基于运动分割的动态SLAM联合优化算法

打开文本图片集

Dynamic SLAM joint optimization algorithm based on motion segmentation

ZhangYutong，Fan Xinyue，Zhou Zhiyuan，Xie Yuanyuan （ScholofCommicatio&IfoatioEnginering，CongqingUniesityofosts&Telecomnitios，ChongingChina）

Abstract：InordertoaddressthechallngeoftraditionalSLAMbeingeasilydisruptedbydynamicobjectsinthescene，this paper proposedavisual SLAMalgorithmsuitable fordynamicenvironments，referrdtoas GMS-SLAM.Inthisapproach，itreplacedsemanticsegmentation withageometric methodbasedonepipolarconstraints，andemployedopticaldilation to further eliminatepotentialcoplanarambiguities.Itmodeleddynamicobjectsusing motionconstraintsandrigidityconstraints.Thisaproachutilizedthegraphoptimizationmodule tooptimizethecamerapose information，andtransformedhumanskeletalmodelingintorigidojects tobeincorporatedintothegraphoptimizationprocess.Aditionally，byexploitingthestrongcorelation betweencameraself-motionandepipolarconstraints，itintegratedthesegmentationnetworkandoptimizationmoduleintoa unifiedframework forjointrefinement.ExperimentalvalidationontheKITTIandhighlydynamicShibuyadatasetsdemonstratesthat GMS-SLAMachievessubstantialimprovements inacuracycompared tostate-of-the-artdynamicvisualalgoritms， suchas DytanVO，and clasical dynamic algorithms ike DynaSLAM，exhibiting superiorlocalization performance in dynamic environments.

Key Words：dynamic SLAM；graph optimization；motion segmentation；rigid constrains

0 引言

SLAM系统帮助机器人在未知环境中寻找自身位置，构建轨迹地图以达到识别整个环境的目的。（剩余12666字）