Improving the Robustness to the Uncertainty of Initial Depth Estimation in Monocular SLAM

Recently, the inverse depth parameterization has been widely used in monocular simultaneous localization and mapping (SLAM) within the standard extended Kalman filter (EKF) framework. However, the feature depth is not able to be estimated at one observation. In fact, the feature depth and its standard deviation are initialized empirically, which may affect the convergence of the standard EKF. In order to improve the performance, a modified covariance extended Kalman filter (MVEKF) is proposed in this paper. Loop closure tests are performed to compare the proposed method with the standard EKF method and the results show that the MVEKF method is more robust to the uncertainty of the initial depth estimation while the computational complexity remains about the same.