攻击行为估计---基于神经网络的人体分割与行为识别研究综述(IJEME-V9-N1-2)

I.J.EducationandManagementEngineering,2019,1,9-19PublishedOnlineJanuary2019inMECS()DOI:10.5815/ijeme.2019.01.02Availableonlineat:AComprehensiveReviewonNeuralNetworkBasedHumanSegmentationandActionRecognitionA.F.M.SaifuddinSaifa,Md.AkibShahriarKhana,AbirMohammadHadia,RahulPrashadKarmokera,JoyJulianGomesaaFacultyofScienceandTechnology,AmericanInternationalUniversity–Bangladesh(AIUB),Dhaka,BangladeshReceived:11October2018;Accepted:17December2018;Published:08January2019AbstractHumanactionrecognitionhasbeenatalkedtopicsincemachinevisionwascoined.Withtheadventofneuralnetworksanddeeplearningmethods,variousarchitecturesweresuggestedtoaddresstheproblemswithinacontext.Convolutionalneuralnetworkhasbeentheprimarygo-toarchitectureforimagesegmentation,flowestimationandactionrecognitioninrecentdays.Astheproblemitselfisanextendedversionofvarioussub-problems,suchasframesegmentation,spatialandtemporalfeatureextraction,motionmodelingandactionclassificationasawhole,somemethodsreviewedinthispaperaddressedsub-problemsandsometriedtoaddressasinglearchitecturetotheactionrecognitionproblem.Whilebeingasuccess,convolutionneuralnetworkshavedrawbacksinitspoolingmethods.CapsNet,ontheotherhand,usessquashingfunctiontodeterminetheactivation.AlsoitaddressesspatiotemporalinformationwiththenormalizedvectormapswhileCNN-basedmethodsextractsfeaturemapforspatialandtemporalinformationandlateraugmenttheminafusionlayerforcombiningtwoseparatefeaturemaps.Criticalreviewofpapersprovidedinthisworkcancontributesignificantlyinaddressinghumanactionrecognitionproblemasawhole.IndexTerms:CapsuleNetwork,NeuralNetwork,ImageSegmentation,FlowEstimation,ActionRecognition.©2019PublishedbyMECSPublisher.Selectionand/orpeerreviewunderresponsibilityoftheResearchAssociationofModernEducationandComputerScience.1.IntroductionActionrecognitioninvideoscanhavearadicalimpactonhumanlife.Numerousattemptshavebeentakentosolvetheactionrecognitionchallenges.Duetohugecollaborativeeffortsincomputervisioncommunity,*Correspondingauthor:E-mailaddress:[e-mail],[e-mail],[e-mail],[e-mail],[e-mail]10AggressiveActionEstimation:AComprehensiveReviewonNeuralNetworkBasedHumanSegmentationandActionRecognitionsimpleactionsofwaving,standingetc.fromKTHandWeizmanndatasetarenowconsideredasobsoletechallengesandthecommunityhasmovedontosolvemorecomplexactionslikesportsandhumaninteractions.However,despitehavingthepotentialtoimprovesecurityandsurveillanceapplications,therehasnotbeenmuchimprovementinregardtotheviolentsceneandaggressivebehaviordetectionwhichisaspecialcaseofactionrecognition.Previousworksonactionrecognitionheavilyreliedontheusageofhard-codedtechniquessuchasMoSIFT,OpticalFlowandDenseTrajectory.Thesehard-codedtechniquesarecomputationallyexpensivewhileofferinglowperformance.InrecenttimesafterthesuccessofAlexNet,awaveofworksapproachedtheproblemfromanewviewpointusingconvolutionalneuralnetworks.Thoughbeingincisiveinimageclassificationtasks,ConvolutionalNeuralNetworksdidnotfarewellimmediatelyagainstalreadyestablishedmethodsinactionrecognition.DifferenttypesoffusiontechniquesusingbothdensefeaturesandCNNimprovesperformance.Two-streamnetworksand3D-CNNusingmotionfeaturessuchasopticalflowandRNNinconjunctionwiththeaforementionedtechniquesalsogaveaboostinperformance.Buttheseapproacheshavesomeseveredisadvantageslikemax-poolingwhichinmostcasessuppressestinybutimportantfeaturesand,susceptibletoadversarialattacks.Thoughthesemethodswork,theydonotprovideanyinsightintohowtheinnermechanismfunctions.ThenewlyproposedCapsNetarchitecturecanhelptobridgethegapasthisparticularsystemfollowsapart-to-wholeapproachandproducesvectoroutputs,unlikeCNNwhichhasscalaroutputs.Capsulesareparticularlygoodathandlingdifferenttypesofvisualstimulusandencodingthingslikepose(position,size,andorientation),deformation,velocity,albedo,hue,textureetc.thatisnotpossibleforCNN.Capsulesencapsulateall-importantinformationaboutthestateofthefeaturetheyaredetectinginvectorform.Therestofthepaperisorganizedasfollows.Section2discussesthechallengesofactionrecognitionandprovidesaconciseviewofabroadrangeoftechnologiesandapproachesthatareusedtosolvetheproblem.Insection3numerousmethodsrelatedtoactionrecognitionarereviewed.Section4elaboratestheframeworksusedinthemethoddescribedinsection3.Section5providesdetailsontheexperimentalsettingsandperformanceofthemethods.Insection6keyfindingsfromthemethodsaresummarized.Section7concludesthepaperemphasizingtheimpactoftheproblem.2.CoreBackgroundStudyHumanactionrecognitionisanintegralprobleminspatiotemporalinformationextraction,fusion,learninganddetectionfromvideostreams,bothinstaticandespeciallyinalivefeedanalysis.Numerousstudieshavebeenconductedbasedonhard-codedfeatureextraction,poseestimation,frameanddynamicsandalsoasneuralnetworklearningproblem.Theproblemindiscussionisaddressedbysub-problemsthatinclude:framepreprocessing(ifany),featureextraction(bothspatialandtemporal),learningthefeat