| This thesis addresses the problem of 3D position and orientation (pose) estimation using measurements from a monocular camera and an inertial measurement unit (IMU). While the algorithmic formulation of the problem is generic enough to be applied to any intelligent agent that moves in 3D and possesses the sensor modalities mentioned above, our implementation of the solution is particularly targeted to robots operating in underwater environments. The algorithmic approach used in this work is based on statistical estimators, and in particular the extended Kalman filter (EKF) formulation, which combines measurements from the camera and the IMU into a unique position and orientation estimate, relative to the starting pose of the robot. Aside from estimating the relative 3D trajectory of the robot, the algorithm estimates the 3D structure of the environment. We present implementation trade-offs that affect estimation accuracy versus real-time operation of the system, and we also present an error analysis that describes how errors induced from any component of the system affect the remaining parts. To validate the approach we present extensive experimental results, both in simulation and in datasets of real-world underwater environments accompanied by ground truth, which confirm that this is a viable approach in terms of accuracy. |
