基于分段匹配点云位置识别方法研究.pdfVIP

2017IEEEInternationalConferenceonRoboticsandAutomation(ICRA)

Singapore,May29-June3,2017

SegMatch:SegmentBasedceRecognitionin3DPointClouds

RenaudDubéDanielDugasElenaStummJuanNietoRolandSiegwartCesarCadena∗

—cerecognitionin3Ddataisachallenging

taskthathasbeencommonlyapproachedbyadaptingimage-

basedsolutions.Methodsbasedonlocalfeaturessufferfrom

ambiguityandfromrobustnesstoenvironmentchangeswhile

methodsbasedonglobalfeaturesareviewpointdependent.

WeproposeSegMatch,areliablecerecognitionalgorithm

basedonthematchingof3Dsegments.Segmentsprovide

agoodcompromiseweenlocalandglobaldescriptions,

incorporatingtheirstrengthswhilereducingtheirindividual

drawbacks.SegMatchdoesnotrelyonassumptionsof‘perfect

segmentation’,orontheexistenceof‘objects’intheenvi-

ronment,whichallowsforreliableexecutiononlargescale,

unstructuredenvironments.Wequantitativelydemonstratethat

SegMatchcanachieveaccuratelocalizationatafrequencyof

1HzonthelargestsequenceoftheKITTIodometrydataset.WeFig.1:Anillustrationofthepresentedcerecognitionframework.The

furthermoreshowhowthisalgorithmcanreliablydetectandreferencepointcloudisshownbelow(inwhite),andthelocalpointcloud

closeloopsinreal-time,duringonlineoperation.Inaddition,theisalignedabove.Coloursareusedtoshowthepointcloudsegmentation,

sourcecodefortheSegMatchalgorithmismadepubliclyandsegmentmatchesareindicatedwithgreenlines.

1

available.