Computer Vision Based High-Fidelity Mapping and 3D Reconstruction for Civil Infrastructure Inspection

Abstract

Globally, infrastructure is deteriorating due to aging structures and delays in timely and effective rehabilitation. As a result, there has been a growing demand for efficient and scalable methods to assess the condition of civil infrastructure. Traditional inspection practices, which rely heavily on manual visual assessments, are time-consuming, labour-intensive, and prone to human error. As a result, there is growing interest in automating traditional inspection processes. A common approach involves using 3D reconstructions of civil structural components through off-the-shelf Structure-from-Motion (SfM) software to create digital twins for measurement and analysis. However, this method faces several challenges. First, the black-box nature of current SfM software used for 3D reconstruction is not optimized for the visually degenerate surfaces common in civil infrastructure and lacks transparency for error diagnosis in case of a failed reconstruction. Secondly, existing methods often fail to provide end-to-end support for extracting measurements that comply with structural inspection manuals.

This thesis proposes an open-access, end-to-end framework for enabling vision-based structural inspection through high-fidelity 3D reconstructions. The motivation is to address key technical and scientific challenges in adopting computer vision tools for infrastructure assessment by providing field-deployable and standards-compliant solutions that enhance both visualization and quantification of structural conditions. The methodologies developed in this thesis support inspections needed at small, medium and large-scales of structural components.

The first and second parts of this thesis address small-scale inspection. In the first part, high-fidelity reconstructions from smartphone-based LiDAR sensors are utilized to extract concrete surface roughness profiles. Point cloud processing methods are calibrated against existing subjective field tools used by inspectors to ensure compatibility and enable classification of roughness profiles within current inspection frameworks.

The second part addresses deployment challenges by introducing a reconstruction tool that only uses images. This enables small-scale reconstruction in environments where LiDAR or high-end equipment is unavailable. By removing hardware constraints and validating the proposed tools through field deployment, this thesis demonstrates the practical feasibility of an open, vision-based inspection workflow for performing surface roughness measurements.

The third part focuses on medium-scale inspection. An image-based 3D reconstruction pipeline is developed, followed by integration with an interactive segmentation algorithm. The AI-based segmentation method is integrated with 3D models to detect and quantify a common defect, concrete spalling, in accordance with structural inspection standards.

Finally, a large-scale multi-resolution map (MRM) reconstruction workflow is developed for constructing 3D maps with varying resolutions by integrating LiDAR sensor-based 3D maps (coarse resolution) with maps built from images (fine resolution). This method uses a novel image-based localization algorithm to precisely align two maps into a cohesive 3D point cloud. To facilitate MRM, an in-house, cost-effective and portable backpack-based scanner and mapper is designed to collect large-scale colourized LiDAR maps.

Experimental results are presented for each method, including field tests, demonstrating the accuracy and utility of the proposed system for real-world inspections. The major contribution of this work is bridging the gap between academic research and practical implementation in infrastructure inspection, advancing toward a more intelligent, scalable, and accessible inspection paradigm.