SESGMS: A Welding Seam Extraction Method Based on Spherical Gaussian Mixture Sampling

Accurate seam extraction is crucial for optimizing welding robot performance. Current 3D weld detectors often struggle with geometric approximation challenges and fail to adapt effectively across different weld types, leading to missed or redundant detections that require meticulous threshold adjustments. This paper proposes a weld seam detection method for accurate and robust extraction of weld seam points. Firstly, a neighborhood-weighted spherical projection algorithm is proposed to enhance local feature distinctions within point clouds, improving recognition accuracy of weld seam points. Secondly, a novel Local Reference Frame (LRF) algorithm based on spherical projection is designed to ensure rotational isotropy and facilitate feature comparability across diverse point cloud orientations. In LFR, rotational Spherical Gaussian (SG) closed-form is utilized for hybrid sampling, aiming to achieve efficient, stable, and isotropic 2D transformations of point cloud features. Finally, angle-independent feature vectors are generated by kernel density estimation with a nearest neighbor algorithm to assess symmetry, improving the extraction accuracy of weld. Experimental results on three types of workpieces and seven angular weld models demonstrate that our method excels in adaptive weld seam extraction, particularly in accurately identifying weld seam points.