AttenD算法说明

IssuesrelatedtothedriverdistractiondetectionalgorithmattendKatjaKircher&ChristerAhlströmSwedishNationalRoadandTransportResearchInstitute(VTI)AbstractDriverdistractionisacontributingfactortomanycrashes.Areal-timedistractionwarningsystemhasthepotentialtomitigateorcircumventmanyofthesecrashes.objective:thoroughlydescribethedistractiondetectionalgorithmAttenDexplainthetheoryunderlyingdifferentdesignchoicesSummary:visualdistractioninreal-timebasedonsinglelongglancesaswellasrepetitiveglancescoreideaofthealgorithm2-secondtimebufferlooksawayfromtheroad↓looksbackattheroad↑thebufferrunsempty→distracted.IntroductionTypicalexamplesforwardcollisionwarnings(FCW)lanedeparturewarnings(LDW)NavigationsystemsToavoidthosecriticalsituationsaltogetherearlierwarnings(highfalsealarms)takingdriverstateintoaccount(drowsy,intoxicatedordistracted)MethodsfordriverstatemonitoringSteeringwheelmovementslateralpositionbreathanalysersAutomaticeyetrackingremoteeyetrackersGlancebehaviourandtrafficsafetyWikman,NieminenandSummalalargerlateraldisplacements→longerin-carglancesTijerina,ParmerandGoodmanglancefrequencytoin-cardevicescorrelatedwithtasktimeandlaneexceedancesanincreasednumberofshortglancesleadtolargerdecrementsintrafficsafetySomeliteratureseveralglancesawayfromtheroad→moredetrimentalonesingleglanceoftheconsecutiveglancesdriversdonotusuallyincreasethesingleglancedurationformoredifficultorlongertasks,butratherincreasethenumberofglancesawayfromtheroadglancedurationrepeatedglanceDriverDistractionDetectionAlgorithmsVictordeterminethedegreeofdistraction:thepercentageoftimethatthedriver’sgazeisfixatedontheroadcentreduringthelastminuteDonmez,B.,L.N.Boylecurrentoff-roadglancedurationduringthelastthreesecondsZhangandSmithathreesecondsmovingaverageofoff-roadglancedurationpredicteddriverdistractionAconceptfromthe100-carstudyApositiverelationshipbetweentheoccurrenceofcrashesandnearcrashesandaccumulatedoff-roadglancedurationsofmorethantwosecondswithinthelastsixsecondswasfound.Fletcher,L.andA.ZelinskyacounterinsteadofabufferTheAttenDalgorithmassumptiondriver’sattentionisdirectedtowardsthesameobjectsasthegazeglancestothemirrorandthespeedometerarenecessaryforsafedriving.Onlywhentheyarelongerthanonesecondaretheytreatedasdistractions.Whenlookingbacktotheroadthedriverneedstoadaptphysiologicallytolong-distancefocusing.Thisprocessisassumedtolast0.1seconds.FRDTheAttenDalgorithmnotrackingisavailable→headdirectionvectorBuffer=0.4s→outsideofFRDBuffer0.4s→ifheaddirectionvector20degreeSummaryofthedifferentthresholdsandparametersusedbyAttenDAflowchartdepictingtheAttenDalgorithmDiscussionofdifferentfeaturesoftheattendalgorithmdistributionofgazecaseshaveonepeakheaddirectiondistributionisbimodaleyegazedataclearlycontainmoreaccurateandmoredetailedinformationthanheadtrackingdatarawgazecaseswithoutfirstsegmentingthedataintofixationsandsaccadesreductionincomputationalcomplexityThesizeofthebufferiscurrentlysettotwosecondsthebuffersizecouldevenbechangedtoroadtypeglancesawayfromtheroadmorethantwosecondsleadtounacceptablelanedeviationsanddoublestheoddsofacrashprolongedvisualtimesharingisdetrimentaltoattentionnavigationsystemdriverneeds0.1secondsdelaycheckthemirrorsandthespeedometerwassetto1sphysiologicalnaturelikearefocusingoftheeyesmeanvaluesfoundintheliteratureDiscussionofdifferentfeaturesoftheattendalgorithmPossiblemodificationstothealgorithmFusingdatafromtheeyetrackerwithvehicledataisthemoststraightforwardextensionofAttenD.Thealgorithmworkswithgazecasesonlyanddoesnotconsiderfixationsandsaccades.（cognitivedriverdistraction→thesaccadespeedanddistance↑）ExtendingorreducingtheFRDcanbemadebothpermanentlyordependingonenvironmentalfeatures.injunctionsandinurbanareas→FRD↑；motorway→FRD↓Warningstrategy1）warnedforanydistractionoccurrenceAdvantage:getthedrivertolookawayfromtheroadlessfrequentlyDisadvantage:mayannoythedriversomuchthatthesystemisturnedoff2）warningwhenacriticalsituationarisesAdvantage:warningswoulddecreasedrasticallyDisadvantage:Thedriverdoesnotgetanyinformationabouttherequiredaction,decisionprocessmightuseupvaluabletime.situationmightgetevenworseduetoapossiblepanicreactionWarningstrategy3）nottowarnfordistraction,adaptthetimingandpossiblyintensityofotherwarningstothecurrentdriverstategivethedriverinstantfeedbackonwhattodoinacriticalsituation.falsealarmsreducedbywarningthedriverlaterwhenthedriverisdeemedtobeattentive.Itisthinkablethatthethirdstrategyofadaptingothersupportsystemstodriverstatewillbeemployedasthemainstrategy.Thedriverwillstillbewarnedforextremedistractionevents,evenifnoimminentdangercouldbeobserved.certaincriteria:Speed50km/hInhibitwhenpressedthebrakeorperformedseveresteeringmanoeuvresInterval15sDiscussionoffundamentalissuesrelatedtodriverdistractionmanyresearchersoptforgivingthedriverasecondarytasktofulfil(visualorauditorynature).Shortcomings:Thedriverdoesnottrytomaximiseperformanceonbothtasks,butbasicallyforgetsaboutdriving.Themotivationfordoingcomplicatedandmeaninglesscomputa