UC San Diego

In the fields of VR, AR, and autonomous driving, it is critical to track the accurate location of an agent using cameras. This thesis dives into the problem of using ordered image sequences for localization, known as visual odometry. The lines of research can be categorized into two main group, geometry-based methods and deep learning-based methods. Geometry- based methods have been explored for over a decade, which yield robust real-time prediction in both outdoor and indoor environments. In recent years, deep learning-based methods show the potential to outperform geometry-based methods in localization. However, they are yet to be proved as accurate in variety of scenes.

In this thesis, we first dive into a complete geometry-based pipeline and point out the key factors for a robust system. Second, we design a deep learning-based camera pose estimation pipeline with geometric constraints, which generalizes better than the learning-based baselines under two datasets. In the end, we explore the possibility of enhancing deep learning prediction based on geometric optimization. The thesis plots a road for combining both methods by thorough comparison. By leveraging the advantages of geometry-based and learning-based methods, the future of a robust visual odometry system can be anticipated.