Low-Drift RGB-D SLAM with Room Reconstruction Using Scene Understanding

Room reconstruction task is very important for robot motion planning and navigation. Existing indoor dense mapping algorithms are inefficient in cluttered and occlusion environments because the re- constructed building environment consists of unmeaningful plane fragments. In this paper, we present an architecture for online, incremental room reconstruction which combines an accurate RGB-D SLAM and room layout understanding. We proposed an efficient scene understanding method, which detects room's corners to infer the wireframes and layout planes of room from single RGB-D image, even if the parts of the room are occluded. Moreover, the 3D global features (wireframes and layout planes of the building) can also improve the accuracy of state estimation, especially in geometric indoor environments. These 3D global features are treated as global consistent landmarks, it efficiently bounds the trajectory drift with the travel length increasing. On a public ICL-NUIM dataset, our algorithm achieves higher accuracy than other state- of-arts, and it also builds a geometrically meaningful map.