Non-Linear Factor Recovery for Visual-Inertial SLAM

Abstract

This paper proposes to use nonlinear factors to recover odometry information and use it in the optimization of global consistent map construction. In the front-end, the pixel matching is carried out by the direct method assisted with IMU information. Then the reprojection error and IMU error are minimized to obtain the initial pose estimation of robot. In the back-end, we use a fix-size optimization window to optimize mapping. When new frames are added, we marginalize the old state. We use a set of nonlinear factors to approximate the marginal distribution, and combine it with loop-closing constraints to construct a globally consistent map. Finally, the performance of the system is verified on the open dataset EuRoC, and conduct experiments in a real environment. The results show that the method improves the accuracy and robustness of mapping.