MN-SLAM: Multi-networks Visual SLAM for Dynamic and Complicated Environments

Abstract

Simultaneous Localization and Mapping (SLAM) have been proposed for many years, which is widely used in the autonomous robot field. However, there are some troubles in dealing with the outdoor and dynamic environment. We present a Multiple Networks SLAM (MN-SLAM) to cope with those arduous tasks. There are three networks in this system which includes four parts: processing, tracking, local map, and loop closing. Meanwhile, we combine the re-matching mechanism with the moving consistency check to battle with complex environments. Our system was allowed to work in the KITTI dataset and TUM dataset for Stereo and RGB-D sensors. The results show the rotation error and translation error can be improved by this method.