基于动态耦合与空间数据关联的RGB-D SLAM算法

An RGB-D SLAM Algorithm Based on Dynamic Coupling and Spatial Data Association

为了解决动态环境下视觉SLAM(同步定位与地图创建)算法定位与建图精度下降的问题,提出了一种基于动态耦合与空间数据关联的RGB-D SLAM算法。首先,使用语义网络获得预处理的语义分割图像,并利用边缘检测算法和相邻语义判定获得完整的语义动态物体;其次,利用稠密直接法模块实现对相机姿态的初始估计,这里动态耦合分数值的计算在利用了传统的动态区域剔除之外,还使用了空间平面一致性判据和深度信息筛选;然后,结合空间数据关联算法和相机位姿实时更新地图点集,并利用最小化重投影误差和闭环优化线程完成对相机位姿的优化;最后,使用相机位姿和地图点集构建八叉树稠密地图,实现从平面到空间的动态区域剔除,完成静态地图在动态环境下的构建。根据高动态环境下TUM数据集测试结果,本文算法定位误差相比于ORB-SLAM算法减小了约90%,有效提高了RGB-D SLAM算法的定位精度和相机位姿估计精度。

In order to solve the problem of accuracy decline of positioning and mapping of the visual SLAM(simultaneous localization and mapping) algorithm in a dynamic environment, an RGB-D SLAM algorithm is proposed based on dynamic coupling and spatial data association. Firstly, the semantic network is used to obtain pre-processed semantic segmentation images, the edge detection algorithm and adjacent semantic judgment are combined to obtain the complete semantic dynamic objects. Secondly, the initial camera attitude is estimated in the dense direct method module, and wherein the dynamic coupling scores are calculated not only by the traditional dynamic region elimination, but also by the spatial plane consistency and the depth information screening. Further, the map point set is updated in real time by the spatial data association algorithm and the camera pose, and then pose of the camera is optimized by minimizing the reprojection error and the closedloop optimization process. Finally, the octree dense map is constructed by using the camera pose and the map point set to eliminate all the dynamic region, from plane to space, and the static map is constructed in dynamic environment. According to the test results on TUM data set in a high-dynamic environment, the positioning error of the proposed algorithm is reduced by about 90% compared with that of ORB-SLAM algorithm, and the positioning accuracy and the camera pose estimation accuracy of RGB-D SLAM algorithm are effectively improved by the proposed algorithm.