonly,theyawnanalysisapproachproposedinthispaper describesacompletesystemforyawndetectioninvideosequences. Inthisresearch,genuineyawningfatiguevideodataisusedfor training,testingandevaluation.ThedataisfromtheStrathclyde FacialFatigue(SFF)videodatabase,whichwasexplicitlycreated toaid this research,and which contains series of facial videos of sleep deprived fatigued individuals. The, ethically approved, sleepdeprivationexperimentswereconductedattheUniversity ofStrathclyde,withtheaidoftwentyvolunteersundercontrolled conditions.Eachofthe20volunteerswassleepdeprivedfor peri-odsof0,3,5and8honseparateoccasions.Duringeachsessionthe participants’faceswererecordedwhiletheywerecarryingouta seriesofcognitivetasks.

Theremainderofthispaperisorganisedasfollows.Section2

discussestherelatedworkofthisresearch,whileSection3explains theSFF database.Section4 discusses theoverall system,while

http://dx.doi.org/10.1016/j.bspc.2015.02.006

(2)

of the mouth opening [7–9]. The algorithms used must there-forebeabletodifferentiatebetweennormalmouthopeningand yawning.Yawndetectioncanbegenerallycategorisedintothree approaches;featurebased,appearancebased,andmodelbased.In thefeaturebasedmethod,themouthopeningcanbemeasured byapplying severalfeatures suchascolour, edges andtexture, whichareabletodescribetheactivitiesofthemouth.Commonly, therearetwoapproachestomeasurethemouthopening,mainly bytrackingthelipsmovementandbyquantifyingthewidthof themouth.Sincethemouthisthewiderfacialcomponent,colour isoneof theprominentfeatures usedtodistinguishthemouth region.Yao et al. [10] transfer RGB colour spaceto LaB colour spaceinordertousethe‘a’component,whichisfoundas accept-ableforsegmentingthelips. Thelipsrepresenttheboundaryof themouthopeningthat istobe measured.YCbCr colourspace waschosen by Omidyeganeh [11] for their yawning detection. In that work,the claim is that colour spaceis able toindicate themouthopeningarea, whichis thedarkestcolour region,by setting a certain threshold. Yawning is detected by computing theratioofthehorizontalandverticalintensityprojectionofthe mouthregion.Themouthopeningisdetectedasyawnwhenthis ratiois aboveasetthreshold. Implementationusingthecolour propertiesiseasybut,tomakethealgorithmrobust,several chal-lengesmustbeaddressed.Thelipscolourisdifferentforeveryone, andduetothedifferentlightingconditionsitisnecessarytoensure thatthethresholdvalueisadaptedtothechanges.

Edgesareanothermouthfeatureabletorepresenttheshape ofthemouthopening.Aliouaetal.[12]extractedtheedgesfrom thedifferencesbetweenthelipsregionandthedarkestregionof themouth.Thewidthofthemouthis measuredinconsecutive frames, and yawn is detectedwhen themouth is continuously openingwidelymorethanasetnumberoftimes.Whenusingthe edgestodetermineyawn,thechallengeisthatthechangesofthe edgesareproportionaltotheilluminationchanges,makingit dif-ficulttosettheparameterofanedgedetector.Inordertolocate accuratelythemouthboundary,theactivecontouralgorithmcan beusedasimplementedin[8].However,thecomputationalcost ofthisapproachisveryhigh.Themouthcornershavealsobeen employedbyXiaoetal.[9]todetectyawn.Themouthcornerscan

bereferencepointsinordertotrackthemouthregion.Yawning isclassifiedbasedonthechangeofthetextureinthecornerof themouthwhilethemouthisopenedwidely.Moreover,asinthe caseimplementedformeasuringeyeactivities,Jiménez-Pintoetal.

[13]havedetectedtheyawnbaseduponthesalientpointsinthe mouthregion.ThesesalientpointsintroducedbyShiandTomasi

[14]areabletodescribethemotionofthelipsmovement,andyawn isdeterminedbyexaminingthemotioninthemouthregion.In

[7,11]yawnisdetectedbasedonahorizontalprofileprojection. Theheightofthemouthopeninginprofileprojectionrepresents themouthyawn.However,thisapproachisinefficientforbeard andmoustachebearingindividuals.

Inappearancebasedmethod,statisticallearningmachine algo-rithmsareappliedanddistinctivefeaturesneedtobeextractedin ordertotrainthealgorithms.Lirongetal.[15]combinedHaar-like featuresandvariancevalueinordertotraintheirsystemtoclassify thelipsusingsupportvectormachine(SVM).Withtheadditional integrationofaKalmanfiltertheiralgorithmisalsoabletotrack thelipregion.Linglingetal.[16]alsoappliedHaar-likefeaturesto

yawnclassificationinthreestates:mouthclosed,normalopen,and wideopen.Themouthregionfeaturesareextractedbasedonthe RGBcolourproperties.Forthisapproachahugenumberofdatasets forimagesisrequiredinordertoobtainadequateresults.

Themodelbasedapproachrequiresthatthemouthorlipsare modelledfirst.Anactiveshapemodel(ASM)forinstance,requires asetoftrainingimagesand,ineveryimage,theimportantpoints thatrepresentthestructureofthemouthneedtobemarked.Then, themarkedsetofimagesisusedfortraining usingastatistical algorithmtoestablishtheshapemodel.

García et al.[18] and Anumaset al. [19] train thestructure usingASM, and theyawn ismeasuredbased onthelips open-ing.Hachisukaetal.[20]employactiveappearancemodel(AAM),

whichannotateseachessentialfacialpoint.Foryawningdetection threepointsaredefinedtorepresenttheopeningofthemouth.In thisapproachalotofimagesarerequiredsinceeverypersonhas differentfacialstructure.

3. StrathclydeFacialFatigue(SFF)database

TheStrathclydeFacialFatiguedatabasewasdevelopedinorder toobtaingenuinefacialsignsoffatigue.Mostofthefatiguerelated researchershaveusedtheirowndatasetsforassessingtheir devel-opedsystem.Forexample,Vuraletal.[21]developedtheirfatigue databasebyusingvideofootagefromasimulateddriving environ-ment.Theparticipantshadtodriveforover3handtheirfaceswere recordedduringthattime.Fanetal.[22]recruitedforty partici-pantsandrecordedtheirfacesforseveralhoursinordertoobtain facialsignsoffatigue.Someresearchersonlytesttheiralgorithms based on video footage of persons who pretend to experience fatiguesymptoms.Forexample,researchersin[8,11,12]detected yawnbasedonthewidthofthemouthopeningandthealgorithms werenottestedonsubjectsexperiencingnaturalyawning.TheSFF databaseprovidesagenuinemediumoffacialfatiguevideofootage fromsleepdeprivationexperimentswhichinvolvedtwenty partic-ipants.

Thesleepdeprivedvolunteerswere10maleand10femalewith agesrangingbetween20to40yearsold.Eachvolunteerhadto per-formaseriesofcognitivetasksduringfourseparateexperimental sessions.Priortoeachsession,theparticipantsweresleepdeprived for0h(nosleepdeprivation),3,5and8hundercontrolled condi-tions.Thecognitivetasksweredesignedtotest(a)simpleattention, (b)sustainedattentionand(c)theworkingmemoryofthe partici-pants.Thesecognitivetasksareassociatedwithsensitivitytosleep lossandaredesignedtoacceleratefatiguesigns.Fig.1showssome examplesofvideofootageintheSFFdatabase.

4. Systemoverview

Ablockdiagramofthenewyawnanalysissystemisshownin

Fig.2.Thesystembeginswithaninitialisationoperationcarriedout inthefirstframeofthevideo.Thisconsistsoffaceacquisitionand regioninitialisationalgorithms.Duringfaceacquisition,theface, eyeandmouthdetectionalgorithmsareperformedsequentially beforearegionofinterestinitialisationalgorithmisexecuted.

(3)

M.M.Ibrahimetal./BiomedicalSignalProcessingandControl18(2015)360–369

Fig.1. ExampleimagesfromvideofootageSFF.

Fa

Eye

Mouth Detection Face Detection ace acquisition

Regions Interest Initialisation

Yawn analysis operations Detection

Fig.2.Yawnanalysissystemoverview.

tobetrackedinallthefollowingframes.TheFMRregionisused inmeasuringthemouthopeningandtodetectwhenthemouth iscovered,andFDRisusedinmeasuringthedistortionsduring yawning.

4.1. Regionofinterestevaluation

Asis indicatedin[32] afacial regionof interestneedstobe establishedfirst.Fig.3depictsafaceacquisitionoperationwhere theeyesandmoutharesequentiallydetectedusingaViolaJones technique[23,24].Trainingisperformedthroughacascade clas-sifierappliedtoeverysub-window intheinputimagein order toincreasedetectionperformanceandreducecomputationtime. TheSFF video footageprovided thelargenumber of face, eyes and mouthdata neededtotrain the Haar-likefeatures [25,26]. Inthepresentmethod,thefaceregionisfirstdetectedandthen, theeyesandmouthareasareidentifiedwithinit.Theacquisition ofthedistinctive distances betweenthesefacialcomponents is effectedthroughanthropometricmeasurements,thusaccounting forhumanfacialvariability.Thisinformationisthenusedinthe formationofthetwofollowingregionsofinterest.

4.1.1. Focusedmouthregion(FMR)

Thisisformedbaseduponthedetectedlocationsoftheeyesand mouth.Thedistancebetweenthecentreoftheeyes(ED),andthe distancebetweenthecentreofmouthandthemid-pointbetween theeyes(EMD)areobtained(Fig.3(a)).Theratioofthesedistances

REMisthencomputedas:

REM= ED

EMD (1)

Fig.3. (a)Anthropometricmeasurement;(b)focusedmouthregion(FMR);(c) focuseddistortionregion(FDR).

FromthisinformationthecoordinatesoftheFMR(Fig.3(b))are empiricallydefinedas:

x1

y1

=

xR

yR+0.75EMD

(2)

x2

y2

=

xL

yL+0.75EMD+0.8ED

(3)

where(xR,yR)and(xL,yL)arethecentrepointsoftherighteyeiris andlefteyeiris,respectively.TheFMRdependsonthelocationand distanceoftheeyes.Whenthefacemovesforwardtheeye dis-tanceincreasesandsodoestheFMR.Conversely,theFMRreduces followingabackwardmove.

4.1.2. Focuseddistortionregion(FDR)

Thefocuseddistortionregion(FDR)isaregioninthefacewhere changesoffacialdistortionoccur.Theregion,whichismostlikely toundergodistortionsduringyawning, wasidentifiedbased on conductedexperimentsusingtheSFFdatabase.Thecoordinatesof theFDR(Fig.3(c)),areempiricallydefinedas:

x1

y1

=

xR

yR−0.75EMD

(4)

x2

y2

=

xL

yL+0.75EMD

(5)

Thethresholdvalues0.75and0.8usedinEqs.(2)–(5)arebased onobservationtests,undertakenontheSFFdatabase.

4.2. Yawnanalysisoperations

(4)

Fig.4. AlineartransformthatremapsintensitylevelinputimageinbetweenG′ min andG′

max.

0 0 10 20 30 40

(a)

50 100 150 200 250 00 50

1000 2000 3000

(b)

100 150 200 250

Fig.5.(a)Enhancedimagehistogram;(b)cumulativehistogram.

openingmeasurement, coveredmouthdetection,and distortion detection.

4.2.1. Mouthopeningmeasurement

Yawningisaninvoluntaryactionthatcausesapersontoopen themouth widely and breathe in. The technique in this paper detectsyawningbaseduponthewidestareaofthedarkestregion betweenthelips.Aspreviouslydiscussedwhenappliedtechniques arebasedonintensitiesandcolourvalues,thechallengeistoobtain anaccuratethresholdvaluewhichwilladapttomulti-skinandlips colouration,aswellastodifferentlightingconditions.Theproposed algorithmappliesanewadaptivethresholdinordertomeetthis challenge.

Theadaptivethresholdvalueiscalculatedduringthe initialisa-tionoperationaftertheFMRisformed.Duetothedependenceof themethodontheintensityvalueoftheregion,itisnecessaryto improvethefaceimageinordertocontrastenhancethedarkest andthebrightestfacialregions.Theenhancementisimplemented bytransformingthehistogramoftheimagetospreadthelevelof colourvaluesevenlybutwithoutchangingtheshapeoftheinput histogram.Thisiscomputedusingthefollowing transformation

[27]:

g′(x,y)=G ′

max−G′

min

Gmax−Gmin[g(x,y)−Gmin]+G

min (6)

ThelineartransformationillustratedinFig.4stretchesthe his-togramoftheinputimageintothedesiredoutputintensitylevel ofthehistogrambetweenG′

minandG′max,whereGminandGmaxare

theminimumandmaximuminputimageintensities,respectively. FromthisenhancedhistogramillustratedinFig.5(a),the cumu-lativehistogramHiiscalculatedasinEq.(7),thusgeneratingthe

Fig.6. Focusedmouthregion(FMR).

Fig.7.FMRsegmentationusingadaptivethreshold.

totalnumberofindexpixelsinthehistogrambinsbetween0and 255asillustratedinFig.5(b).

Hi= i

j=1

hj (7)

Consequently,theinitialadaptivethresholdiscalculatedby ref-erencetothedarkestareaasshowninFig.6.Onaverage,thedarkest areaconstitutesapproximately5%,orless,oftheentireobtained regionofinterest.Basedonthispercentage,thedarkestareainthe histogramissituatedbetween0andXintensitylevels.The inten-sityvalueXcanbeobtainedbyselectingthefirst5%ofthetotal numberofpixelscontainedintheenhancedhistogram(Fig.5(a)) andreadingthecorrespondingright-mostintensityvaluefromthe

x-axis.

Then,theadaptivethresholdTavalueiscalculatedasfollows:

Ta=

X−Pmin

Pmax−Pmin (8)

wherePmaxandPminarethemaximumandminimumpixelvalues,

respectively.

AftertheadaptivethresholdTa iscomputed,theFMRis seg-mentedandthemouthlengthLMisobtainedasillustratedinFig.7. ThelengthofLMmustbewithintherangeindicatedinthefollowing equation:

0.7ED<LM<0.9ED (9)

IfLMisoutsidethisrange,thevalueTaisincreasedordecreased accordingly,instepsof0.001.Thenthesegmentationprocessis repeateduntilthevalueofLMiswithinrangeasinEq.(9).

Theadaptivethresholdisappliedtothenextsequenceofframes forsegmentingthedarkestregionbetweenthelips.In orderto determinetheyawning,theheightofthemouthHM,basedonthe darkestregioninFMR,ismeasuredasshowninFig.8.Then,the YawnRatio(RY)oftheheightHMovertheheightofFMRiscomputed asfollows:

RY =

HM

FMRheight (10)

(5)

M.M.Ibrahimetal./BiomedicalSignalProcessingandControl18(2015)360–369

Fig.8. Themeasurementofhighofmouthopening.

Fig.9.(a)Uncoveredmouth;(b)coveredmouth.

Fig.10.ExampleofthebasicLBPoperator.

4.2.2. Coveredmouthdetection

Itisnotunusualtooccludethemouthwiththehandduring yawning.ThismeansthatinformationfromSection4.2.1canno longerbeusedtodetectyawning.Theapproachfirstintroducedin

[32]onstillimagescanalsobeusedinvideosequencefootage.The approachinvolvesthattheFMRisexaminedtodeterminewhether theregioniscoveredornot.Fig.9displaysimagesofFMR(a)not coveredand(b)covered.Fromtheseimagesthedifferenttexture oftheFMRregionsisclearlyobviousinthetwocases.Inorderto differentiatebetweenthesetwocases,localbinarypatterns(LBP) areused toextractthe FMRtexturepattern. LBPfeatures have beenprovenhighlyaccuratedescriptorsoftexture,arerobustto themonotonicgrayscalechanges,aswellassimpletoimplement computationally[28].

Localbinarypatterns(LBP)wasoriginallyappliedontexture analysistoindicatethediscriminativefeaturesoftexture.With ref-erencetoFig.10,thebasicLBPcodeofthecentralpixelisobtained bysubtractingitsvaluefromeachofitsneighboursand,ifanegative numberisobtained,itissubstitutedwitha0,elseitisa1.

TheLBPoperatorislimitedsinceitrepresentsasmallscale fea-turestructurewhichmaybeunabletocapturelargescaledominant features.Inordertodealwiththedifferentscaleoftexture,Ojalaet al.[29]presentedanextendedLBPoperatorwheretheradiusand samplingpointsareincreased.Fig.11showstheextendedoperator

Fig.11.ExampleofextendedLBPoperator.(a)(8,1);(b)(16,2)and(c)(24,3)

neigh-bourhoods.

wherethenotation(P,R)representsaneighbourhoodofPsampling pointsonacircleofradiusR.

TheLBPresultcanbeformulatedindecimalformas:

LBPP,R(xc,yc)=

P−1

P=0

s(iP−ic) 2P (11)

whereic andiPdenotegraylevelvaluesofthecentralpixeland

Prepresentsurroundingpixelsinthecircleneighbourhoodwith radiusR.Functions(x)isdefinedas:

s(x)=

1

, if x≥ 0

0, if x<0 (12)

TheLBPoperator,as formulatedinEq. (11),hasa rotational effect.Hence,iftheimageisrotated,then,thesurroundingpixelsin eachneighbourhoodaremovingaccordinglyalongtheperimeter ofthecircle.InordertoremovethiseffectOjalaetal.[29]proposed arotationalinvariant(ri)LBPasfollows:

LBPri P,R=

minROR

LBP

P,R,i

i=0, 1,..., P−1

(13)

whereROR(x,i)performsacircularbitwiserightshiftontheP-bit numberxwithitime.ALBPuniformpattern(u2),LBPu(P,R2 )wasalso

proposed.TheLBPiscalleduniformwhenitcontainsamaximum oftwobitwisetransitionsfrom0to1orviceversa.Forexample, 1111111(0transition)and00111100(2transitions)areboth uni-form,but100110001(4transitions)and01010011(6transitions) arenotuniform.AnotherLBPoperator,whichisacombinationof arotationalinvariantpatternwithauniformpattern,isLBPriu2

(P,R).

Thisoperator isobtainedbysimplycountingthe1’sinuniform patterncodes.Allnon-uniformpatternsareplacedinasinglebin. AnexampleofanLBPhistogramofthethreetypesofLBPoperators forcoveredanduncoveredmouthregionisshowninFig.12.

AclassifierthenusestheLBPfeaturestodetectmouthcovering, asexplainedinSection6.2.

4.2.3. Distortionsdetection

Yawning cannot be concluded solely on the detection of a coveredmouthregion.Itispossiblethatthemouthiscovered acci-dentallyorevenintentionallyforreasonsunrelatedtoyawning.To minimisethepossibilityoffalseyawndetection,themetricof dis-tortionsinaspecificregionoftheface,mostlikelytodistortwhile yawning,isalsoused[32].Thisregionisidentifiedbasedon obser-vationsandtheexperimentsconductedusingthevideo footage fromtheSFFdatabase.

Theregionmostaffectedbydistortionduringyawningisthe focuseddistortionregion(FDR)(Section4.1.2).FromFig.13bandc itisobviousthattherearesignificantvisualdifferencesintheFDR areabetweennormalconditionsandduringyawning.

ToevaluatethechangesoccurringinFDR,edgedetectionisused toidentifythedistortions.Hereanyedge detectorcanbeused. However,basedonexperimentscarriedout,theSobeloperator[30]

(6)

Fig.12.LBPhistogramfornotcoveredmouthandcoveredmouthregions.

Fig.13.(a)Focuseddistortionregion(FDR);(b)normalconditioninFDR;(c) yawn-inginFDR.

intensitygradientateachpointintheregionofinterest.Then,it providesthedirectionofthelargestpossibleincreasefromlightto darkandtherateofchangeinthehorizontalandverticaldirections. TheSobeloperatorrepresentsapartialderivativeoff(x,y)andof thecentralpointofa3×3areaofpixels.Thegradientsforthe hor-izontalGxandverticalGydirectionsfortheregionofinterestare thencomputedas[31]:

Gx=

f(x+1, y−1)+2f(x+1, y)+f(x+1, y+1)

f(x−1, y−1)+2f(x−1, y)+f(x−1, y−1)

(14)

Gy=

f(x−1, y+1)+2f(x, y+1)+f(x+1, y+1)

f(x−1, y−1)+2f(x, y−1)+f(x+1, y−1)

(15)

TheFDRdistortionsaredetectedbasedonthegradientsinboth directionsandtheresultsareshowninFig.14.Thenumberofedges detectedundernormalconditions,asillustratedinFig.14(a),is sig-nificantlydifferentcomparedtoyawning(Fig.14(b)).Inorderto identifythedistortionintheFDRduringyawning,thesumFDRSAD

oftheabsolutevaluesobtainedinEq.(16)isusedtocomputethe numberofwrinklesintheregion.ThenormalisedFDRSADis calcu-latedinEq.(17)whereWandHdenotethewidthandheightofthe FDR,respectively.

Fig.14. FDRwithinputimageandedgesdetectedimage.(a)Normal;(b)yawn.

Fig.15.PlotofnormalisedFDRSADvaluesduringyawning.

Duringyawning,thenumbersofthedetectededgesincreaseas showninFig.15.

FDRSAD =

i,j

I1(i,j)−I2(x+i,y+j) (16)

Normalised FDRSAD= FDRSAD

255×W×H (17)

5. Yawnanalysis

Ingeneral,theactionofyawningcanbecategorisedintothree situations:(a)mouthnot-coveredandwideopen,(b)mouthwide open for a short time before it gets covered and, (c) mouth totallycoveredwithinvisiblemouthopening.Thesesituationsare based ontheobservations madefrom thevideo footageof the SFFdatabase.Ayawnanalysisalgorithm, asdepictedinFig.16, wasthereforedevelopedtodealwiththesecasesbyintroducing theaforementionedprocedures,i.e.mouthopeningmeasurement, mouthcovereddetectionanddistortionsdetection.

Inthisalgorithm,thestatusofmouthopening,mouthcovered, anddistortionsareexaminedperiodically.Thechosentimeframe formeasurements,basedonobservationsfromtheSFFdatabase, is30s—atimeintervalreachedbyexaminingtheSFFvideodata. Thisassumesthatyawningdoesnotoccurtwicewithinthistime period.

The algorithm starts byfirst examining themouth opening. Yawningisidentifiedthroughthemouthopeningforlongerthana fixedinterval,usuallybetween3and5s.Ifthisshowsnoyawning, thenthemouthcovereddetectionistriggered.Thestatusof distor-tionsinFDRisonlyexaminedwhenmouthcoveringisdetected.

(7)

M.M.Ibrahimetal./BiomedicalSignalProcessingandControl18(2015)360–369

Cov

mo

dete

M cov

Disto

dete

Dist

YA

Yes No

Yes No

vered

outh

ection

outh vered

ortion

ection

tortion

AWN

No Yes

Mouth

Openin

g

(mo)

mo

>

YAWN

NOT YAWN

NOT YAWN

Fig.16.Yawnanalysisalgorithm.

6. Experimentalresults

Theperformanceof thedevelopedalgorithms wastested by usingthefullSFFvideofootage.Theresultspresentedherewerethe outcomeofsubstantialreal-timeexperimentationusingthevideo sequencesfromthewholeofthedatabase.

6.1. Mouthopeningmeasurement

Themouthopeningmeasurementalgorithmsaretestedusing theSFFdatabase.Inordertoevaluatetheaccuracyofthemouth openingsegmentationusingtheproposedadaptivethreshold,the actualinput image is compared withthe segmented region of mouthopening.Fig.17showsexamplesofmouthopeningimages pairedwiththeoutputimageofmouthopeningsegmentation.A totalof145mouthopeningimagesundervaryinglighting condi-tionswereusedfortesting.Fromthesetests,itwasshownthatthe algorithmwasabletosegmentthedarkestregionbetweenthelips asrequiredfordetectingyawning.

AsdiscussedinSection4,yawningisdeterminedbymeasuring theheightofthesegmentedregionofthemouthopening.

Inordertoobtaintheapplicablethresholdvaluethatrepresents yawning,25oftherealyawningscenesfromtheSFFdatabasewere used.Fromtheyawningscenesitwasfoundthattheminimum valueofRYwhichdenotesyawningis0.5.Thisratiovalueisapplied inmouthopeningmeasurementsduringyawnanalysis.

Fig.17.Mouthopeningsegmentationimages.

6.2. Mouthcovereddetection

Inordertodetectthecoveredmouth,theFMRisexaminedin everyvideoframe.TheLBPfeaturesareextractedfromtheregion, andthelearningmachinealgorithmdecidesthestatus.The clas-sifierstested,todeterminetheirsuitabilityforuseintheanalysis algorithm,comprisedsupportvectormachines(SVM)andneural networks(NN).

A total of 500covered mouth images and 100non covered mouthimagesofsize50×55pixelhavebeenusedtotraintheNN andSVMclassifiers.TheLBPoperatorsusedforfeatureextraction were:rotationalinvariant(ri),uniform(u2),anduniformrotational

invariant(riu2).Theradiioftheregionsexaminedwereeither8,16, or24pixel.

Classificationwasperformedon300images.Fourdifferent ker-nelfunctionsweretestedfortheSVMclassifierandtwodifferent networklayersfortheNN.Analyticallythesewere:SVMwitha) radialbasisfunction(SVM-rbf),(b)linear(SVM-lin),(c)polynomial (SVM-pol),and (d)quadratic (SVM-quad). The neuralnetworks testedhad5(NN-5)and10neurons(NN-10),respectively,intheir hiddenlayer.

Itis usualtopresent suchresultsusinggraphsofsensitivity (ordinate)against1-specificity (abscissa),where sensitivity and specificityaregiven,respectively,by:

Sensitivity= TP TP+FP

(18)

Specificity= TN TN+FN

(19)

whereTprepresentsthenumberoftruepositives,Fpthenumberof falsepositives,Tnthenumberoftruenegatives,andFnthenumber offalsepositives.Sensitivityrepresentstheabilityofthedeveloped algorithmtoidentifythecoveredmouth.Thespecificityrelatesto theabilityofthealgorithmtodetectthenon-coveredmouth.For maximumaccuracyandminimumnumberoffalsepositives, max-imumvaluesofsensitivityandspecificityarerequired.TheROC curvesfromtheresultsoftheclassificationsareshowninFig.18.

ThelegendsshowninFig.18(a–c)indicatetheclassifiersused ineachcaseandtheirrespectiveoutputs.

(8)

Fig.18.ROCcurvesforLBPoperator:(a)rotationalinvariant(ri);(b)rotational invariantwithuniformpattern(riu2);(c)uniformpattern(u2)withzoom-ininsert ofupperleftcorner.

6.3. Yawnanalysis

Yawnanalysisisacombinationofmouthopeningmeasurement, mouthcovereddetection,anddistortionsdetection.

Theseoperationsareexaminedevery30s.Formouthopening measurements,thebestthresholdvaluethatrepresentsyawningis

Fig.19.Plotsofyawningdetectionwithmouthnotoccluded.

0.5.BasedontheresultsobtainedandshowninFig.19,formouth coveredclassificationtheLBPuniformpattern(u2)withradius16 andaneuralnetworkclassifierwith5neuronsinthehiddenlayer arechosentodeterminethestatusofFMR.

Genuine yawning images from sequences of 30 video clips fromtheSFFdatabaseareusedforevaluationofthealgorithm’s performance. Thesevideos contain10 yawning sequences with non-coveredmouthscenes,14 withcovered mouthscenesand 6sceneswithbothsituationspresent.Examplesofthedetection resultsineachsituationofyawningareshowninFigs.19–20.

Whenthemouthisnotoccluded(Fig.19),yawningisdetected whentheheightofthemouthopeningisequaltoorgreaterthanthe thresholdvalue.Theyawningperiodismeasuredfromthe begin-ningoftheintersectionpoint betweenthethresholdvalueand theratioofthemouthopeningheighttoFMRheightRY.When theRY valueis belowthethreshold(Fig.20), thiscouldbedue tomouthbeingcoveredduringyawning.Inthiscase,themouth covereddetectionpartofthesystemistriggered.Ifmouth occlu-sionisdetected,facialdistortionduringtheperiodofocclusionis checked.Yawningisassumedtooccurwhendistortionincreases substantiallywithinthatperiod.

Whenthemouthisopenedwideoverashortperiodoftime beforeitisoccluded(Fig.21),thevalueofRYexceedsthe thresh-oldvaluefora shortperiodbeforeitdropsquickly.Inthiscase, themouthocclusiondetectionwillbetriggeredwhentheRYvalue dropsbelowthethreshold.Ifocclusionisdetected,thentheperiod ofyawningismeasuredfromthebeginningofthefirstintersection ofRYwiththethresholdvalue,untiltheendofthemouthocclusion detection.

(9)

M.M.Ibrahimetal./BiomedicalSignalProcessingandControl18(2015)360–369

Fig.20.Plotsofyawningdetectionwithmouthoccluded.

Fig.21.Plotsofyawningdetectionwithmouthinitiallynotoccludedbutoccluded subsequently.

7. Conclusion

Ayawnanalysisalgorithmdesignedtodealwithbothoccluded and uncovered mouth yawning situations was presented. This algorithmisacombinationoftechniquesformouthopening mea-surement, mouth covered detection and distortions detection.

The mouth opening is measured based on the darkest region

betweenthelips.Thedarkestregionissegmentedusinganadaptive thresholdwhichisintroducedinthispaper.Fordetectingmouth occlusion,anLBPuniformoperatorisappliedtoextractfeaturesof themouthregion,andtheocclusionisevaluatedusinganeural net-workclassifier.Toensurethatthemouthocclusionisindeeddue toyawningandnootherreason,thewrinklesofspecificregionson thefacearemeasured.Yawningisassumedtohaveoccurredwhen distortionsintheseregionsaredetected.

Inthepresentedyawnanalysisalgorithm,thetworegionsFMR andFDRaremonitoredevery30s. Theanalysisresultsproduce thestatusofyawning,aswellastheperiodofyawning.Genuine sequencesofyawningfromtheSFFdatabasevideofootageareused fortraining,testingandevaluatingtheperformanceofthe devel-opedalgorithms.Basedontheconductedexperimentsandresults obtained,theoveralldevelopedsystemdevelopedindicatedavery encouragingperformance.

Acknowledgements

ThisworkwassupportedbytheScottishSensorSystemsCentre (SSSC),theMinistryforHigherEducationofMalaysiaandUniversiti TeknikalMalaysia,Melaka(UTeM).TheSFFdatabasedevelopment wasfundedbyaBridgingTheGap(BTG)grantfromtheUniversity ofStrathclyde.

Theworkwhich ledtothedevelopmentoftheSFF database wasincollaborationwiththeschoolofPsychologicalandHealth Sciences,UniversityofStrathclyde,theGlasgowSleepCentre, Uni-versityofGlasgow,BrooksBellLtd.

References

[1]D. f. T. UK, “Reported Road Casualties in Great Britain: 2010 Annual Report”DepartmentforTransportUK,UnitedKingdom,StatisticalRelease-29 September2011.

[2]P. Jackson, C. Hilditch, A. Holmes, N. Reed, N. Merat, L. Smith, Fatigue and Road Safety: A Critical Analysis of Recent Evidence, Department for Transport, London, 2011, pp. p88.

[3]R. Resendes, K.H. martin, Saving Lives Through Advanced Safety Technology, Faderal Highway Administration, US, Washington, DC, 2003.

[4]Daimler, Drowsiness-Detection System Warns Drivers to Pre-vent Them Falling Asleep Momentarily, Daimler, 2012, Available: http://www.daimler.com/dccom/0-5-1210218-1-1210332-1-0-0-1210228-0-0-135-0-0-0-0-0-0-0-0.html.

[5]M.MichaelLittner,C.A.Kushida,D.DennisBailey,R.B.Berry,D.G.Davila,M. Hirshkowitz,Practiceparametersfortheroleofactigraphyinthestudyofsleep andcircadianrhythms:anupdatefor2002,Sleep26(2003)337.

[6]N.Mabbott,ARRBPro-activefatiguemanagementsystem,RoadTrans.Res.12 (2003)91–95.

[7]Y.Lu,Z.Wang,Detectingdriveryawninginsuccessiveimages,in: Bioinfor-maticsandBiomedicalEngineering,2007.ICBBE2007.The1stInternational Conferenceon,2007,pp.581–583.

[8]M.Mohanty,A.Mishra,A.Routray,Anon-rigidmotionestimationalgorithmfor yawndetectioninhumandrivers,Int.J.Comput.VisionRob.1(2009)89–109.

[9]F. Xiao, Y. Bao-Cai, S. Yan-Feng, Yawning detection for monitoring driver fatigue, in: Machine Learning and Cybernetics, 2007 International Conference on, 2007, pp. 664–668.

[10]W. Yao, Y. Liang, M. Du, A real-time lip localization and tacking for lip reading, in: Advanced Computer Theory and Engineering (ICACTE), 2010 Third Interna-tional Conference on, 2010, V6-363-V6-366.

[11]M. Omidyeganeh, A. Javadtalab, S. Shirmohammadi, Intelligent driver drowsi-ness detection through fusion of yawning and eye closure, in: Virtual EnvironmentsHuman–ComputerInterfacesandMeasurementSystems (VEC-IMS),2011IEEEInternationalConferenceon,2011,pp.1–6.

(10)

ITSC’09. 12th International IEEE Conference on, 2009, pp. 1–6.

[17]Y. Yang, J. Sheng, W. Zhou, The monitoring method of driver’s fatigue based on neural network, in: Mechatronics and Automation, 2007. ICMA 2007. Interna-tional Conference on, 2007, pp. 3555–3559.

[18]H. García, A. Salazar, D. Alvarez, Á. Orozco, Driving fatigue detection using active shape models, Adv. Visual Comput. 6455 (2010) 171–180.

[19]S. Anumas, Kim Soo-chan, Driver fatigue monitoring system using video face images & physiological information, in: Biomedical Engineering International Conference (BMEiCON), 2011, 2012, pp. 125–130.

[20]S.Hachisuka,K.Ishida,T.Enya,M.Kamijo,Facialexpressionmeasurementfor detectingdriverdrowsiness,Eng.Psychol.Cogn.Ergon.6781(2011)135–144.

[21]E.Vural,M.Cetin,A.Ercil,G.Littlewort,M.Bartlett,J.Movellan,Drowsydriver detectionthroughfacialmovementanalysis,in:Human–ComputerInteraction, 2007,pp.6–18.

[22]X.Fan,B.-C.Yin,Y.-F.Sun,Dynamichumanfatiguedetectionusing feature-levelfusion,in:Springer(Ed.),ImageandSignalProcessing,Springer,2008,pp. 94–102.

[28]G. Zhao, M. Pietikainen, Dynamic texture recognition using local binary pat-terns with an application to facial expressions, in: Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2007, pp. 915–928, 29.

[29]T. Ojala, M. Pietikainen, T. Maenpaa, Multiresolution gray-scale and rota-tion invariant texture classification with local binary patterns, in: Pattern Analysis and Machine Intelligence, IEEE Transactions on, 2002, pp. 971–987, 24.

[30]R.O. Duda, P.E. Hart, Pattern Classification and Scene Analysis, vol. 3, Wiley, New York, NY, 1973.

[31]J.-Y. Zhang, Y. Chen, X.-x. Huang, Edge detection of images based on improvedSobel operatorand geneticalgorithms, in:ImageAnalysisand Signal Processing, 2009. IASP 2009. International Conference on, 2009, pp.31–35.

Gambar

Fig. 3. (a) Anthropometric measurement; (b) focused mouth region (FMR); (c)focused distortion region (FDR).
Fig.  5(b).
Fig. 9. (a) Uncovered mouth; (b) covered mouth.
Fig. 12. LBP histogram for not covered mouth and covered mouth regions.
+4

Referensi

Dokumen terkait

KEPALA BAGIAN PENGELOLA DATA ELEKTRONIK SEKRETARIAT DAERAH KABUPATEN MALANG.. NOMOR : 180 / / KEP / 421.024 / 2014

Kegiatan pelaksanaan PPL dilakukan dengan keterlibatan mahasiswa sebagai praktikan untuk mengajar. Praktikan mendapatkan tugas sesuai dengan bidang masing-masing

In its efforts to identify what science is required to develop microbial forensics further, the committee used a generous deinition of “science,” including research to

Proses Pembelajaran Pada Implementasi Kurikulum 2013 Tematik Terpadu di SD ………..………. Implementasi Penilaian Autentik Dalam Kegiatan Pembelajaran di SD …..……… 199

RUU KIP mengharuskan semua badan publik untuk mengumumkan informasi publik secara berkala, paling sedikit enam bulan sekali; mengumumkan secara serta-merta suatu informasi yang

KEPOLISIAN NEGARA REPUBLIK INDONESIA DAERAH NUSA TENGGARA BARAT1. BIRO SARANA DAN PRASARANA Jalan Langko

PEKERJAAN REHABILITASI GEDUNG AULA PADA PSBL DHARMA GUNA BENGKULU TAHUN ANGGARAN 2015.. Nomor :

dan Current rasio (rasio lancar) Merupakan Rasio yang digunakan untuk mengukur kemampuan perusahaan dalam membayar kewajiban jangka pendeknya dengan menggunakan aktiva