自动驾驶论文0618 Uncertainty-Aware Fast Curb Detection Using Convolutional Networks in Point Clouds.pdfVIP

2021IEEEInternationalConferenceonRoboticsandAutomation(ICRA2021)

May31-June4,2021,Xian,China

Uncertainty-AwareFastCurbDetectionUsing

ConvolutionalNetworksinPointClouds

YounghwaJung,MinguJeon,ChanKim,Seung-WooSeo,andSeong-WooKim

Abstract—Curbdetectionisanessentialfunctionofau-

tonomousvehiclesinurbanareas.However,curbsaredifficult

todetectincomplexurbanenvironmentsinwhichmany

dynamicobjectsexist.Additionally,curbsappearinavarietyProposed

ofshapesandsizes.PreviousstudieshavebeenbasedontheMethod

traditionalpipeline,whichconsistsoftheextractionandaggre-

gationofhand-craftedfeaturesthatarethenfedtoclassifiers.

However,thissequentialprocessisinefficientanddesigningthe

hand-craftedfeaturesisacomplexprocess.Recently,thiskind

ofprocesshasbeenreplacedbyDeepNeuralNetworks(DNN),InputPredictionUncertainty

inwhichclassifiersandfeaturesarelearnedfromlarge-scale

data.VeryfewworkshaveexploitedDNNforthecurbdetectionFig.1.Leftsideofthefigureshowsadrivingenvironmentwherethe

problem.Mostworksusemulti-modalsensor-basedmethodspurple-dottedboundingboxrepresentsthedynamicobjectsandtheredarea

thatcombineimagesandaccumulated3Dpointcloudsfromreferstotheoccludedregion.Theproposedmethodusestheencodedpoint

LIDAR.However,theseapproachesrequiresynchronizationcloudstoestimatethelocationofthecurbs(lightblue)andtheuncertainty

andcalibrationbetweensensors.Inaddition,theydonoto