Edge and Corner Detection for Unorganized 3D Point Clouds with Application to Robotic Welding

Abdullah Al Mamun, Syeda Mariam Ahmed, Yan Zhi Tan, Chee-Meng Chew, Abdullah Al Mamun 0002, Fook Seng Wong

TL;DR

This paper introduces novel algorithms for detecting edges and corners in unorganized 3D point clouds, crucial for tasks like robotic welding. Their edge detection method leverages local symmetry and adaptive density, while their corner detection clusters curvature vectors. These methods achieve higher accuracy than existing approaches and enable direct weld seam generation from point clouds, bypassing traditional CAD models.

What Problem Does This Paper Address?

Robotic welding, a rapidly growing application, relies heavily on identifying precise welding paths. Traditionally, this involves using 3D CAD models for offline planning, which can be inflexible and difficult to adapt to unseen or complex welding tasks. While 3D point clouds offer richer information than 2D images, extracting reliable edge and corner features from them is challenging. Existing methods for edge detection in unorganized point clouds often involve computationally expensive mesh reconstruction or graph-based analysis, or they struggle with unseen data. Similarly, 3D corner detection methods often require organized structures or depth images and may not accurately pinpoint edges. This paper aims to bridge this gap by proposing robust and efficient edge and corner detection algorithms specifically designed for unorganized 3D point clouds, with a direct application to automating weld seam generation.

Methods

The paper proposes two main novel algorithms:

  • Edge Detection: This method is based on the mean shift algorithm and evaluates the local symmetry around a query point. Instead of finding density modes, it measures the shift in the centroid of neighboring points. If the shift is significant relative to the local point density (measured by the distance to the nearest neighbor), the point is classified as an edge. This approach is notable for being independent of surface normals, leading to higher accuracy.
  • Corner Detection: This algorithm builds upon the edge detection by first identifying edge points. It then clusters the curvature vectors of these edge points. The key idea is that at a corner, curvature vectors from different surfaces meeting at that point will exhibit distinct orientations. The algorithm analyzes the geometric statistics of these clusters, specifically the angular variation between curvature vectors and the size of the clusters, to classify points as corners. This method focuses on edge points, reducing computational load, and uses curvature vector clustering to identify corners, distinguishing it from methods that rely on cluster counts.

Datasets

  • RGB-D Semantic Segmentation Dataset [20]: Used for evaluating the edge detection algorithm. This dataset provides 3D meshes and YAML files for various scenes containing common grocery products.
  • ShapeNet Dataset [14]: Specifically, 3D models of washers from this dataset were used to evaluate the corner detection algorithm.
  • Panel Workpiece: A 3D point cloud of a panel workpiece, captured using an Asus Xtion sensor, was used to demonstrate the application of the proposed algorithms for automatic weld seam detection.

Key Insights and Ideas

  • Local Symmetry for Edge Detection: The insight that local symmetry (or lack thereof) in a neighborhood is a strong indicator of an edge is fundamental. This avoids the need for explicit normal estimation, which can be noisy or computationally intensive.
  • Curvature Vector Clustering for Corners: The idea that corners represent points where multiple distinct surfaces (and thus distinct curvature vector orientations) converge is powerful. Clustering these vectors and analyzing their geometric properties provides a robust way to identify these convergence points.
  • Direct Point Cloud to Weld Seam: The most significant insight is the demonstration that robust edge and corner detection directly from point clouds can eliminate the need for traditional CAD models in robotic welding path planning, paving the way for more adaptable automation.
  • Adaptive Density Handling: The use of an adaptive density-based threshold for edge detection is crucial for handling the varying point densities common in real-world 3D scans.

Key Results

  • Edge Detection: The proposed Mean Shift (MS) based edge detector consistently achieved higher precision and recall compared to state-of-the-art algorithms (EV, D+HC, D+SC) on the RGB-D semantic segmentation dataset. Visual analysis showed the MS method produced more precise edges closely matching ground truth.
  • Corner Detection: The algorithm demonstrated high accuracy on 3D washer models. For the panel workpiece application, the proposed corner detector was shown to detect desired corners more accurately than Harris 3D, which is crucial for weld seam generation.
  • Computational Performance: While not explicitly optimized for speed, the paper notes that the proposed algorithms are applicable to unorganized point clouds. For the panel workpiece, the proposed corner detector achieved a precision-recall curve comparable to Harris 3D, with the proposed method being more precise in generating fewer false positives.

Limitations and Future Work

  • Computational Speed: The paper acknowledges that while the algorithms are effective, the current implementation (in MATLAB) is not highly optimized for speed. Future work aims to improve this by implementing the algorithms in C++.
  • Parameter Tuning: Like most algorithms, the proposed methods have parameters (e.g., k, λ, ɛ, θ₁, θ₂) that require tuning. While the paper provides guidance and shows good performance across ranges, optimal tuning for diverse datasets might still be necessary.
  • Corner Type Specificity: The paper suggests that the proposed approach can be used to select specific types of corners more accurately than Harris 3D, implying potential for further refinement in distinguishing between different corner types.
  • Weld Seam Generation: While the paper demonstrates the detection of edges and corners for weld seam generation, the full automation of generating the path itself from these detections is an area for further development.

Summary updated

Abstract

In this paper, we propose novel edge and corner detection algorithms for unorganized point clouds. Our edge detection method evaluates symmetry in a local neighborhood and uses an adaptive density based threshold to differentiate 3D edge points. We extend this algorithm to propose a novel corner detector that clusters curvature vectors and uses their geometrical statistics to classify a point as corner. We perform rigorous evaluation of the algorithms on RGB-D semantic segmentation and 3D washer models from the ShapeNet dataset and report higher precision and recall scores. Finally, we also demonstrate how our edge and corner detectors can be used as a novel approach towards automatic weld seam detection for robotic welding. We propose to generate weld seams directly from a point cloud as opposed to using 3D models for offline planning of welding paths. For this application, we show a comparison between Harris 3D and our proposed approach on a panel workpiece.
BibTeX
@inproceedings{mamun2018edge,
  author = {Abdullah Al Mamun and Syeda Mariam Ahmed and Yan Zhi Tan and Chee-Meng Chew and Abdullah Al Mamun and Fook Seng Wong},
  title = {{Edge and Corner Detection for Unorganized 3D Point Clouds with Application to Robotic Welding}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2018},
  doi = {10.1109/iros.2018.8593910},
}
Figure