动态特征剔除的无人系统视觉/惯性导航方法

To reduce the impact of dynamic environment on the localization accuracy and stability of visual/inertial navigation system, a dynamic feature removal visual/inertial navigation method is proposed in this paper. Based on the VINS framework, this method uses structural similarity as the cost volume to generate an end-to-end network for dynamic regions detection. Symmetric optical flow screening is then performed on the identified dynamic regions to remove non-consistent outliers and further eliminate dynamic features that affect localization. The cost function is constructed by fusing visual and inertial measurements, and the nonlinear optimization method is used to estimate the unmanned system states effectively. The experimental results show that the visual/inertial navigation method with dynamic feature removal has good localization accuracy and stability, the position root mean square error is 0. 081 and 1. 982 m on EuRoC publicly available datasets and real scenario data respectively, which is only 35. 5% and 24. 9% of VINS. This method can provide accurate position information in complex application environment, and has good practical value in the navigation of low-cost unmanned systems.