Semi-direct visual SLAM for stereo cameras: system design and validation
Simultaneous Localization and Mapping (SLAM) requires an autonomous vehicle in an unknown environment to learn about the environment, generates a map and localize itself at the same time. To solve this problem, various types of sensors are equipped to gather information. Nowadays, with the development of computer vision, Visual-SLAM (VSLAM) that relies on cameras becomes a major topic. Specifically, stereo cameras can provide additional depth information along with regular RGB information. In this thesis, state-of-the-art VSLAM methodologies are reviewed and evaluated over standard vision benchmarks. Then a novel semi-direct VSLAM system for stereo cameras is proposed. It utilizes direct image alignment for camera pose estimation, and indirect methods to optimize poses and landmarks. The system maintains a sparse point cloud map and allows loop closing and relocalization when tracking is lost. Further experiments validate that it can achieve competitive accuracy with higher efficiency comparing to other VSLAM methods.