An Improvement Algorithm for OctoMap Based on RGB-D SLAM

For indoor mobile robots, building an accurate and compact map is an indispensable prerequisite for different navigation tasks. OctoMap can achieve compact 3D maps, and is suitable for robot navigation. But in the raw OctoMap, there are many sparse outliers, which affect the robot navigation. In this paper, we use RGB-D SLAM system to estimate the pose of robots and build dense 3D point cloud maps in real time. We propose a sparse outliers removal algorithm based on K-Nearest Neighbor and Gaussian distribution to remove sparse outliers of the 3D point cloud maps. And then, we build a more accurate and compact OctoMap from the filtered point cloud maps for the robot navigation. We perform several experiments in a real laboratory environment and some indoor sequences of the TUM RGB-D datasets. Experiments verify that our algorithm can successfully remove the obvious outliers in the OctoMap as well as reduce the memory consumption while preserving the main OctoMap.