VINS-Vehicle: A Tightly-Coupled Vehicle Dynamics Extension to Visual-Inertial State Estimator

Abstract

In this paper, we propose VINS-Vehicle, a novel tightly-coupled vehicle dynamics extension to visual-inertial navigation system (VINS) framework. Degenerate motions, such as uniform linear motions or uniform circular motions, which are most common for a ground vehicle, are not observable for a monocular VINS. Therefore, VINS cannot be applied to vehicles, due to difficulties in initialization and low accuracy. To address this limitation, we extend VINS to tightly coupled with pre-integrated high-frequency motion information based on a two degree-of-freedom (DOF) vehicle dynamics model. By loosely aligning the structure from motion (SfM) results, pre-integrated IMU measurements and motion information, the system can be robustly initialized. A tightly-coupled, sliding window optimization method is proposed to obtain an accurate visual-inertial-dynamics odometry result. The experiments show that the system achieved significantly higher positioning accuracy compared with existing VINS methods. Moreover, the proposed method is robust in a texture-less underground parking lot and dynamic outdoor environments.