基于 BoW 模型的增强 RGB-D SLAM 算法-《计算机技术与发展》

文章信息/Info

Author(s):: JING Shu-xu; LU Peng-yu; ZHAI Xiao-hui; GAO Tao; School of Information Engineering,Chang’an University,Xi’an 710064,China

Keywords:: RGB-D SLAM; BoW model; PROSAC algorithm; back-loop detection; pose-graph optimization; grid map

摘要:: 即时定位与地图构建(simultaneous localization and mapping,SLAM) 被认为是机器人自主运动的核心技术。针对目前的 RGB-D SLAM 算法实时性和鲁棒性差的问题，提出了一种增强的 RGB-D SLAM 算法。提取 RGB 图像的 ORB 特征描述子,然后利用 BoW( bag of word) 模型缩小特征描述子的匹配范围从而提高算法的实时性; 接着采用 PROSAC 算法结合PnP 算法解算初始相机位姿并通过非线性优化的方式得到优化的相机位姿; 利用 BoW 模型结合关键帧技术和结构一致性几何约束提高回环检测的鲁棒性;采用通用图优化工具 g2o 对位姿图进行优化,得到全局一致的位姿和点云;最后采用贪心三角化算法将点云转换成网格地图。针对 Fr1 数据集, 该算法的平均定位误差为 0. 079 7 m,每帧数据平均处理时间为0.04 s。与 RGB-D SLAM 原始算法相比,该算法具有良好的实时性和鲁棒性,可以满足机器人实时 SLAM 的要求。

Abstract:: Simultaneous localization and mapping (SLAM) is considered to be the core technology for autonomous motion of robots. Anenhanced RGB-D SLAM algorithm is proposed to solve the problem of poor real-time performance and robustness of the current RGB-D SLAM algorithm. The ORB feature descriptor of RGB image is extracted,and then the BoW (bag of word) model is used to reducethe matching range of the feature descriptor to improve the real-time performance of the algorithm. Secondly,PROSAC algorithm andPnP algorithm are used to calculate the initial pose of the camera,and the nonlinear optimization is used to optimize the initial pose toobtain a high precision pose. The BoW model is combined with keyframe technique and structure - consistent geometric constraints toimprove the robustness of loop detection. The general graph optimization tool g2o is used to optimize the pose-graph to get the globallyconsistent pose and point cloud. Finally,greedy-triangulation algorithm is used to transform point cloud into grid map. For the Fr1 dataset,the average RMSE is 0. 079 7 m, and each frame costs 0. 04 s. Compared with the original algorithm of RGB - D SLAM, thisalgorithm has better real-time performance and robustness,which can meet the requirements of robot real-time SLAM.