|M.Sc Student||Aviv Tsahi|
|Subject||Inertial Navigation Aiding Using Multiple View Feature|
|Department||Department of Aerospace Engineering||Supervisors||Professor Pinchas Gurfil|
|Dr. Hector Rotstein|
|Full Thesis text|
Navigation, understood as the problem of computing the location, velocity and orientation of a craft, is one of the main problems in many engineering disciplines. Historically, navigation was limited to rather large undertakings, such as computing the position of a ship in the ocean or an air vehicle flying between waypoints. More recently, the progress in robotics and autonomous platforms has increased the interest in navigation, since it became clear that the ability to navigate is a key property towards autonomous behavior. Finally, the revolution introduced by the Global Positioning System (GPS) has turned navigation into a commodity, and the need to increase availability and reliability is a key problem with potential economic and practical impacts.
In recent years, the use of onboard cameras became popular for a large number of applications. It is thus interesting to try using the data acquired by the onboard camera to aid inertial navigation. A particularly interesting field of research is Relative Vision Aided Navigation (R-NAV), because it does not require any prior information regarding the observed scene. The current research focuses on aiding inertial navigation by using a wide Field of View (FOV) camera in an unknown environment.
The first part derives an Extended Kalman Filter (EKF) formulation of the problem while rigorously verifying that the KF assumptions are valid to the R-VAN problem; for example, the correlation between the land feature estimation error and the camera position error needs to be considered. We note the highly nonlinear nature of the problem due to the camera projection model.
The second part of the research evaluates a popular algorithm, mostly used for 3D scene reconstruction, called Bundle Adjustment (BA). We integrate the inertial navigation into the BA formulation while taking into consideration the inertial navigation error model. We note that the revised BA algorithm is in fact the Maximum A-Posteriori (MAP) estimator for the R-VAN problem, making it the asymptotically optimal solution of the problem.