MLC-SLAM: MASK LOOP CLOSING FOR MONOCULAR SLAM

Bo Han∗ and Li Xu∗

Keywords

Dynamic SLAM, monocular SLAM, loop closing

Abstract

Most monocular simultaneous localization and mapping (SLAM) systems achieve good performance under the static environment assumption, while they are unable to deal with dynamic environments. This paper proposes a novel approach, mask loop closing, to handle dynamic objects by defining the static rate, relying more on those static scene parts. The proposed approach introduces semantic information into a visual vocabulary that is trained offline with semantic masks produced by the segmentation neural network. A monocular SLAM system (MLC-SLAM) combining the direct and feature-based modules is built to test the performance of mask loop closing, achieving fast and robust tracking. This system is evaluated on the monocular visual odometry dataset published by technical university of munich (TUM) (TUM mono VO), and results show that MLC-SLAM achieves better performance than the state-of-theart monocular SLAM, in terms of trajectory accuracy and system robustness.

Important Links:



Go Back