Automatic Extraction of Welding Trajectories for V-Groove Welds Based on 3-D Active Vision Imaging

Technical Summary: Automatic Extraction of Welding Trajectories for V-Groove Welds Based on 3-D Active Vision Imaging

Problem Statement
The development of automatic welding technology is critical for addressing rising labor costs and the hazards associated with manual welding. However, traditional welding robots relying on precise position teaching or offline programming suffer from low efficiency, particularly in unstructured environments. A significant challenge lies in extracting welding trajectories from reflective materials where environmental light interference, uneven surface reflections, and laser scattering degrade the quality of visual data. Existing methods, such as fixed thresholding, standard edge detection, and deep learning approaches, often struggle with noise, residual laser stripes, or errors under reflective characteristics, necessitating a more robust solution for planar welds on reflective workpieces.

Methodology
The paper proposes a three-dimensional active vision-based methodology that integrates feature fusion to automatically extract welding trajectories. The process is divided into three primary stages:

1. Saliency Feature Fusion-Based Laser Stripe Segmentation (SFF-LSM):
   To address background reflection interference, the authors developed the SFF-LSM model. This approach combines two distinct saliency maps:

   • Color Distance Saliency Map: Calculated by subtracting the mean RGB channel values from the input image to minimize the influence of background reflections on the laser stripes.
   • Local Saliency Map: Generated by decomposing RGB channels using 2D wavelet decomposition (db5 filter) to capture detailed features across horizontal, vertical, and diagonal directions. These feature maps are reconstructed and linearly fused.
   • Fusion and Thresholding: The color distance and local saliency maps are multiplied to suppress background noise while preserving stripe contours. An iterative thresholding algorithm is then applied to segment the laser stripes accurately from the background.

2. Precise Stripe Center Extraction with Secondary Smoothing:
   Once segmented, the center of the laser stripe is extracted using a Gray-Gravity algorithm to determine initial coordinates. To ensure the continuity and fluidity of the extracted line, a Savitzky-Golay filter is applied for secondary smoothing. This step utilizes least squares polynomial fitting within a sliding window to refine the centerline, addressing the jaggedness often found in template or extremum methods and the curvature errors seen in Steger's algorithm.

3. Automatic Welding Trajectory Extraction:
   The final trajectory is derived from the 3D point cloud data of the weld seam. The method identifies the lowest points of the V-groove weld seam along the X-axis. Outliers are removed by calculating the standard deviation of the distance from the initial point cloud to a fitted line; points exceeding twice the standard deviation are rejected. A Least Squares fitting algorithm is then applied to the remaining valid points to generate the final welding trajectory.

Key Contributions

• Robust Segmentation: The introduction of the SFF-LSM model effectively suppresses background interference and uneven illumination, enabling accurate segmentation of laser stripes on highly reflective surfaces where traditional methods fail.
• Enhanced Centerline Accuracy: The combination of the Gray-Gravity algorithm with Savitzky-Golay smoothing provides a more stable and continuous centerline extraction compared to extremum, template, or standard Steger methods, particularly in areas with high curvature or noise.
• Integrated 3D Trajectory Extraction: The paper presents a complete pipeline from 3D imaging to trajectory fitting, specifically tailored for V-groove welds, demonstrating the ability to filter point cloud data and fit trajectories with high precision.

Experimental Results
Experiments were conducted using a system comprising a CCD camera, a 650nm line laser, and a six-degree-of-freedom robot on V-groove weld specimens.

• Segmentation and Robustness: The SFF-LSM model successfully filtered background noise and laser speckle interference. Centerline extraction tests on flat surfaces and under varying Gaussian noise levels (variances 0.1, 0.2, 0.4) showed the proposed method had the lowest variance (0.2498) and standard deviation (0.4998) compared to Steger, Gray-gravity, Template, and Extreme Value methods.
• Trajectory Accuracy: The maximum fitting errors for the extracted welding trajectory were measured as 0.51 mm in the X direction, 0.40 mm in the Y direction, and 0.29 mm in the Z direction.
• Efficiency: The centerline extraction process was completed in approximately 0.72 seconds, indicating suitability for real-time applications.

Significance
The paper claims that the proposed methodology provides an effective means of overcoming interference caused by reflections and background noise in complex welding scenes. By achieving a maximum trajectory error of less than 0.51 mm, the research lays a theoretical and practical foundation for the feature extraction and tracking of planar weld seams on reflective materials. This contributes to the advancement of automatic welding systems capable of operating in unstructured environments without the need for extensive offline programming or precise position teaching.