A 3-Component Inverse Depth Parameterization for Particle Filter SLAM

Abstract : The non-Gaussianity of the depth estimate uncertainty degrades the performance of monocular extended Kalman filter SLAM (EKF-SLAM) systems employing a 3-component Cartesian landmark parameterization, especially in low-parallax configurations. Even particle filter SLAM (PF-SLAM) approaches are affected, as they utilize EKF for estimating the map. The inverse depth parameterization (IDP) alleviates this problem through a redundant representation, but at the price of increased computational complexity. The authors show that such a redundancy does not exist in PF-SLAM, hence the performance advantage of the IDP comes almost without an increase in the computational cost.