Pose graph for improved monocular visual odometry

Abstract

In this paper the monocular visual odometry algorithm augmented with pose graph optimization is presented. The algorithm was tested using five different combinations of feature extractors and descriptors and evaluated using two challenging datasets from KITTI database. The main result of this study is that the implementation of pose graph optimization may lead to reduction of position error ranging between 1.53% to 76.05%. The error reduction depends on a feature type and dataset used.