Accurate real-time SLAM based on two-step registration and multimodal loop detection

Abstract

To improve the accuracy and efficiency of 3D LiDAR mapping, this paper improves the traditional point cloud registration and loop-detection methods and proposes a new scheme for accurate and real-time simultaneous localization and mapping systems. Cluster constraints are introduced into the front-end point cloud registration, and a two-step point cloud registration algorithm combining normal distribution transform and iterative closest point is proposed to speed up the point cloud registration. The effectiveness of this registration algorithm is fully validated in three sets of point cloud registration experiments. A loop detection process based on a height global descriptor for multimodal fusion is designed at the back end. The point cloud is encoded by this descriptor and fused with camera image information to generate a 1D operator to improve the search efficiency of loop closure frames. The proposed method is extensively evaluated on the KITTI data set and tested in playground periphery, fountain and parking lot environments. The results show that the accuracy of the proposed method outperforms the state-of-the-art LOAM and LeGO-LOAM in all three different scenarios.