Visual prior map assisted monocular location algorithm based on 3D spatial lines

Abstract

When the Visual-Inertial Odometry (VIO) is started, its Inertial Measurement Unit (IMU) lacks acceleration incentive, which will result in poor orientation estimation accuracy during initialization, or even initialization failure. Therefore, a visual priori map-assisted monocular location algorithm based on 3D spatial straight lines is proposed. Firstly, the monocular image data of the surrounding environment were extracted through the Line Segment Detection algorithm (LSD), and high precision 2D line features were selected according to the length of the line and the number of surrounding point features. The 3D spatial lines of the surrounding environment were obtained using the line and surface intersection method. Construct a visual prior map with 3D spatial straight lines. Secondly, the constructed visual prior map is used as the online monocular VIO pose estimation for the global map. Based on the straight-line feature matching algorithm and the 3D space straight line depth information as additional constraints, the 2D straight-line feature in the monocular VIO's current field of vision is matched with the 3D space straight line in the visual prior map. The matching results were used as global constraints to optimize the monocular VIO pose. Tests on EUROC and TUM common data sets show that the 3D spatial straight line based visual prior map can effectively correct the pose during the monocular VIO initialization stage. Compared with the VINS-Mono localization algorithm, this algorithm can effectively improve the pose estimation accuracy during VIO initialization and reduce the overall trajectory positioning error.