Deep learning-enhanced smart ground robotic system for automated structural damage inspection and mapping

摘要

Ground robotic systems are essential for structural inspection, enhancing mobility, efficiency, and safety while minimizing risks in manual inspections. These systems automate 3D mapping and defect assessment in aging. However, current robotic platforms often require the integration of various sensors and complex parameter tuning, raising costs and limiting efficiency. This paper proposes a streamlined unmanned ground vehicle-based inspection platform, integrating only LiDAR and a low-cost monocular camera. Operated via the Robot Operating System, the platform deploys efficient instance segmentation, Simultaneous Localization and Mapping, and fusion algorithms, eliminating complex tuning across environments. A self-attention-enhanced YOLOv7 algorithm is proposed for accurate damage segmentation with limited datasets, while an enhanced KISS-ICP (Keep It Small and Simple-Iterative Closest Point) algorithm is developed to optimize point cloud odometry for efficient mapping and localization. By introducing camera-LiDAR information fusion, the proposed platform achieves structural mapping, damage localization, quantification, and 3D visualization. Laboratory and full-scale bridge tests demonstrated its high accuracy, efficiency, and robustness. • Robust UGV-based automated inspection platform for the assessment of structural damage. • Efficient framework solely uses LiDAR and a camera to achieve damage localization, quantification and visualization. • Improved YOLOv7 for enhanced instance segmentation minimizes need for custom datasets. • An approach derived from the state-of-the-art odometry strategy is proposed for SLAM. • Verification through rigorous laboratory and full-scale bridge experimentations.