A novel weld seam extraction method with semantic segmentation and point cloud feature for irregular structure workpieces

Abstract

The intricate surfaces of irregular structure workpieces present significant challenges for robotic welding path planning. To facilitate robot welding without teaching and programming, this paper proposes a weld seam extraction method that integrates semantic segmentation with point cloud features. This approach effectively harnesses the RGB-D information captured by an area array structured light camera. Initially, the K-Net model is employed for semantic segmentation of the workpiece using two-dimensional image features. This segmentation facilitates the coarse localization of the weld seam by extracting edges based on the segmented image mask, thus setting the stage for detailed weld seam extraction using point cloud features. Subsequent improvements in the DLP structured light vision imaging process allow for accurate edge point cloud reconstruction and adaptive extraction of ROI based on shape extension. The extraction of weld seam feature points is then performed using the LOBB feature extraction method, followed by a polynomial fitting of the weld seam using a least-squares approach. Experimental results indicate that the maximum error in weld seam extraction is less than 1.2 mm, with a root mean square error of less than 0.7 mm, and the algorithm completes its task in under 5 s, demonstrating the method’s efficiency and precision.