Abstract:

Aiming at solving the problem that the accuracy and robustness of the current visual-inertial simultaneous localization and mapping (SLAM) algorithm are not high due to lack of the loop-closing detection, this paper presents a keyframe-based visual-inertial SLAM algorithm for monocular and stereo cameras. The visual-inertial SLAM system provides locally consistent trajectory through a visual-inertial odometry. Then the global consistent trajectory is achieved through loop-closing detection and non-linear optimization in parallel, thus a continuous global map is    created. Moreover, this algorithm has the ability to perform relocations in a map that has been obtained and to       continue the subsequent building. The experiments are carried out using the EuRoC dataset to evaluate the accuracy, relocalization and run time of the algorithm. Experiment results show that, compared with the current visual-inertial SLAM algorithm, the proposed method can reduce error accumulation and the drift, relocate the camera and continue to build the map on the basis of the previous map.

Key words: simultaneous localization and mapping (SLAM), inertial measurement unit (IMU), keyframe, non-linear optimization, loop-closing detection, relocalization

摘要：

针对目前视觉惯性SLAM（simultaneous localization and mapping）算法因缺少闭环检测而造成算法的准确性以及鲁棒性不高的问题，提出了一种适用于立体相机和单目相机的基于关键帧技术的视觉惯性SLAM算法。通过视觉惯性里程计提供局部连续轨迹，通过非线性优化技术和闭环检测技术得到并行的全局连续轨迹，从而建立连续的全局地图。此外，算法具有在已获得的地图中进行重定位，并可以继续进行后续建图的能力。采用EuRoC数据集评价了算法的准确性、重定位能力以及运行时间。实验结果表明，与目前视觉惯性SLAM算法相比，该算法可以减少误差累积，减少漂移，重定位相机位置以及在已构建地图基础上继续构建地图。

关键词: 即时定位与地图构建（SLAM）, 惯性测量单元（IMU）, 关键帧, 非线性优化, 闭环检测, 重定位