Optimization-Based, Simplified Stereo Visual-inertial Odometry With High-Accuracy Initialization

This paper proposes a robust and highly efficient feature-based visual-inertial odometry (VIO) approach. In order to save the computational resource, a simplified stereo visual model is applied to reduce the dimension of visual measurements. Moreover, the speed of feature matching is improved by using prior information from the inertial sensor. And through the marginalization, optimization is limited in two sliding windows, which can meet the need for the real-time application. In addition, a high accuracy initialization method and the homogeneous extraction of image features are introduced to ensure excellent pose tracking. The proposed VIO system is implemented on open datasets to show its merit compared with other state-of-the-art algorithms. In addition, we also perform this system on a low-cost stereo visual-inertial sensor and validate its practicability and superiority. Furthermore, the comparative experiment shows that the proposed algorithm has a higher accuracy than the monocular VIO and a shorter running time than the stereo VIO.