Automotive visual odometry

In the future, autonomous driving will essentially rely on the continuous availability of vehicle localization. Thus, a robust system for ego-pose estimation is of great interest. Such a system has to be composed of redundant subsystems to provide uninterrupted information with high precision. Due to the increasing number of built-in vehicle cameras, the objective of this thesis is the development of an automotive visual dead reckoning system, known as visual odometry to support existing localization frameworks. To identify critical points of visual odometry, we simulate different system-internal calibration errors as well as scenario-based influences and their effects on the pose estimation quality of visual odometry. The scenario-based error sources include the presence of dynamic objects and measurement errors of observed landmarks that are used for the estimation of the ego-motion. Both of these error sources are referred to as outliers. Based on the simulations, we analyze the properties of the most common measure for outlier identification -- the reprojection error. As will be revealed, the reprojection error judges landmarks depending on their respective coordinate. We derive that this leads to a potential loss of nearby landmarks, which are essential for good pose estimation. Based on these findings, we present a method that increases robustness and precision. An adaption of this idea is subsequently presented for outlier detection in monocular systems. We proceed with a detailed description of our developed visual odometry algorithm and its evaluation. Finally, we present a purely vision-based method for visual global localization. This approach achieves to robustly and efficiently localize the car in a 1000 km² map.