NDF-SLAM: LiDAR SLAM based on neural distance field for registration and loop closure detection

Abstract

LiDAR SLAM (Simultaneous Localization and Mapping) is commonly used by robots in outdoor environments. However, as the scene expands, errors in point cloud alignment gradually increase, requiring loop closure detection to correct them. To address this problem, we propose point cloud alignment, loop closure detection, and a novel loss function based on a neural distance field (NDF). First, we acquire and optimize the NDF from the point cloud, then use the NDF in conjunction with the LM (Levenberg–Marquardt) algorithm to achieve dynamically adjusted point cloud alignment, effectively reducing the cumulative error. Second, we extract the NDF from two frames that may form a loop closure, comparing their similarities using position coding and GICP (Generalized Iterative Closest Point) to achieve reliable loop closure detection, thereby further reducing the error. In cases where loop closure detection cannot be applied to reduce errors, we propose incorporating intensity information as neural features into the NDF and introducing an intensity-based loss function. Furthermore, we incorporate intensity as a constraint in point cloud registration, resulting in more robust registration. We conducted extensive experiments on the KITTI and Newer College datasets, achieving an average absolute trajectory error (ATE) root mean square error (RMSE) of 0.88 m on the KITTI dataset and 0.178 m on the Newer College dataset. Our method demonstrates higher accuracy compared to existing state-of-the-art LiDAR SLAM techniques.