Robust RGB-D SLAM in Dynamic Environments Using Geometry and Semantic Information

Enabling dynamic SLAM with cameras is challenging due to the reliance on ego-motion for dynamic object handling, which creates a paradox since accurate ego-motion calculation requires the removal of dynamic features. This paper presents a solution with a geometry-based motion detection module that uses point correlation and semantic data from real-time instance segmentation to identify moving objects. For efficiency, this module operates on keyframes while a template-based method tracks known objects. The system also eliminates invalid map points for consistent mapping and improved localization. Extensive evaluation in public TUM dataset demonstrates that it surpasses other dynamic RGB-D SLAM systems in accuracy and robustness.