About the Event
Localization and mapping is essential for autonomous robots. An accurate map is necessary for efficient path planning and reliable localization is required for navigating through the planned path successfully. This thesis focuses on developing high availability localization and mapping algorithms that provide more accurate estimates for a given amount of resources compared to state-of-the-art algorithms.
Firstly, this thesis presents a real-time Simultaneous Localization and Mapping (SLAM) system, AprilSAM, for solving large scale mapping problems. AprilSAM achieves lower errors than state-of-the-art SLAM algorithms for a given amount of computation time.
Autonomous robots need maps to perform path planning and localize them in a global reference frame. AprilSAM can be applied online to build such maps for new environments, but it takes lots of resources (e.g. time). This limits the availability of algorithms for many real-world applications such as search and rescue in which robots may be called upon to operate in a novel environment. For a never visited indoor environment, this thesis presents FLAG, a factor graph based localization system that provides global positioning based on floor plans. The novelty of the method is that it does not require additional map building as floor plans already exist for indoor localization.
Pose estimation in a prior map is a common online localization method for robot navigation. The accuracy of a localization system directly depends on the quality of the prior map. A high-resolution map provides a detailed representation of the world, but it is susceptible to feature aliasing since a sensors view is noisy and often ambiguous. Such aliasing leads to poor localization performance. To address this issue, this thesis presents MOSS, a machine learning-based map optimization algorithm that produces compact maps supporting more robust localization than state-of-the-art approaches.