Towards robust visual-inertial SLAM

Publication Type:
Thesis
Issue Date:
2024
Full metadata record
In recent decades, the development of autonomous navigation systems for mobile robots has been a key area of research. Among the solutions gaining prominence, the Monocular Visual-Inertial Navigation System (VINS) stands out for its compact size, cost-effectiveness, and robustness, addressing challenges in this domain. Achieving optimal performance within resource constraints requires a delicate balance between computational efficiency and estimation accuracy. Choosing a Visual-Inertial SLAM (VI-SLAM) approach for VINS holds substantial significance, encompassing two primary categories: filtering-based methods and optimization-based methods. These methods offer versatile strategies tailored to specific application needs and resource constraints. In this thesis, Compressed-MSCKF (Comp-MSCKF) is introduced as a filtering-based approach. This method effectively incorporates loop closure constraints for long-term navigation based on MSCKF. It achieves this by partitioning the extensive map into local and global maps, ensuring that the global map is updated whenever the local boundary changes. This approach leads to updates limited to O(N_L^2), where N_L represents the size of the local map—typically smaller than the total number of states N. To further enhance system accuracy and robustness, a novel optimization-based method called Parallax Visual-Inertial SLAM (PVI-SLAM) is then proposed. This approach leverages the parallax angle for feature parametrization, combining feature observations and preintegrated inertial measurement unit (IMU) data to formulate a nonlinear least squares problem. By doing so, it adeptly avoids singularity issues linked to problematic features, enabling PVI-SLAM to outperform VI-SLAM methods using XYZ parametrization. Incorporating Gaussian Process (GP)-based preintegration and using the observation ray as an objective function contribute to additional performance improvements. These enhancements not only address challenges posed by traditional methods but also elevate PVI-SLAM, bestowing it with superior robustness and accuracy. However, the high-dimensional nonlinear optimization problem does not always ensure convergence, and even when it does, reaching the global minimum is not guaranteed. Additionally, it poses a significant computational burden, especially in large-scale scenarios with a very large number of poses. To tackle these challenges, a linear submap joining method using the Linear SLAM framework is proposed. In this approach, local submaps are constructed using the PVI-SLAM method, seamlessly joined through a combination of linear least squares and nonlinear coordinate transformations. This technique aims to enhance computational efficiency and overall system robustness, making it well-suited for challenging and resource-intensive scenarios. A comprehensive series of quantitative analyses was conducted on a range of challenging datasets, validating the effectiveness of the proposed VI-SLAM algorithms.
Please use this identifier to cite or link to this item: