Tunnel Deformation Inspection via Global Spatial Axis Extraction from 3D Raw Point Cloud

Cheng Yi, Dening Lu, Qian Xie, Jinxuan Xu, Jun Wang

TL;DR

This paper introduces a novel method for accurately inspecting tunnel deformation from raw 3D LiDAR point clouds, even with significant noise and occlusions. By iteratively fitting cross-sections to extract the tunnel's global spatial axis and then segmenting it, the approach provides a robust and detailed deformation analysis crucial for structural safety.

What Problem Does This Paper Address?

Ensuring the structural stability and safety of large-scale tunnels is a critical but challenging task. Traditional inspection methods are time-consuming and may lack the necessary detail. While 3D LiDAR scanning offers a promising alternative for capturing tunnel geometry as point clouds, the raw data often suffers from occlusions, missing areas, and noise. Furthermore, the actual spatial axis of a tunnel can deviate from its design axis, complicating the extraction of accurate cross-sections for deformation analysis. Existing methods often struggle with these imperfections, leading to limited accuracy in assessing the global deformation state of long tunnels. This paper aims to bridge this gap by developing a robust method that can accurately extract tunnel deformation information from imperfect raw point cloud data.

Methods

The proposed method consists of three main stages:

  1. Tunnel Axis Extraction: This is the core innovation. Instead of relying on a predefined design axis, the algorithm iteratively refines the tunnel's spatial axis. It starts with the design axis to generate initial cross-sections. These cross-sections are then fitted with circles, and the centers of these circles are used to approximate a B-Spline curve, which becomes the updated axis. This "circle fitting and B-Spline approximation" process repeats until the change between successive iterations is below a defined threshold, ensuring convergence to the actual tunnel axis. The use of RANSAC for circle fitting makes the process robust to noise and outliers.
  2. Tunnel Segmentation: Once the actual tunnel axis is extracted, the point cloud is segmented. This involves defining tunnel segment rings along the axis and then segmenting these rings into meaningful pieces. This stage leverages geometric features and normal vector analysis to distinguish between structural elements and accessories, enabling more precise analysis.
  3. Section Deformation Analysis: With the segmented tunnel structure, a metric is designed to evaluate the displacement between actual tunnel cross-sections and their designed counterparts. This allows for quantitative assessment of deformation, providing insights into the tunnel's structural health.

Datasets

The paper evaluates its method using real-world LiDAR data from the Shenzhen Metro Line 1 and Line 3 in China. These tunnels are in operation and have been for eight years. The data was collected using a FARO Focus3D 330 laser scanner. The experiments specifically test the method's performance on data with significant occlusions, noise, and outliers, as well as on large-scale datasets (over 5 km and 1 billion points). No new benchmark datasets were introduced.

Key Insights and Ideas

  • Iterative Axis Refinement: The most significant insight is the shift from relying on a fixed design axis to an iterative process that refines the tunnel's spatial axis based on the actual scanned geometry. This makes the method robust to deviations from the design.
  • RANSAC for Robustness: Employing RANSAC for circle fitting in cross-section extraction is crucial for handling noisy LiDAR data, which is a common issue in tunnel environments due to dust, water, and accessories.
  • Geometric Fidelity over Prior Assumptions: The method prioritizes using the raw geometric information from the point cloud rather than making strong assumptions about the tunnel's shape (like perfect circles or ellipses). This allows it to capture more complex and irregular deformations.
  • Segmentation for Detailed Analysis: The segmentation approach, which breaks down the tunnel into smaller, meaningful pieces, is key to enabling detailed deformation analysis at a granular level, identifying localized issues.

Key Results

  • The proposed method demonstrates accurate tunnel axis extraction and cross-section segmentation even in the presence of significant occlusions (Figure 8) and severe noise/outliers (Figure 9).
  • It successfully handles large-scale tunnel data (over 5 km) with both straight and curved sections (Figure 10).
  • Quantitative comparisons (Table 1, Figure 12, Figure 13) show that the proposed method achieves lower RMSE and higher PGP (Proportion of Good Points) compared to traditional circle fitting, ellipse fitting, and a recent piecewise fitting method [20]. For instance, in Figure 8 data, the method achieved an RMSE of 0.006 and PGP90% of 97%, outperforming other methods.
  • The method is shown to be effective in practical applications like section change inspection over time (Figure 14) and adjacent section step measurement (Figure 15), providing credible deformation analysis.

Limitations and Future Work

  • Point Sparsity: The paper acknowledges that point sparsity, caused by factors like instrument limitations and environmental conditions, can lead to inaccurate segmentation and missing key geometric information.
  • Future Work: The authors plan to exploit and integrate upsampling techniques into their framework to address the limitations posed by point sparsity.

Summary updated

BibTeX
@inproceedings{yi2020tunnel,
  author = {Cheng Yi and Dening Lu and Qian Xie and Jinxuan Xu and Jun Wang},
  title = {{Tunnel Deformation Inspection via Global Spatial Axis Extraction from 3D Raw Point Cloud}},
  booktitle = {Sensors},
  year = {2020},
  url = {https://openreview.net/forum?id=xXfsuzqNq8},
}
Figure