Accurate and Robust RGB-D Dense Mapping with Inertial Fusion and Deformation-Graph Optimization

RGB-D dense mapping has become more and more popular, however, when encountering rapid movement or shake, the robustness and accuracy of most RGB-D dense mapping methods are degraded and the generated maps are overlapped or distorted, due to the drift of pose estimation. In this paper, we present a novel RGB-D dense mapping method, which can obtain accurate, robust and global consistency map even in the above complex conditions. Firstly, the improved ORBSLAM method, which tightly-couples RGB-D information and inertial information to estimate the current pose of robot, is firstly introduced for accurate pose estimation rather than traditional frame-to-frame method in most RGB-D dense mapping methods. Besides, the TSDF (Truncated Signed Distance Function) method is used to effectively fuse depth frame into a global model, and to keep the global consistency of the generated map. Furthermore, since the drift error is inevitable, a deformation graph is constructed to minimize the consistent error in global model, to further improve the mapping performance. The performance of the proposed RGB-D dense mapping method was validated by extensive localization and mapping experiments on public datasets and real scene datasets, and it showed strongly accuracy and robustness over other state-of-the-art methods. What's more, the proposed method can achieve real-time performance implemented on GPU.