Intelligent Visual Seam Tracking System

Welding is an important field of industrial manufacturing research.,welding has a wide range of applications in the connection of metal materials,which is widely used in industrial fields.In these industries,stainless steel as the "blood of industry" is extremely important.However,in practice,due to welding reasons,resulting in partial welding or burning holes in the steel pipe,resulting in the final product safety problems.In response to the above problem,a system based on binocular vision was designed to identify and track the weld seam,and experiments were conducted on the welding of the splice seam of seamless steel pipes.First,the overall external control composition of the system are introduced,including a three-axis motion platform realized by two sliding tables,a welding gun,a binocular vision sensor and a control instrument.A vision sensing system was designed,the embedded hardware was designed,and the structural parameters and effective field of view of the vision system were designed for theoretical analysis.Next,the calibration algorithm of vision system is designed.The calibration function of opencv library is used to calibrate the projection transformation relationship between monocular camera parameters and binocular camera parameters.ORB algorithm is used to achieve stereo matching algorithm to obtain 3D coordinates.The hand-eye calibration algorithm is used to calculate the transformation relationship between the welding gun and the camera.Then,Network based on semantic segmentation was designed,and two semantic segmentation algorithms based on supervised and semi-supervised models were designed to extract the weld target,respectively.Finally,a stainless steel pipe welding test environment was constructed to test each of the two mentioned semantic segmentation algorithms,compare their performance,and then deploy them in a real project.The experiments show that the overall weld deviation is within ±0.1mm,the overall weld line is smooth and the weld quality is good,which verifies the reliability of the overall system design and the overall applicability of the semantic segmentation algorithms for the environment.