Bearings-only localization and mapping.

摘要

In many applications, mobile robots must be able to localize themselves with respect to environments which are not known a priori in order to navigate and accomplish tasks. This means that the robot must be able to build a map of an unknown environment while simultaneously localizing itself within that map. The so called Simultaneous Localization and Mapping or SLAM problem is a formulation of this requirement, and has been the subject of a considerable amount of robotics research in the last decade.; This thesis looks at the problem of localization and mapping when the only information available to the robot is measurements of relative motion and bearings to features. The relative motion sensor measures displacement from one time to the next through some means such as inertial measurement or odometry, as opposed to externally referenced position measurements like compass or GPS. The bearing sensor measures the direction toward features from the robot through a sensor such as an omnidirectional camera, as opposed to bearing and range sensors such as laser rangefinders, sonar, or millimeter wave radar.; A full solution to the bearing-only SLAM problem must take into consideration detecting and identifying features and estimating the location of the features as well as the motion of the robot using the measurements. This thesis focuses on the estimation problem given that feature detection and data association are available. Estimation requires a solution that is fast, accurate, consistent, and robust.; In an applied sense, this dissertation puts forth a methodology for building maps and localizing a mobile robot using odometry and monocular vision. This sensor suite is chosen for its simplicity and generality, and in some sense represents a minimal configuration for localization and mapping.; In a broader sense, the dissertation describes a novel method for state estimation applicable to problems which exhibit particular nonlinearity and sparseness properties. The method relies on deterministic sampling in order to compute sufficient statistics at each time step in a recursive filter. The relationship of the new algorithm to bundle adjustment and Kalman filtering (including some of its variants) is discussed.