Fast Recurrent Fully Convolutional Networks for Direct Perception in Autonomous Driving

《Fast Recurrent Fully Convolutional Networks for Direct Perception in 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、FastRecurrentFullyConvolutionalNetworksforDirectPerceptioninAutonomousDrivingEric(Yiqi)HouSaschaHornauerKarlZipserBerkeleyDeepDriveUniversityofCalifornia,BerkeleyBerkeleyDeepDrive109McLaughlinHall(ICSI)1947CenterSt109McLaughlinHallBerkeley,CA94720-1720Berkeley,CA94704Berkeley,CA94704

2、eric.hou@berkeley.edusaschaho@icsi.berkeley.edukarlzipser@berkeley.eduAbstractDeepconvolutionalneuralnetworks(CNNs)havebeenshowntoperformextremelywellatavarietyoftasksin-cludingsubtasksofautonomousdrivingsuchasimageseg-mentationandobjectclassification.However,networksde-signedforthese

3、taskstypicallyrequirevastquantitiesoftrainingdataandlongtrainingperiodstoconverge.Weinvestigatethedesignrationalebehindend-to-enddrivingnetworkdesignsbyproposingandcomparingthreesmallandcomputationallyinexpensivedeepend-to-endneuralnetworkmodelsthatgeneratedrivingcontrolsignalsdi-rec

4、tlyfrominputimages.Incontrasttopriorworkthatseg-Figure1.F-RFCNArchitecturementstheautonomousdrivingtask,ourmodelstakeonanovelapproachtotheautonomousdrivingproblembyuti-lizingdeepandthinFullyConvolutionalNets(FCNs)withandrequirehighamountsofcomputationalpower.3Dob-recurrentneuralnetsa

5、ndlowparametercountstotacklejectdetection[5],forinstance,takes0.36sforinferenceacomplexend-to-endregressiontaskpredictingbothsteer-aloneonaTitanX,orlessthan3inferencespersecond,ingandaccelerationcommands.Inaddition,weincludeandreliesonamodifiedbutstillparameter-heavyVGG-layersoptimize

6、dforclassificationtoallowthenetworksto16.Whileextremelyaccurate,withmodern-dayprocessorimplicitlylearnimagesemantics.Weshowthattheresult-limitations,suchmodelsaretooheavytoreachbetterthaningnetworksuse3xfewerparametersthanthemostre-humanreactiontimesof100ms(10FPS)comparedtotra-centcom

7、parableend-to-enddrivingnetwork[2]and500xditionalsystemslikeLIDAR(24FPS)[1]andalgorithmsfewerparametersthantheAlexNetvariationsandconvergebasedontraditionalmachinelearning(100FPS)[20]thatarXiv:1711.06459v2[cs.CV]20Nov2017bothfasterandtolowerlosseswhilemaintainingrobust-achievehigherF

8、PSattheexpen