A practical 3D bearing-only SLAM algorithm

This article presents a bearing only 3D SLAM algorithm which has the same complexity and optimality as the usual extended Kalman filter used in classical SLAM. We especially focus on the landmark initialization process, which relies on visual point features tracked in the sequence of acquired images: a probabilistic approach to estimate their parameters is presented. This induces a particular structure of the filter architecture, in which are memorized a set of past robot poses. Simulations are made to compare the influence of some parameters required by our approach, and results with an indoor robot and an airship are presented.