VISUAL SLAM — MOBILE ROBOT LOCALIZATION WITH ENVIRONMENT MAPPING

In this paper we present an improved implementation of Visual Slam based on the use of the Scale Invariant Features Transform to solve the problem of corrispondences between two views of the same scene. The improvements described in this paper include the use of odometry estimates to predict the location of the acquired points in the next position, and the Method of Kanatani to solve the absolute orientation of the robot. Furthermore the probabilistic algorithm RanSaC is used to provide outliers—-rejection, allowing to keep only reliable triangulated points. Finally a sparse—-map of these elements is updated at every step, exploiting an Octree—- based data structure. Examples of the performance of the algorithm and estimates of its running time are discussed.