Dynamic multi-feature RGBD-SLAM algorithm based on region mapping depth optimization
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
万方数据
维普
针对RGBD-SLAM在特征提取环节,由深度相机得到的深度图易出现深度不连续以及突然出现动态物体造成轨迹误差增大的问题,提出一种基于区域映射深度优化的动态多特征 RGBD-SLAM 算法.利用轻量级的语义分割网络获取 RGB 帧中的语义信息,将得到的语义掩码映射到其对应的深度图像中,并在掩码映射后的区域内进行近邻修复以完成深度图的优化.为减少动态物体对SLAM系统轨迹精度的影响,通过迭代最近点求解相机位姿,结合多视图几何得到物体位姿、运动估计矩阵以及动态视觉误差,进而估计物体运动状态并剔除动态物体.根据双向映射背景修复模型补全剔除区域的静态信息,并进行点线面特征的提取以完成定位与建图的任务.在公开数据集 TUM 中进行验证,实验结果表明所提算法相较于ORB-SLAM3、RGBD-SLAM、DS-SALM的平均绝对轨迹误差分别减少了78.2%、81.4%以及 17.6%,表现了良好的轨迹精度与构图能力.
In order to solve the problem that the depth map obtained by the depth camera is prone to depth discontinuity and the trajectory error increases due to the sudden appearance of dynamic objects in the feature extraction process of RGBD-SLAM,a dynamic multi-feature RGBD-SLAM algorithm based on region mapping depth optimization is proposed.A lightweight semantic segmentation network is used to obtain the semantic information in RGB frames.The semantic mask obtained is mapped to the corresponding depth image,and the depth map is optimized by the nearest neighbor repair in the mask mapped area.In order to reduce the influence of dynamic objects on the trajectory accuracy of SLAM system,the pose of the camera is solved by iterative closest point,and the pose of the objects,motion estimation matrix and dynamic visual error are obtained by combining multi-view geometry,and then the motion state of the objects is estimated and the dynamic objects are eliminated.According to the bidirectional mapping background repair model,the static information of the eliminated region is completed,and the point,line and surface features are extracted to complete the task of localization and mapping.In the open data set TUM,the experimental results show that the average absolute trajectory error of the proposed algorithm is reduced by 78.2%,81.4%and 17.6%,respectively,compared with ORB-SLAM3,RGBD-SLAM and DS-SALM,which shows good trajectory accuracy and composition ability.
simultaneous localization and map constructiondeep optimizationremove objectsbackground restoration
万方数据
维普
