A novel RGB-D SLAM algorithm based on points and plane-patches

In this work, we present a novel RGB-D SLAM algorithm. The novelty of the proposed algorithm lies in the use of both feature points and plane patches for pose estimation. A plane patch is defined as a small-sized patch constructed by using a feature point with small curvature. The feature points with small curvature are called plane points. The remaining feature points are classified as either smooth points which represent smooth surfaces or structural points which represent edges or corners of some objects. Two criteria (coplanarity and overlap) are used to match the plane patches from two frames. The proposed algorithm is able to weight various types of feature points in order to improve the robustness of SLAM algorithms in favouring plane patches but not the isolated feature points, and maintain the real-time performance of SLAM algorithms in favouring small plane patches, but not the large size of planes. We evaluate the proposed algorithm on multiple benchmark datasets. A large scale experiment is also presented to show the robustness of our algorithm.