Exploiting Attitude Sensing in Vision-Based Navigation with an Airship

In this thesis, an attitude sensor is used to compensate the rotational camera motion, and to generate virtual images to simulate pure translation motion. This compensation is made possible by a calibration routine which determines the rotation between the camera and inertial sensor frames. The objective is to facilitate vision-based tasks such as navigation over smooth terrain, 3D point cloud registration, and image mosaicing. In the rotation-compensated, pure translation case, homographies are reduced to planar homologies. In the particular case where the ground plane is horizontal, the relative pose between two views can also be recovered by directly nding a rigid transformation to register corresponding scene coordinates. These models perform fairly accurately especially when estimating the vertical component. The results include recovering trajectories of hundreds of meters for an airship UAV and comparison with GPS data, and using a 3D graph optimization SLAM package to minimize the drift in visual odometry after recovering other relative pose constraints.