PL-LVI: A LiDAR-Visual-Inertial SLAM System Integrating Visual Point-Line Features

Line features provide more structural information than point features, particularly in low-texture, man-made environments. In this paper, we present PL-LVI, a monocular LiDARvisual-inertial simultaneous localization and mapping (SLAM) system that tightly integrates LiDAR point features with visual line features across all stages of the process. Line features are initially detected by the visual-inertial odometry (VIO), with depth information derived from LiDAR. When feature association fails, triangulation is employed to estimate the missing depth. We then model the point-line reprojection error and optimize the VIO output by minimizing residuals, which serve as the initial guess for LiDAR-inertial odometry (LIO) point cloud registration. Furthermore, we introduce a novel global descriptor that combines the power of point and line features for place recognition, which enhances loop closure detection. Our method is thoroughly evaluated on multimodal datasets, encompassing both indoor and outdoor scenarios, and compared with state-of-the-art techniques. The results demonstrate that PL-LVI significantly improves localization accuracy, global consistency, and robustness in complex environments.