A robust fusion method for RGB-D SLAM

RGB-D cameras are becoming more and more popular in the areas of Simultaneous Localisation and Mapping (SLAM). Visual Feature matching and dense point cloud ICP are the main methods to estimate camera pose in the existing RGB-D SLAM system. Only using visual information, feature matching method can hardly obtain accurate enough result as the appearance features are sparse. Making use of depth information alone, ICP method cannot avoid the result converging to an incorrect local minimum when the environment has poor 3D geometry. In this paper, we propose a robust fusion method, which acquires accurate and robust pose estimation via combining visual information with depth information. The method can obtain relative transformation by minimizing a robust error function, which integrates matching errors from both visual features and dense point clouds. In addition, a loop closure detection mechanism and a pose graph optimizing method are utilized to estimate the globally consistent pose. Therefore the 3D environment can be well reconstructed. Extensive experiments demonstrate that the method can robustly and accurately estimate the camera trajectory for the RGB-D SLAM system, even in the environment with poor 3D geometry or the area with sparse texture.