DRM-SLAM: Robust Visual SLAM Based on Dynamic Region Mask for Dynamic Scenes

Currently, most visual SLAM is based on a strong assumption of static scenes and has low localization accuracy and weak robustness in dynamic scenes. For dynamic scenes, we propose a robust visual SLAM system based on a dynamic region mask, named DRM-SLAM. In DRM-SLAM, we design a dynamic region mask generation module combining semantic segmentation and geometric constraints, which performs pixel-level dynamic region segmentation on RGB image frames captured by visual cameras. The generated dynamic region mask is used to extract static feature points, rejecting the negative influence of dynamic objects in the environment. We evaluate DRM-SLAM using the dynamic scene sequences of the TUM public dataset, and the results show that DRM-SLAM demonstrates excellent positioning accuracy in dynamic scenes while still maintaining robustness.