Browsing by Author "Muller, Christiaan"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Map point selection for hardware constrained visual simultaneous localisation and mapping.
    (Stellenbosch : Stellenbosch University, 2024-03) Muller, Christiaan; Van Daalen, Corne E. ; Stellenbosch University. Faculty of Engineering. Dept. of Electrical and Electronic Engineering.
    ENGLISH ABSTRACT: Simultaneous localisation and mapping (SLAM) plays a vital role in autonomous robotics. Robotic platforms are often resource constrained, and this limitation necessitates resource efficient SLAM implementations. While sparse visual SLAM algorithms offer reasonable accuracy for modest hardware requirements, these more scalable approaches still face limitations when applied to large-scale and long-term scenarios. Existing SLAM approaches require manual adjustment to maintain a compact map of the environment. This dissertation presents an optimisation-based approach to select a good subset of map points for visual SLAM, which discards redundant map points in an automated fashion. Novel information-theoretic approaches are developed as solutions for this optimisation problem. Existing coverage-based approaches for related robotics problems are also adapted to selecting map points for visual SLAM. The first of the novel information-theoretic approaches developed in this dissertation allows for accurate SLAM estimation by considering the full SLAM estimation problem, which allowed the reduction of the map points by more than 60%. This is in contrast to coverage-based approaches, which do not allow for accurate SLAM estimation. This first information-theoretic approach is too computationally expensive to use online and motivates the development of computationally inexpensive alternatives. To this end, this dissertation also proposes two additional information-theoretic map point selection techniques that approximate the SLAM estimation problem. These approximate approaches allow for similar trajectory accuracy to the original approach while reducing computation times from hours to often less than a second. These approximate approaches enable the practical online application of these techniques to reduce maps, while minimising the impact on SLAM trajectory accuracy. Lastly, this dissertation also presents a combined approach that improves on the approximate information-theoretic approach by including a coverage-based model. This combined approach outperforms all alternative map point selection approaches developed in this dissertation. Map point selection approaches are first evaluated offline using a modified version of an existing visual SLAM algorithm (ORB-SLAM 2). Different map point selection approaches are evaluated on practical data from both the EuRoC and KITTI datasets. A follow-up experiment also demonstrates the application of map point selection approaches online.