A 201.4 GOPS real-time multi-object recognitionprocessor is presented with a three-stage pipelined architecture.Visual perception based multi-object recognition algorithm isapplied to give multiple attentions to multiple objects in the inputimage. For human-like multi-object perception, a neural perceptionengine is proposed with biologically inspired neural networksand fuzzy logic circ
36IEEEJOURNALOFSOLID-STATECIRCUITS,VOL.45,NO.1,JANUARY
2010
Fig.5.(a)Workload-awaretaskscheduling.(b)Applieddatabasesize
control.
Fig.6.BlockdiagramofneuralperceptionengineandSPUtask/powermanager.
proportionaltothenumberofinputdescriptorvectorsandthesizeofapplieddatabase.Basedonthese,theexecutiontimeoftheobjectdecisionstagecanbecontrolledbycon guringcoveragerateofdatabase.First,theSTMmeasuresthenumberofdescriptorvectorsfromtheSPUsandcalculatestheexpectedexecutiontimeofthevectormatching.Then,itcomparestheexpectedexecutiontimewiththetargetpipelinetimeandcon guresthedatabasecoveragerateoftheDPtomeetthepipelinetime.However,reducingcoveragerateshouldbecare-fullyperformedbecauseitcandegradetheoverallrecognitionrate.Witha16384-entrydatabasefor50objectsrecognition,correctlymatchedratedegrades0.6%and1.3%,whenthecoveragerateis0.95and0.90,respectively.WiththehelpoftheWATSandADSC,theexecutiontimesofthethreestagescanbebalancedtothetargetpipelinetime,16ms,evenunder
theworkloadvariations.Asaresult,theproposedprocessorachieves60frame/secframe-rateforVGA
(640480)sizedvideoinput.
IV.BUILDINGBLOCKDESIGN
A.NeuralPerceptionEngine
Fig.6showstheblockdiagramoftheNPE.Foref cientROIdetection,theNPEemploysa32-bitRISCcontrollerandthreehardwareengines;motionestimator(ME),visualattentionen-gine(VAE),andobjectdetectionengine(ODE).TheMEisem-ployedtoextractdynamicmotionvectorsbetweentwosequen-tialframesandimplementedbyarrayPEswithafullsearchblockmatchingmethod[13].TheVAEisemployedtoextract