Visual Odometry for Autonomous Vehicles

For Autonomous Vehicles (AVs) to take over, they must reach a high level of reliability. Different aspects are to be considered in the development of well-performing AVs. An important one is the development of a robust and accurate localization for AVs. Localization is the act of determining the vehicle position with respect to its environment. Cameras are available on almost all AVs and mobile robots, and research shows the benefits using them in the localization process. Monocular VO is the process of determining the position and orientation of a vehicle using one camera. Visual odometry systems are usually based on epipolar geometry which embed information on the geometric relation between two views of a scene. In the case of planar scenes, such as empty wide roads, the epipolar geometry fails. Homographies which provide the geometric relation between two views of a mostly planar scene can provide better motion estimates in this case. After a review of the literature and state-of-art of Visual Odometry (VO), a homography-based monocular VO system was developed. This system uses \emph{parallax beams} which is a method that allows to recover the essential matrix without having to discard the previously estimated homography. The developed VO system has been tested on the KITTI dataset and on a custom sequence. The results of these tests show that this system can provide reliable localization.