Abstract
This paper presents an algorithm for feature point extraction from scanning data of large tubular T-joints (a sub-type of a TKY joint). Extracting such feature points is a vital step for robot path generation in robotic welding. Therefore, fast and reliable feature point extraction is necessary for developing adaptive robotic welding solutions. The algorithm is based on a Convolutional Neural Network (CNN) for detecting feature points in a scanned weld groove, where the scans are done using a laser profile scanner. To facilitate fast and efficient training, we propose a methodology for generating synthetic training data in the computer graphics software Blender using realistic physical properties of objects. Further, an iterative feature point correction procedure is implemented to improve initial feature point results. The algorithm's performance was validated using a real-world dataset acquired from a large tubular T-joint.
