Abstract:

Building more detailed 3D maps and estimating more accurate camera poses have always been the goal pursued by Simultaneous Localization And Mapping（SLAM） technology. The above goals are contradictory to real-time requirements, low computational cost, and limited computing resource of SLAM. A novel semi-dense 3D reconstruction method based on the monocular Oriented FAST and Rotated BRIEF-SLAM（ORB-SLAM） by using the line segment features extracted from keyframes is proposed. Specifically, the improved method builds upon ORB-SLAM, which first provides a series of map points and a set of keyframes and their corresponding camera poses information in real-time. A Keyframe Re-Culling（KRC） algorithm is proposed to further reduce redundant keyframes. A line segment extraction algorithm is employed to extract line segments in each keyframe. By adopting a purely geometric constraints method to match 2D line segments from different keyframes to generate a semi-dense 3D scene model. Experimental results show that the novel method runs steadily and reliably, and can quickly perform online 3D reconstruction with low computational cost.

Key words: monocular-vision, semi-dense, 3D reconstruction, Simultaneous Localization And Mapping（SLAM）, keyframes re-culling

摘要：

构建更详细的地图以及估计更精准的相机位姿一直都是同时定位与地图构建（Simultaneous Localization And Mapping，SLAM）技术所追求的目标，但是以上目标与实时性要求、较低的计算代价和受限的计算资源条件是相矛盾的。提出一种在单目ORB-SLAM（Oriented FAST and Rotated BRIEF-SLAM）方法的基础上利用关键帧中提取到的直线特征进行半稠密三维重建的方法。由ORB-SLAM实时提供一组关键帧及其对应的相机位姿信息和一系列地图点，提出一种关键帧再剔除算法进一步减少冗余帧数目，使用直线段提取方法提取各帧中的直线段，使用纯几何约束方法对以上检测得到的直线段进行匹配，生成一个由直线段构成的半稠密三维场景模型。实验结果表明新方法持续稳定的运行，能在低计算代价条件下快速地在线三维重建。

关键词: 单目视觉, 半稠密, 三维重建, 同时定位与地图构建, 关键帧再剔除