Go straight, turn right: Pose graph reduction through trajectory segmentation using line segments

With better hardware and more efficient graph-SLAM solvers, we are able to solve increasingly large mapping problems. An actual implementation of a mapping problem as a pose graph requires a certain amount of discretization of the information coming from odometry. Such discritizations are either sensor dependent or use a minimum distance travelled heuristic to add poses to the graph. In this work, we explore the question: how much information we can discard and still be able to get a correct map estimate using the pose graph formulation. We approximate the robot trajectory by a sequence of lines leading to a reduced representation of the original pose graph. This reduction is carried out by using an incremental algorithm that adds new poses to the reduced graph when the perpendicular distance for the current estimated line exceeds a threshold. The reduced representation allows us to recover a part of (or the full) graph when needed. This is achieved by exposing the reduced graph to the optimizer but at the same time not discarding the original pose graph. We show the application of our proposed method on real world datasets and illustrate the accuracy and efficiency with which a reduced representation can approximate the original pose graph problem.