Visual-lidar Map Alignment for Change Detection in Infrastructure Applications

Abstract

Current human-based infrastructure inspections face several problems. Firstly, human inspectors face safety risks working in hazardous environments and reaching challenging areas in order to inspect an infrastructure asset fully. Secondly, they are time-consuming and costly, requiring significant resources to perform. Lastly, human inspections are subjective and prone to human error, which causes inconsistent assessments. These challenges limit engineers abilities to make accurate assessments of infrastructure health and can make infrastructure management difficult and costly.

Current research in Simultaneous Localization and Mapping (SLAM) has shown promise when applied to automated infrastructure inspections due to its ability to provide accurate 3D maps of infrastructure assets; allowing objective and efficient inspection of assets using the generated 3D models. For the purpose of inspecting infrastructure assets, repeat visits and monitoring changes to the structure are paramount to determining the health of the asset. When generating a 3D map in GPS denied environments (as is typically the case for many infrastructure assets such as bridges), it is typically with respect to the first sensor measurement, meaning multiple scans of the same area will not be in the same reference frame. Many publicly available SLAM solutions do not provide the ability to automatically align 3D scans of the same area from mapping sessions gathered across time (multi-session mapping). When performing repeat inspections of the same asset, data needs to be merged and integrated seamlessly into a single reference frame for accurate association of regions of interest over time.

The central contributions of this thesis are: the implementation of a generic map alignment algorithm utilizing both visual and lidar data for enhanced robustness, and the collection of an infrastructure centric dataset for repeat inspections. These contributions greatly enhance the usefulness of existing automated infrastructure inspection pipelines by enabling the tracking of changes in infrastructure assets over time, allowing for accurate estimation of the degradation infrastructure assets. These contributions also have a significant impact on the general field of SLAM, by decoupling map alignment from SLAM, researchers have more freedom to explore new SLAM approaches without the necessitation of using approaches that incorporate multi-session capability.

To achieve accurate and robust 3D map alignment, this work assumes that an initial trajectory of each scan has been generated by SLAM and that camera and lidar data is available from the scanning session. The map alignment process utilizes techniques from Visual and Lidar relocalization to align maps in a way that is robust to environment change. The map alignment approach was validated on the custom datasets gathered in the Structures Lab at the University of Waterloo and at the Conestogo Bridge in Kitchener, Ontario.