In this paper, methods for estimating one's own position from a video image are considered. A robust two-stage algorithm for reconstructing the scene structure from its observed video images is proposed. In the proposed algorithm, at the feature extraction and matching stage, a random sample based on the neighborhood graph cuts is used to select the most probable matching feature pairs. At the nonlinear optimization stage, an improved optimization algorithm with an adaptive attenuation coefficient and dynamic adjustment of the trust region is used. Compared with the classical Levenberg-Marquard (LM) algorithm, global and local convergence can be better balanced. To simplify the system's decisions, the Schur complement method is used at the group tuning stage, which allows for a significant reduction in the amount of computation. The experiments confirmed the operability and effectiveness of the proposed algorithm.

Keywords: 3D reconstruction，graph-cut, Structure-from-Motion (SfM)，RANSAC，Bundle Adjustment optimization，Levenberg-Marquardt algorithm，Robust feature matching

This paper is devoted to the construction of a robust visual-inertial odometry system for an unmanned vehicle using binocular cameras and inertial sensors as information sources.The system is based on a modified structure of the VINS-FUSION system. Two types of feature points and matching methods are used to better balance the quantity and quality of tracking points. To filter out incorrect matches of key points, it is proposed to use several different methods. Semantic and geometric information are combined to quickly remove dynamic objects. Keypoints of static objects are used to complement the tracking points. A multi-layer optimization mechanism is proposed to fully utilize all point matchings and improve the accuracy of motion estimation. The experimental results demonstrate the effectiveness of the system.

Keywords: robust visual-inertial odometry, localization, road scene, multi-level optimization mechanism