Ground Enhanced RGB-D SLAM for Dynamic Environments

Robust pose estimation and map reconstruction are the basic requirements of the robotics autonomous. In this paper, a static ground feature enhanced SLAM system is proposed for dynamic environments with RGB-D sensors. Compared with the typical point-based SLAM, our designed system extra introduce the ground and other plane constraints to solve the dynamic SLAM. In the front-end, the ground as a special plane feature is detected and tracked, which can provide realiable constraint for the pose estimation in dynamic environments. In the back-end, a point-ground based factor graph is constructed and optimized for more accurate map. Moreover, plane structure is exploited to repair the keyframe dynamic regions, new synthesized keyframes are used to reconstruct the static map for long-term applications. Real world dataset tests demonstrate the effectiveness of our proposed system.