第50卷第2期2023年2月Vol.50,No.2Feb.2023湖南大学学报(自然科学版)JournalofHunanUniversity(NaturalSciences)基于改进ORB-SLAM2算法的RGB-D稠密地图构建韩彦峰†,唐超超,肖科(重庆大学机械与运载工程学院,重庆400044)摘要:针对传统ORB-SLAM2系统无法稠密建图以及缺少可应用于移动机器人在特定场景下导航与路径规划的八叉树地图等问题,提出了一种改进RGB-DORB-SLAM2算法.该算法利用深度信息,通过针孔成像模型计算出关键帧点云的三维空间位置,采用离群点滤波去除多余杂点,体素滤波保留具有特征信息的点云进行拼接,降低地图冗余度,并经过稠密回环处理进一步优化更新关键帧的点云位姿,构建出精确的稠密点云地图并转换为八叉树地图.试验数据表明,相比于RGB-DSLAMV2系统,RGB-DORB-SLAM2系统的全局轨迹误差和相对位姿误差提升有50%以上,均方根误差为0.89%,均值误差为0.76%;在建图性能方面,相比于同类型算法,点云数量平均降低约30%.此外,八叉树地图相比于点云地图,仅占其内存的0.6%,更加满足了高精度、快速性的导航需求.关键词:同时定位与地图构建;体素滤波;离群点滤波;稠密回环;八叉树地图中图分类号:TP242.6文献标志码:ARGB-DDenseMapConstructionBasedonImprovedORB-SLAM2AlgorithmHANYanfeng†,TANGChaochao,XIAOKe(CollegeofMechanicalandVehicleEngineering,ChongqingUniversity,Chongqing400044,China)Abstract:AnimprovedRGB-DORB-SLAM2algorithmisproposedtoaddresstheproblemsthatthetraditionalORB-SLAM2systemcannotbuildadensemapandlackanoctreemapthatcanbeappliedtothenavigationandpathplanningofmobilerobotsinspecificscenarios.Thealgorithmusesdepthinformationtocalculatethe3Dspatialpositionofthepointcloudsinkeyframesthroughthepinholeimagingmodelandusesoutlierfilteringtoremovere⁃dundantclutterandvoxelfilteringtoretainpointcloudswithfeatureinformationforstitchingandreducemapredun⁃dancy.Andthroughdenseloopbackprocessing,thepointcloudposesarefurtheroptimizedandupdatedinkey⁃frames,andanaccuratedensepointcloudmapisconstructedandconvertedintoanoctreemap.TheexperimentaldatashowsthatcomparedwiththeRGB-DSLAMV2system,theglobaltrajectoryerrorandrelativepositionalerroroftheRGB-DORB-SLAM2systemareimprovedbymorethan50%,therootmeansquaree...