基于邻域一致性的点云场景流传播更新方法郑晗1王宁1马新柱2张宏1王智慧1李豪杰11（大连理工大学-立命馆大学国际信息与软件学院辽宁大连116000）2（悉尼大学电气与信息工程系澳大利亚悉尼2006）（zheng_han@mail.dlut.edu.cn）PointCloudSceneFlowPropagationUpdateMethodBasedonNeighborhoodConsistencyZhengHan1,WangNing1,MaXinzhu2,ZhangHong1,WangZhihui1,andLiHaojie11（InternationalSchoolofInformationScienceandEngineering,DalianUniversityofTechnology-RitsumeikanUniversity,Dalian,Liaoning116000）2（DepartmentofElectricalandInformationEngineering,TheUniversityofSydney,Sydney,Australia2006）AbstractSceneflowisa3Dmotionfieldbetweencontinuousdynamicscenes,whichiswidelyappliedinroboticsandautonomousdrivingtasks.Existingmethodsignorethecorrelationofpointcloudpointsandfocusonlyonthepoint-by-pointmatchingrelationshipbetweenthesourcepointcloudandtargetpointcloud,whichisstillchallengingtoestimatethesceneflowaccuratelyatthepointswithinsufficientlocalfeatureinformation,becausethematchingrelationshipdependsentirelyonthefeatureinformationofthepointclouddata.Consideringthecorrelationpropertyofthesourcepointcloud’slocalregions,theNCPUM(neighborhoodconsistencypropagationupdatemethod)isproposedtopropagatethesceneflowfromhigh-confidencepointstolow-confidencepointsinlocalregions,soastooptimizethesceneflowatthepointswithinsufficientlocalfeatureinformation.Specifically,NCPUMconsistsoftwomodules:theconfidencepredictionmodule,whichpredictstheconfidenceofthesourcepointcloudaccordingtotheprioridistributionmapofsceneflow;thesceneflowpropagationmodule,whichupdatesthesceneflowofthelowconfidencepointsetbasedonthelocalareaconsistencyconstraint.WeevaluateNCPUMonbothchallengingsyntheticdatafromFlyingthing3DandrealLidarscansfromKITTI,andexperimentresultsoutperformbypreviousmethodsalargemargininaccuracy,especiallyonKITTIdataset,becausetheneighborhoodconsistencyismoreapplicablewiththeaprioriassumptionsofrealLidarscans.Keywordssceneflow；pointcloud；neighborhoodconsistency；confidence；deeplearning摘要场景流是连续动态场景...
