高性能低功耗实时多目标视觉感知引擎

시각인식가속기를집적한고성능저전력실시간다중물체인식프로세서HighPerformanceLowPowerReal-TimeMulti-ObjectRecognitionProcessorwithVisualPerceptionEngineHighPerformanceLowPowerReal-TimeMulti-ObjectRecognitionProcessorwithVisualPerceptionEngineADVISOR:ProfessorYoo,Hoi-JunByKim,Joo-YoungSchoolofElectricalEngineeringandComputerScienceDivisionofElectricalEngineeringKoreaAdvancedInstituteofScienceandTechnologyAthesissubmittedtothefacultyoftheKAISTinpartialfulfillmentoftherequirementsforthedegreeofDoctorofPhilosophyintheSchoolofElectricalEngineeringandComputerScience,DivisionofElectricalEngineering.Daejeon,Korea2009.11.30Approvedby________________________ProfessorYoo,Hoi-Jun시각인식가속기를집적한고성능저전력실시간다중물체인식프로세서김주영위논문은한국과학기술원박사학위논문으로학위논문심사위원회에서심사통과하였음.2009년11월19일심사위원장유회준(인)심사위원박규호(인)심사위원나종범(인)심사위원한일송(인)심사위원김기응(인)사랑하는가족에게이논문을바칩니다.iDEE김주영.Kim,Joo-Young.HighPerformanceLowPowerReal-TimeMulti-ObjectRecognitionProcessorwithVisualPerceptionEngine.시각인식가속기를집적한고성능저전력실시간다중물체인식프로세서.SchoolofElectricalEngineeringandComputerScience,DivisionofElectricalEngineering.2010.153p.AdvisorProfessorYoo,Hoi-Jun.TextinEnglish20075036AbstractA201.4GOPSreal-timemulti-coreprocessorisproposedtominimizetheenergyconsumptionperframeformulti-objectrecognition.Inalgorithmlevel,wedevisethevisualperceptionbasedobjectrecognitionmodelthatextractstheregions-of-interest(ROIs)ofobjectsfirstandthenperformsdetailobjectrecognitionprocessingconsistingofimageprocessinganddatabasematching.Byreducingtheeffectiveprocessingareaby3times,theproposedobjectrecognitionalgorithmlargelyimprovesenergyefficiency.Toimplementtheproposedvisualperceptionbasedobjectrecognitioninenergyefficientway,weproposetheattentioncontrolledmulti-corearchitectureconsistingoftwoIPlayershavingdifferentroles.Theattention/controlIPsestimatetheglobalworkloadsofoverallimage,e.g.,thenumberofROItiles,andcontrolmultipleprocessingcoresbasedonthemforefficientimageprocessing.Inthisarchitecture,wefindouttheenergyefficientsolutionsforthefourissues:multi-coreorganization,ROIprocessingmodel,ROItaskscheduling,andIPcommunication.Puttingthesolutionstogether,theproposedarchitectureimproves3.7timesenergyefficiencyfromtheconventionalmulti-corearchitecture.Afterthat,wetailortheproposedarchitecturetothree-stagepipelinedarchitecturetomaximizeiithethroughputofobjectrecognition.Inaddition,wealsodesignanapplication-specificnetwork-on-chiptoresolvethedatacommunicationproblemsunderthetargetobjectrecognitiontraffic.Forchipimplementation,wedesignedthreekindsofIPblocksforthreestagesoftheproposedobjectrecognition,visualperception,mainimageprocessing,andpostdatabasematching.Especially,inthedesignofvisualperception,weexploitbio-inspiredneuralnetworksandfuzzylogiccircuitsforhuman-likeROIestimation.WeemploymultipleSIMDprocessorsandvectormatchingacceleratorformainimageprocessingandpostdatabasematching,respectively.Inoverallchip,workload-awaredynamicpowermanagementisperformedforlowpowerobjectrecognition.Asaresult,thefabricatedprocessorachieves60frame/sec496mWobjectrecognitionupto10differentobjectsforVGA(640x480)videoinput.Obtained8.2mJ/frameenergyconsumptionisthelowestamongthepreviousworksand3.2timesofthestate-of-the-art.Theembeddedvisionsystemisdevelopedtooperatethefabricatedchipandthedevelopedsystemisappliedtovariousapplicationssuchasmulti-objectrecognition,trafficsignrecognition,andmobilerobotvision.iiiTableofContentsChapter1Introduction.......................................................................................................11.1.Backgrounds...................................................................................................11.2.Previousworks................................................................................................41.3.Contributionofthiswork................................................................................7Chapter2VisualPerceptionbasedObjectRecognitionAlgorithm...................................92.1.Introduction.....................................................................................................92.2.Previousattentionbasedobjectrecognition...................................................102.3.Visualperceptionbasedobjectrecognition....................................................112.4.Visualperceptionalgorithm...........................................................................122.5.Three-stageobjectrecognitionalgorithm......................................................142.6.Modelverification.........................................................................................16Chapter3AttentionControlledMulti-coreArchitecture..................................................183.1.Introduction....................................................................................................183.2.Previousarchitectures....................................................................................183.3.Attentioncontrolledmulti-corearchitecture..................................................213.3.1.Overallblockdiagram.......................................................