MSeg-SLAM: A Semantic Visual SLAM System for Dynamic Scenes.

Visual Simultaneous Localization and Mapping (VSLAM) technique is crucial for intelligent mobile devices to acquire their current position and pose information. The traditional VSLAM system assumes that all objects in the environment are static by default. But in the physical world, some dynamic objects like humans are unavoidable, which imposes a huge burden on the system. This work proposes a semantic VSLAM system named MSeg-SLAM, which combines the VSLAM system and MSeg semantic segmentation network to reduce the impact of dynamic objects. At the same time, we generate semantic octree maps to optimize the storage space occupied by the dense point cloud map. Compared with the original ORB-SLAM2 system, the absolute trajectory error (ATE) can be reduced by over 93% in high dynamic scenes of the public TUM dataset. The results indicate our system has strong robustness and stability both in datasets and real-world practical applications.