Robust Real-Time Visual Odometry for Autonomous Ground Vehicles

Estimating the motion of an agent, such as a self-driving vehicle or mobile robot, is an essential requirement for many modern autonomy applications. Real-time and accurate position estimates are essential for navigation, perception and control, especially in previously unknown environments. Using cameras and Visual Odometry (VO) provides an effective way to achieve such motion estimation. Visual odometry is an active area of research in computer vision and mobile robotics communities, as the problem is still a challenging one. In this thesis, a robust real-time feature-based visual odometry algorithm will be presented. The algorithm utilizes a stereo camera which enables estimation in true scale and easy startup of the system. A distinguishing aspect of the developed algorithm is its utilization of a local map consisting of sparse 3D points for tracking and motion estimation. This results in the full history of each feature being utilized for motion estimation. Hence, drift in the ego-motion estimates are greatly reduced, enabling long-term operation over prolonged distances. Furthermore, the algorithm employs Progressive Sample Consensus (PROSAC) in order to increase robustness against outliers. Extensive evaluations on the challenging KITTI and New College datasets are presented. KITTI dataset was collected by a vehicle driving in the city of Karlsruhe in Germany, and represents one of the most commonly used datasets in evaluating self-driving algorithms. The New College dataset was collected by a mobile robot traversing within New College grounds in Oxford. Moreover, experiments on custom data are performed and results are presented.Master of Science in Engineering (MSE)Computer Science, College of Engineering and Computer ScienceUniversity of Michigan-Dearbornhttps://deepblue.lib.umich.edu/bitstream/2027.42/136618/1/aladem_thesis.pdfDescription of aladem_thesis.pdf : Thesi