LA-VIO: LiDAR-Assisted Visual–Inertial Odometry With Structural Feature Analysis

Line and vanishing points, as crucial geometric elements in visual scenes, can significantly enhance the accuracy of point-only visual–inertial odometry (VIO) systems. However, existing methods struggle with inaccurate line depth estimation and unreliable vanishing point extraction under varying lighting and dynamic scenes. These issues remain unaddressed even in LiDAR point fusion approaches, which merely incorporate LiDAR point cloud residuals into the system rather than using LiDAR data to optimize line and vanishing point features. To address this, this article proposes a LiDAR-assisted VIO method that targets these geometric feature challenges: it associates image line depth with LiDAR point cloud depth, uses LiDAR-extracted structural plane normals to enhance the reliability of vanishing point extraction, and integrates these optimized line and vanishing point factors into a factor graph for joint pose optimization. Extensive experiments conducted on the KITTI benchmark, M2DGR datasets, and real-world scenarios demonstrate that using LiDAR point clouds to optimize line depth estimation and enhance the reliability of vanishing point extraction, as opposed to merely incorporating LiDAR point cloud residuals, effectively enhances the performance of point-only VIO systems.