Driver Drowsiness Detection Using Artificial Neura

CISRGW-TRICenterforIntelligentSystemsResearchGWTransportationResearchInstituteTheGeorgeWashingtonUniversity,VirginiaCampus,20101AcademicWay,Ashburn,VA20147NDIA3rdAnnualIntelligentVehicleSystemsSymposiumDrivingSimulatorExperiment:DetectingDriverFatiguebyMonitoringEyeandSteeringActivityDr.AzimEskandarian,RiazSayed(GWU)CISRGW-TRIResearchObjectiveConductSimulatorExperimentandAnalyzetheData,tosearchforasystemforautomaticdetectionofdrowsinessbasedondriver’sperformanceCISRGW-TRISignificanceoftheProblem•Drowsiness/FatigueRelatedAccidentData:•NHTSAEstimates100,000drowsiness/fatiguerelatedCrashesAnnually•FARSindicatesanannualaverageof1,544fatalities•Fatiguehasbeenestimatedtobeinvolvedin10-40%ofcrashesonhighways(ruralInterstate)•15%ofsinglevehiclefataltruckcrashes•FatigueisthemostfrequentcontributortocrashesinwhichatruckdriverwasfatallyinjuredCISRGW-TRI•Adrowsy/sleepydriverisunabletodeterminewhenhe/shewillhaveanuncontrolledsleeponset•Fallasleepcrashesareveryseriousintermsofinjuryseverity•Anaccidentinvolvingdriverdrowsinesshasahighfatalityratebecausetheperception,recognition,andvehiclecontrolabilitiesreducessharplywhilefallingasleep•Driverdrowsinessdetectiontechnologiescanreducetheriskofacatastrophicaccidentbywarningthedriverofhis/herdrowsinessSignificanceoftheProblemCISRGW-TRIDriverDrowsinessDetectionTechniques1.Sensingofdriverphysicalandphysiologicalphenomenon–AnalyzingchangesinbrainwaveorEEG–AnalyzingchangesineyeactivityandFacialexpressions•Gooddetectionaccuracyisachievedbythesetechniques•Disadvantages:–Electrodeshavetobeattachedtothebodyofthedriverforsensingthesignals–Non-contacttypesensingisalsohighlydependantonenvironmentalconditionsCISRGW-TRI2.Analyzingchangesinperformanceoutputofthevehiclehardware–Steering,speed,acceleration,lateralposition,andbrakingetc.•Advantages:–Nowires,cameras,monitorsorotherdevicesaretobeattachedoraimedatthedriver–Duetothenon-obtrusivenatureofthesemethodstheyaremorepracticallyapplicableDriverDrowsinessDetectionTechniquesCISRGW-TRIApproachforDrowsinessDetectionandDriverWarningCISRGW-TRIExperiment•ConductedintheVehicleSimulatorLaboftheCISR.GWUVACampus,AshburnVA.•Twelvesubjectsbetweentheagesof23and43•TestScenarioconsistedofacontinuousruralInterstatehighway,withtrafficinbothdirectionsSpeedlimitof55mph.•Morningsession8–10am•Nightsession1–3amCISRGW-TRICISRDrivingSimulatorCISRGW-TRIEyeTrackingEquipmentCISRGW-TRISampleDataFromSimulatorRUN#ZONETIMESPEEDLIMCRASHBCRASHVLANEXBRAKEFORBRAKETAP10350000012.1350000014.2350000016.2350000018.33500000STEERPOSSTEERVARLATPLACELATPLVARSPEEDSPEEDVARSPEEDDEV-0.10-0.09053.710-4.650.20-0.22053.710-4.650.40-0.31053.710-4.6500-0.35053.710-4.65CISRGW-TRILateralPositionofVehicleCISRGW-TRIPowerSpectrumDensityforVehicleLateralPosition00.10.20.30.40.50.60500100015002000TIMEPSDDAY-1DAY-2DAY-3DAY-4ak2k1nTCISRGW-TRISteeringAnglefiltercorrectionforcurvesCISRGW-TRIHypothesis•Thehypothesizedrelationshipbetweendriverstateofalertnessandsteeringwheelpositionisthatunderanalertstate,driversmakesmallamplitudemovementsofthesteeringwheel,correspondingtosmalladjustmentsinvehicletrajectory,butunderadrowsystate,thesemovementsbecomelesspreciseandlargerinamplituderesultinginsharpchangesintrajectory(Planqueetal.1991).CISRGW-TRIAHybridArtificialNeuralNetworkArchitectureWj1UnsupervisedLayer:ClusteringCompetitiveAlgorithmSupervisedLayer:ClassificationFeedforwardAlgorithm28X8CISRGW-TRIHybridArtificialNeuralNetworkArchitectureUnsupervisedSupervisedAdaptiveNetworkWInputOutputDesiredOutputErrorAdaptiveNetworkWInputOutputCISRGW-TRIANNTrainingforUnsupervisedCompetitiveLayer1.Initializetheweightvectorrandomlyforeachneuron.2.PresenttheinputvectorX(n).3.ComputethewinningneuronusingtheEuclideandistanceasametric.WhereWi=[w1,w2,….w8]Tistheweightvectorofneuroni.biisthebiastostoptheformationofdeadneurons.CISRGW-TRIANNTrainingCompetitiveLayerContinued•Nnumberoftimeaneuronwinsincompetitivelayer•andarelearningconstantsando(n)istheoutcomeofthepresentcompetition(=1ifneuronwins&else=0).•Ciinitiallysettosmallrandomvalue4.UpdatetheweightvectorofthewinningneuronWi*only.5.Continuewithstep(2)twountilchangeintheweightvectorsreachesaminimumvalue.CISRGW-TRIANNTrainingCompetitiveLayerContinued•Thecompetitivealgorithmmovestheweightvectorsofalltheneuronsclosertothecenteroftheclusters.•Eachneuron(orsetofneurons)ofthecompetitivelayerrepresentsacluster.•TheOutputoftheneuronis1ifitwinsthecompetitionand0ifitlosses.•TheOutputoftheCompetitivelayerisann-dimensionalbinaryvectorT(n)=[t1,t2,……..,tn]T.CISRGW-TRIANNTrainingforsupervisedfeedforwardlayer•Step1:Initializethesynapticweightsandthethresholdstosmallrandomnumbers.•Step2:Presentthenetworkwithanepochoftrainingexemplars•Step3:ApplyInputvectorX(n)totheinputlayerandthedesiredresponsed(n)totheoutputlayerofneurons.TheoutputofeachneuroniscalculatedasCISRGW-TRIANNTrainingContinuedCISRGW-TRIANNTrainingContinued•N=No.oftrainingsetsinoneepoch•=Learningrateparameter•=Momentumconstant•Step5:Iteratethecomputationbypresentingnewepochsoftrainingexamplesuntilthemeansquareerror(MSE)comput