基于图优化的移动机器人视觉SLAM

针对目前移动机器人视觉SLAM（simultaneous localization and mapping）研究中存在的实时性差、精确度不高、无法稠密化建图等问题，提出了一种基于RGB-D数据的实时 SLAM算法。在本算法前端处理中，采用了鲁棒性与实时性更好的ORB特征检测。利用 RANSAC 算法对可能存在的误匹配点进行剔除完成初始匹配，对所得内点进行PNP求解，用于机器人相邻位姿的增量估计。在后端优化中，设计了一种遵循图优化思想的非线性优化方法对移动机器人位姿进行优化。同时结合闭环检测机制，提出了一种点云优化算法，用于抑制系统的累积误差，进一步提升位姿与点云的精确性。实验验证了本文所提方法能够迅速、准确地重构出稠密化的三维环境模型。

Visual-SLAM for mobile robot based on graph optimization

In the present research on the visual SLAM (Simultaneous Localization and Mapping)of mobile robot, some defects as inferior real-timeness, low accuracy and hard to densified mapping exist, therefore, the paper proposed a real-time SLAM algorithm based on RGB-D data. In the front-end processing of the algorithm, the ORB feature detection with better robustness and real-timeness was adopted. RANSAC algorithm was utilized to get rid of the possible mismatch points and complete the initial match. For the obtained inner point, PNP solution was carried out for using as the increment estimate of the adjacent pose of robot. In the rear-end optimization, a nonlinear optimization method obeying image optimization thought was used for optimizing the pose of a moving robot. In addition, in combination with the closed-loop detection mechanism, a point cloud optimization algorithm was proposed to suppress the cumulative error of the system and further improve the accuracy of pose and point cloud. The experimental results show that the proposed method can reconstruct the dense 3D environment model quickly and accurately.