Inertial vision pose estimation using non-linear observers

This thesis considers the problem of obtaining a high quality estimate of pose (position and orientation) from a combination of inertial and vision measurements using low cost sensors. The novelty of this work is in using non-linear observers designed on the Lie group SE(3). This approach results in robust estimators with strong local and almost-global stability results and straightforward gain tuning. A range of sensor models and observer designs are investigated, including • Creating a cascaded pose observer design using component observers for orientation and position, combining pose measurements reconstructed from vision with inertial measurements of angular velocity and linear acceleration. • Combining inertial measurement of angular and linear velocity with pose measurements reconstructed from vision, and designing an observer using a decomposition of SE(3) into separate orientation and position components. • Considering the case where the inertial measurements of angular and linear velocity are corrupted by slowly time varying biases and developing an observer for both pose and velocity measurement bias directly on SE(3)× se(3). • Eliminating the need for a pose measurement reconstruction by designing an observer operating directly on the measurements of landmark bearing from a vision sensor. The bearing measurements are combined with unbiased measurements of linear and angular velocity measurements to obtain an observer for pose evolving on SE(3). Throughout these works, particular attention is given to designing observers suitable for implementation with multiple independent measurement devices. Care is given to avoid coupling of independent measurement noise processes in estimator dynamics. Further, the final observers proposed are suitable for a multi-rate asynchronous implementation, making timely use of measurements from sensors whose measurement rates may differ by over an order of magnitude.