Dynamically Estimating Pose from Invariant Feature Measurements

Authors:   Duncan Baird
File name: baird-robot-eye-thesis.txt

Abstract:

    Complex robotic tasks such as autonomous exploration and grasping demand
    the co-operation of sensors and actuators.  In order to integrate sensor
    measurements and actuator control schemes we must determine the rigid body
    transformations that relate the native co-ordinate frames of these
    devices.  Equivalently, we need to estimate the relative pose of sensors
    and actuators in the system.

    We examine the problem of determining the pose of a robot-mounted
    range-finding camera, and present a class of solutions motivated by the
    idea that mobile camera calibration is best addressed by an ongoing
    dynamic estimation process.  We use range measurements and known robot
    kinematics to provide the estimate of camera pose which is maximally
    consistent with the available data.  Our scheme uses scene features that
    are often present in typical workcell scenes and that are easily and
    reliably extracted.  We develop several formulations of the principle, and
    present experimental results for both simulated and real data sets.