Line-Features-Based Pose Estimation Method for the Disc Cutter Holder of Shield Machine

Zhe Xie, Guoli Zhu, Dailin Zhang, Dandan Peng, Jinlong Hu, Yueyu SunSensors 2023 · DOI

TL;DR

This paper introduces a novel line-feature-based pose estimation method for the disc cutter holder of shield machines, crucial for automating replacement. The method addresses challenges like poor illumination and rounded edges by proposing a "rounded edge model" and an adaptive region-growing approach for robust contour extraction, leading to improved accuracy and efficiency in a demanding industrial environment.

What Problem Does This Paper Address?

Automating the replacement of disc cutters on shield machines is a significant challenge in tunneling construction. This process is currently manual, time-consuming, and often hazardous, especially in the high-humidity and high-pressure environments of some shield machines. Accurate visual positioning is key to enabling this automation, but it's hindered by several factors: the limited space for mounting cameras on the replacement robot, the often-polluted and unevenly lit conditions within the tunnel, and the specific geometric features of the disc cutter holder itself, which has prominent line features but lacks distinct corner points, and is prone to mud and defects. Existing vision-based methods struggle with these combined challenges, leading to inaccurate pose estimation.

Methods

The proposed method tackles pose estimation in two main stages: feature acquisition and pose calculation.

1. Feature Acquisition (Line Extraction):

  • Rounded Edge Model: To address the blurring effect of rounded corners on edges, a "rounded edge model" is developed. This model helps to more accurately identify edge points that correspond to the 3D model of the workpiece, even when the image gradient is not sharp. This is achieved by analyzing the grayscale profile of edges, distinguishing between "strongly illuminated" and "weakly illuminated" edges, and using geometric relationships to pinpoint the true edge.
  • Adaptive Region Growing for Contour Extraction: To handle uneven illumination and extract the inner contour of the disc cutter holder, an adaptive threshold region-growing method is employed. This method dynamically adjusts the growth threshold based on the local grayscale average, ensuring better contour separation compared to fixed-threshold methods. The image is downsampled to improve processing speed.
  • Edge Search Box Generation: After contour extraction, a method is proposed to generate a specific "edge search box" for each identified edge. This involves classifying edge points as horizontal or vertical based on gradient differences and then creating a bounding box that is expanded to cover the relevant search area.
  • Preprocessing-Based RANSAC Linear Fitting: The extracted edge points within the search boxes are then processed using a modified RANSAC algorithm. This preprocessing includes "Continuity Processing" (to remove noise and defects), "Linearity Processing" (to filter out irregular segments), and "Co-linear Processing" (to retain only segments that form consistent lines). This pre-filtering significantly improves the efficiency and accuracy of the RANSAC fitting, reducing iteration time and fitting error.

2. Pose Calculation:

  • MinPnL Algorithm: For calculating the 6D pose (position and orientation) of the disc cutter holder, the paper evaluates both PnP (Perspective-n-Point) and PnL (Perspective-n-Line) algorithms. Through simulation, the MinPnL algorithm is found to be more stable and accurate than ASPnP, especially under noisy conditions, making it the preferred choice for this application.

Datasets

The paper does not explicitly mention training datasets. For evaluation, it uses:

  • Simulated Data: A virtual camera setup is used to generate synthetic images of the disc cutter holder with varying levels of Gaussian noise added to line endpoints. This allows for controlled testing of the pose estimation algorithms.
  • Real-World Experimental Data: Experiments are conducted using a real disc cutter holder, a Yaskawa robot arm, a CCD camera, and a Leica laser tracker for ground truth measurements. The experiments involve simulating "clay pollution" by manually blocking parts of the disc cutter holder's contour to test the robustness of the line detection.

Key Insights and Ideas

  • Rounded edges are a significant challenge: Traditional edge detectors struggle with the gradual transitions of rounded corners, leading to inaccurate edge point localization. The proposed "rounded edge model" directly addresses this by incorporating geometric assumptions about how light reflects from these surfaces.
  • Adaptive processing is crucial for illumination: In industrial settings, consistent lighting is rare. The adaptive region-growing method is a key innovation for reliably extracting contours despite varying illumination conditions.
  • Preprocessing RANSAC is more than just outlier removal: The multi-stage preprocessing (continuity, linearity, co-linearity) effectively cleans the data before RANSAC, not just by removing outliers but by identifying and preserving genuine linear features, leading to substantial improvements in speed and accuracy.
  • Line-based pose estimation (PnL) can be more robust than point-based (PnP) in certain scenarios: The simulation results suggest that for this specific problem, where line features are prominent and corners are scarce, PnL methods like MinPnL offer better stability and accuracy, particularly when dealing with imperfect line detections.

Key Results

  • Improved Line Detection: The proposed method successfully extracts complete straight lines even under conditions of no pollution, light pollution, and heavy pollution (simulated by clay). It outperforms traditional methods like CannyLines and EDLines, which struggle to extract complete lines and often produce extraneous ones.
  • Pose Estimation Accuracy: In experimental verification using a laser tracker, the proposed method achieves an attitude measurement accuracy of 0.6° and a displacement measurement accuracy of 0.9 mm. This meets the required accuracy for disc cutter replacement.
  • Efficiency Gains: The preprocessing-based RANSAC linear fitting significantly reduces iteration time (from 38.251 ms to 8.810 ms) and fitting error (from 0.482 to 0.408 pixels) compared to standard RANSAC.
  • Robustness: The MinPnL algorithm demonstrates higher stability and lower error compared to ASPnP in simulations, especially when noise is introduced.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while the method is robust to mud and defects, extremely severe contamination might still pose challenges.
  • The accuracy verification relies on relative pose changes due to the difficulty of obtaining absolute pose measurements in the experimental setup.
  • Future Work:
    • Further improvements in the positional calculation algorithm, such as exploring vanishing point-based methods or enhancing the PnL algorithm.
  • Developing pose estimation methods for multi-rectangular combinations of targets.
  • Investigating ways to further improve measurement accuracy.

Summary updated

BibTeX
@article{xie2023line,
  author = {Zhe Xie and Guoli Zhu and Dailin Zhang and Dandan Peng and Jinlong Hu and Yueyu Sun},
  title = {{Line-Features-Based Pose Estimation Method for the Disc Cutter Holder of Shield Machine}},
  journal = {Sensors},
  year = {2023},
  doi = {10.3390/s23031536},
}
Figure