A Cross Structured Light Sensor and Stripe Segmentation Method for Visual Tracking of a Wall Climbing Robot

Zhang Liguo, Liguo Zhang, Jianguo Sun, Guisheng Yin, Jing Zhao, Qilong HanSensors 2015 · DOI

TL;DR

This paper introduces a novel Cross Structured Light (CSL) sensor and a robust laser stripe segmentation algorithm for accurate weld line tracking by wall climbing robots in challenging outdoor environments. The method effectively handles noise and illumination variations, enabling precise 3D reconstruction and navigation for industrial inspection tasks.

What Problem Does This Paper Address?

Inspecting weld lines on large industrial structures like wind turbine towers often requires robots to operate outdoors. This presents significant challenges for vision-based sensors, particularly structured light systems, due to varying sunlight, shadows, and reflections. These environmental factors can severely degrade the quality of laser stripe images, making accurate weld line detection and tracking difficult. Existing methods often struggle with robustness in such noisy conditions, impacting the reliability of automated inspection and robot navigation.

Methods

The authors propose a two-pronged approach:

  • Cross Structured Light (CSL) Sensor: This sensor design projects two orthogonal laser planes, creating a cross-shaped light pattern. This allows for simultaneous detection of both horizontal and vertical weld lines, crucial for navigating intersections. The CSL sensor is designed to be robust to ambient noise.
  • Robust Laser Stripe Segmentation Algorithm: This algorithm comprises three main steps:
  • Preprocessing: An adaptive monochromatic value space is computed using a minimum entropy model to enhance the laser stripe and suppress noise. This involves finding optimal weights for the R, G, and B color channels.
  • Minimum Entropy Deconvolution (MED): This technique is applied to recover and enhance the laser stripe signal, effectively restoring its impulsive nature and improving contrast against the background. MED aims to maximize the kurtosis of the signal.
  • Center Point Localization: Steger's method, a combination of Gaussian filtering and edge detection, is used to accurately extract the center points of the segmented laser stripe.

Datasets

Not discussed in the paper. The experiments appear to use real-world captured video data from a wall climbing robot operating on a wind power tower.

Key Insights and Ideas

  • CSL for Intersections: The cross-shaped laser pattern is a clever way to simultaneously detect both horizontal and vertical weld lines, which is essential for robots navigating complex geometries like T-intersections.
  • Monochromatic Space for Noise Reduction: Transforming the image into a specific monochromatic space (e.g., R-G) proved more effective than using the raw R component or grayscale for noise suppression and signal enhancement.
  • MED for Robustness: Minimum Entropy Deconvolution is a powerful tool for recovering the true laser stripe signal from noisy images, making the segmentation more robust to environmental disturbances. It effectively enhances the "impulsive" nature of the laser stripe.
  • MED vs. Traditional Methods: The MED approach demonstrated superior robustness and accuracy compared to methods like Center of Mass, Linear Approximation, and spline approximations, especially in challenging conditions.

Key Results

  • Accuracy:
    • Camera calibration resulted in RMS measurement errors within 0.120 mm.
  • In field experiments, the absolute measurement error of the weld line was less than 0.3 mm.
  • Robot navigation experiments showed an average tracking error of less than 2 mm.
  • Speed: The proposed algorithm processes a color image (640x480 pixels) in approximately 20 ms on average, making it suitable for real-time applications.
  • Robustness: The system demonstrated good performance in various challenging conditions, including high illumination, varying light intensity, reflections, and shadows, as shown in Figure 18.
  • Comparison: The MED approach outperformed other methods (CM, LA, QA, AA) in terms of average error and robustness, particularly for horizontal laser stripes. While vertical laser stripe segmentation was more challenging, MED still provided smoother results.

Limitations and Future Work

  • Vertical Stripe Segmentation: The paper notes that segmenting vertical laser stripes in the R component space is particularly challenging and time-consuming, making some methods (LA, QA, AA, MED) infeasible for practical use in this specific scenario. However, the MED approach still offered smoother results than LA.
  • Computational Cost: While MED is faster than some other methods, the paper mentions that selecting a smaller region of interest could further increase running speed.
  • Future Work: The paper implicitly suggests that further optimization for vertical stripe segmentation could be beneficial. The general methodology based on MED is presented as applicable to a broader range of industrial measurement problems.

Summary updated

BibTeX
@article{liguo2015cross,
  author = {Zhang Liguo and Liguo Zhang and Jianguo Sun and Guisheng Yin and Jing Zhao and Qilong Han},
  title = {{A Cross Structured Light Sensor and Stripe Segmentation Method for Visual Tracking of a Wall Climbing Robot}},
  journal = {Sensors},
  year = {2015},
  doi = {10.3390/s150613725},
}
Figure