Towards SLAM-Centric Inspection of Infrastructure

Abstract

The inspection and maintenance of civil infrastructure are essential for ensuring public safety, minimizing economic losses, and extending the lifespan of critical assets such as bridges and parking garages. Traditional inspection methods rely heavily on manual visual assessments, which are often subjective, labor-intensive, and inconsistent. These limitations have driven the development of robotic-aided inspection techniques that leverage mobile robotics, sensor fusion, computer vision, and machine learning to enhance inspection efficiency and accuracy.

Despite advancements in robotic-aided inspection, existing works often focus on isolated components of the inspection process—such as improving data collection or automating defect detection—without providing a complete end-to-end solution. Many approaches utilize robotics to capture 2D images for inspection, but these lack spatial context, making it difficult to accurately locate, quantify, and track defects over multiple inspections. Other works extend this by detecting defects within images; however, without a robust 3D representation, defects cannot be precisely geolocated or measured in real-world dimensions, limiting their utility for long-term monitoring. While some studies explore 3D mapping for inspection, the majority rely on image-only Structure-from-Motion, which is known to be unreliable for generating dense and accurate maps, or are restricted to mapping along 2D surfaces, thereby failing to capture the full complexity of infrastructure assets.

This thesis introduces a novel SLAM (Simultaneous Localization and Mapping)-centric framework for robotic infrastructure inspection, addressing these limitations by integrating lidar, cameras, and inertial measurement units (IMUs) into a mobile robotic platform. This system enables precise and repeatable localization, 3D mapping, and automated inspection of infrastructure assets.

Three key challenges that hinder the development of a practical SLAM-centric inspection system are identified and addressed in this work. The first challenge pertains to the design and implementation of SLAM-centric robotic systems. This thesis demonstrates how sensor selection and configuration can be optimized to simultaneously support both highaccuracy SLAM and high-quality inspection data collection. Additionally, it establishes a robotic platform-agnostic design, allowing for flexibility across different infrastructure inspection applications. The second challenge involves precise and reliable calibration of camera-lidar systems, particularly when sensors have non-overlapping fields of view as is the case with the proposed inspection systems. To address this, a novel target-based extrinsic calibration technique is developed, leveraging a motion capture system to achieve high-precision calibration across both sensing modalities. This ensures accurate sensor fusion, yielding geometrically consistent inspection outputs. The third challenge is the development of a complete end-to-end inspection methodology. This research implements state-of-the-art online camera-lidar-IMU SLAM, with an added offline refinement process and a decoupled mapping framework. This approach enables the generation of high-quality 3D maps that are specifically tailored for infrastructure inspection by prioritizing accuracy, density, and low noise in the map. Machine learning-based defect detection is then integrated into the pipeline, coupled with a novel 3D map labeling method that transfers visual and defect information onto the 3D inspection map. Finally, an automated defect quantification and tracking system is introduced, allowing for defects to be monitored across multiple inspection cycles—completing the full end-to-end inspection workflow.

The proposed SLAM-centric inspection system is validated through extensive real-world experiments on infrastructure assets, including a bridge and a parking garage. Results demonstrate that the system generates highly accurate, repeatable, and metrically consistent inspection data, significantly improving upon traditional manual inspection methods. By enabling automated defect detection, precise localization, and long-term defect tracking within a robust 3D mapping framework, this research represents a paradigm shift in infrastructure assessment—transitioning from qualitative visual inspections to scalable, data-driven, and quantitative condition monitoring.

Ultimately, this thesis advances the field of robotic infrastructure inspection by presenting a comprehensive SLAM-centric framework that integrates state-of-the-art sensing, calibration, and mapping techniques. The findings have broad implications for the future of automated infrastructure management, providing a foundation for intelligent inspection systems that can enhance the efficiency, reliability, and safety of civil infrastructure maintenance worldwide.