Abstract:

Feature-based visual SLAM（Simultaneous Localization and Mapping） has some problems such as poor real-time and robustness. In this paper, an improved ORB feature extraction method based on quad-tree is proposed, and a robot RGB-D SLAM algorithm including frontend, backend and map construction is designed. In the frontend, quad-tree is used to extract uniform ORB features, and hamming distances between descriptors are calculated to achieve feature matching. According to the idea of random sampling consistency algorithm, EPNP（Efficient Perspective-N-Point） and iterative closest point method are combined to obtain accurate robotic poses through multiple iterations. To reduce the cumulative error, the keyframe loop closing is adopted to optimize the pose graph based on bundle adjustment so that a globally consistent 3D map is constructed. Experiments on TUM datasets and multi-caterpillar omnidirectional mobile robot are carried out to verify the validity of the presented approach. The results meet the real-time and stability requirements. The feasibility and effectiveness of the algorithm are proved.

Key words: RGB-D Simultaneous Localization and Mapping（SLAM） algorithm, quad-tree, ORB, keyframe, loop closing, bundle adjustment

摘要：

基于特征的视觉同时定位与地图构建（Simultaneous Localization and Mapping，SLAM）存在实时性和鲁棒性差等问题，提出一种改进的基于四叉树的ORB特征提取方法，设计包含前后端及地图构建的机器人RGB-D SLAM算法。在前端使用四叉树方法完成ORB特征的均匀提取，计算描述子间汉明距离实现特征匹配。根据随机采样一致性算法思想，结合EPNP（Efficient Perspective-N-Point）和迭代最近点法求解位姿，获取多次迭代后的准确位姿。采用关键帧进行回环检测，并且基于光速法平差优化位姿图，从而构建全局一致的3D地图，达到减少累积误差的目的。通过TUM数据集和多履带式全向移动机器人进行对比验证，实验结果满足实时性和稳定性要求，证明了算法的可行性和有效性。

关键词: RGB-D SLAM算法, 四叉树, ORB, 关键帧, 回环检测, 光速法平差