LiDAR-Aided Visual-Inertial Odometry Using Line and Plane Features for Ground Vehicles

With the conventional classification as edge and planar features, LiDAR point cloud tends to support visual-based odometry by focusing on visual point features depth estimation, while ignoring high dimensional visual features, i.e. line and plane. This paper proposes a novel light-weight visual-inertial odometry for ground vehicles and aerial vehicles with the help of a small portion of LiDAR measurements, which establishes correspondence between visual line as well as plane features and LiDAR point cloud. Specifically, proposed pipeline recovers depth of vertical and ground line via fitting points and line triming, which can avoid estimated depth drift generated by visual line triangulation. Furthermore, statistical information grid (STING) structure is adopted to detect plane using undistorted LiDAR points, while screened 3D mesh produced by 2D Delaunay triangulation are applied to determine correspondence between point as well as line features and plane. This strategy not only makes it more efficient to detect accurate surface but also avoids mis-assignment of features to plane. Both public dataset and man-mad data are implemented to verify progressiveness of proposed pipeline through comparison with state-of-the-art algorithm and ablation study.