MultiNet Multi-Modal Multi-Task Learning for Autonomous Driving

《MultiNet Multi-Modal Multi-Task Learning for Autonomous Driving》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、MultiNet:Multi-ModalMulti-TaskLearningforAutonomousDrivingSauhaardaChowdhuri1TusharPankaj2KarlZipser3RCAbstract—Severaldeeplearningapproacheshavebeenap-USBHubReceiverSSDpliedtotheautonomousdrivingtask,manyemployingend-to-enddeepneuralnetworks.Autonomousdrivingiscomplex,utilizingmultiplebehav

2、ioralmodalitiesrangingfromlanechangingtoturningandstopping.However,mostexistingap-proachesdonotfactorinthedifferentbehavioralmodalitiesofthedrivingtaskintothetrainingstrategy.ThispaperdescribesatechniqueforusingMulti-ModalMulti-TaskLearning,whichwedenoteasMultiNetwhichconsidersmultiplebehavi

3、oralmodalitiesasdistinctmodesofoperationforanend-to-endautonomousdeepneuralnetworkutilizingtheinsertionofmodalinformationassecondaryinputdata.Usinglabeleddatafromhoursofdrivingourfleetof1/10thscalemodelcars,weArudino#2Arudino#1NVIDIAtraineddifferentneuralnetworkstoimitatethesteeringangleLi

4、PoZEDStereoSensorandMotorandJetsonTX1BatteryCameraanddrivingspeedofhumancontrolofacar.WeshowthatdisplayControllerServoControllerineachcase,MultiNetmodelsoutperformnetworkstrainedFig.1:CarDiagramonindividualtasks,whileusingafractionofthenumberofparameters.I.INTRODUCTIONdifferenttask.Ifinstead

5、thenetworkisgiventhetaskoftranscribingtextintwomodes:oneforaudio,andtheotherMostcurrentresearchondrivingwithDNNshasfocusedforvideorecordings[8].Thisisanexampleofmulti-modalonasingledrivingmodality,e.g.lanefollowingorobstaclelearningastherearemultiplemodesofrunningthenetworkavoidance[1],[2],[

6、3],[4].Weconsidertheseapproacheseitherofwhichcanbeusedduringevaluation.asSingleTaskLearning(STL),astheyfocusontrainingtoWorkonmulti-modallearninghaspredominantlybeenperformanindividualtask.focusedinfieldsotherthanroboticsorlocomotion;e.g.Multi-tasklearning(MTL)researchhasshownthattrain-speech

7、recognitionwithaudioandvideo[8],[7].Withiningonsidetasksrelatedtothemainoperationofadeeptheseworks,itiscommonforDNNstobegiveninputthatneuralnetworkcanenhanceitslearningcapabilities[5],[6].couldcorrespondtoanyormultiplemodesofoperation.Thesesidetask