动态场景中基于特征点几何约束的3D SLAM算法  

3D SLAM algorithm based on geometric constraints of feature points in dynamic scenarios

作　　者：刘铭坚 罗景文 秦世引[2] LIU Mingjian;LUO Jingwen;QIN Shiyin(School of Information Science and Technology,Yunnan Normal University,Kunming 650500,China;School of Automation Science and Electrical Engineering,Beihang University,Beijing 100191,China)

机构地区：[1]云南师范大学信息学院,昆明650500 [2]北京航空航天大学自动化科学与电气工程学院,北京100191

出　　处：《北京航空航天大学学报》2024年第9期2872-2884,共13页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家自然科学基金(62063036);云南师范大学博士科研启动项目(01000205020503115)。

摘　　要：针对动态场景中动态物体会导致机器人在进行位姿估计时引入大量动态误差的问题,提出一种利用特征点间几何约束来剔除动态特征点的移动机器人3D同步定位与地图构建(SLAM)算法。利用当前帧的ORB特征点与上一帧特征点生成的地图点进行投影匹配,通过引入Delaunay三角剖分法构建能够表示2帧间、匹配地图点间几何关系的三角网。利用相邻2帧地图点的几何关系变化检测出动态特征点,考虑到静态特征点可能被误检测为动态特征点而导致特征点缺失的情况,在相邻2帧匹配时提取更多的特征点以实现静态特征点的补偿,进而剔除动态特征点,实现对移动机器人位姿的精确估计。在此基础上,通过引入滑动窗口提取关键帧并完成闭环检测,从而构建出精确的3D稠密地图。在多组公开数据集上进行仿真实验及室内动态场景下的实验,结果表明,所提算法能够有效剔除动态特征点,提高移动机器人在动态场景中位姿估计的精度和地图的一致性。The dynamic objects will cause a large number of dynamic errors in the pose estimation of robots in dynamic scenarios.To address this issue,a 3D simultaneous localization and mapping(SLAM)algorithm for mobile robots was presented by using geometric constraints between feature points to eliminate dynamic feature points.First,the ORB feature points of the current frame and the map points generated by feature points of the previous frame were used for projection matching,and the Delaunay triangulation method was introduced to construct a triangle net that could represent the geometric relationship between the matching map points of the two frames.Then,the dynamic feature points were detected according to the geometric relationship changes of the map points in the adjacent two frames.Since the static feature points may be incorrectly detected as dynamic feature points,which thus brings about the loss of feature points,more feature points were extracted during the matching of the adjacent two frames,so as to compensate for static feature points.Then,the dynamic feature points were eliminated,and the pose of the mobile robots was estimated accurately.On this basis,a sliding window was introduced to extract key frames and complete closed-loop detection,and thus an accurate 3D dense map was constructed.The results of simulation experiments on multiple sets of public datasets and the experiments in the indoor dynamic scenarios show that the proposed algorithm in this paper can effectively eliminate the dynamic feature points and improve the accuracy of the pose estimation of mobile robots in dynamic scenarios and the consistency of the map.

关 键 词：动态场景 同步定位与地图构建 移动机器人 几何约束 三角剖分 

分 类 号：TP242.6[自动化与计算机技术—检测技术与自动化装置] TB553[自动化与计算机技术—控制科学与工程]