Computational Comparison of Different Optimization Methods for Visual/Inertial Integrated Positioning

The visual/inertial integrated positioning system (VINS) is highly expected for autonomous systems due to the complementariness between the visual and inertial measurement unit (IMU) measurements. In general, the VINS is achieved by combining the visual factor and the IMU factors into an optimization problem which is then solved iteratively. However, due to the nonlinearity of the visual measurement model, the nonlinear optimization method is critical, such as the Gauss-Newton, Levenberg-Marquardt, and Dogleg methods. In particular, it is still unclear how these nonlinear optimization methods differ from each other in terms of computational accuracy and efficiency. To fill this gap, this paper proposes to perform a detailed comparison of the listed three methods in the VINS using the open EuRoC MAV dataset collected onboard a Micro Aerial Vehicle (MAV).