MISD-SLAM: Multimodal Semantic SLAM for Dynamic Environments

Simultaneous localization and mapping (SLAM) is one of the most essential technologies for mobile robots. Although great progress has been made in field SLAM recent years, there are a number challenges dynamic environments high-level semantic scenes. In this paper, we propose novel multimodal system (MISD-SLAM), which removes objects reconstructs static background with information. MISD-SLAM builds three main processes: instance segmentation, pixels removal, 3D map construction. An segmentation network used to provide knowledge surrounding level. The ORB features located on predefined removed directly. way, effectively reduces impact precise pose estimation. Then, combining multiview geometry constraint K -means clustering algorithm, our undefined but moving pixels. Meanwhile, dense point cloud information reconstructed, recovers without corruptions objects. Finally, evaluate by comparing ORB-SLAM3 state-of-the-art systems TUM RGB-D datasets real-world indoor environments. results indicate that method significantly improves accuracy robustness, especially high-dynamic