EKF-based SLAM fusing heterogeneous landmarks

Visual SLAM (Simultaneous Localization and Mapping from Vision) concerns both the spatial and temporal fusion of sensory data in a map when moving a camera in an unknown environment. This paper concerns the construction of landmarks-based stochastic map, using Extended Kalman Filtering in order to fuse new observations in the map, when considering heterogeneous landmarks. It is evaluated how this combination allows to improve the accuracy both on the map and on the camera localization, depending on the parametrization selected for points and straight lines. It is analyzed using a simulated environment, so knowing perfectly the ground truth, what are the better landmark representations. Experiments on image sequences acquired from a camera mounted on a mobile robot, were already presented: it is detailed here a new front end where segment matching has been improved.