Research on a real-time pose estimation method for a seam tracking system

In the process of computer vision-based seam tracking, although strong noise interference exists such as that arising from arc and splash in the welding process, the tracking effect has been significantly improved by noise reduction, feature point probability estimation and other methods. However, in the process of automatic tracking of a seam tracking system, the robot's pose cannot be adapted to various welding conditions in real time, resulting in decreased weld quality. To enhance the adaptability and real-time estimation of the robot's pose during welding, this paper proposes a real-time pose estimation method for a seam tracking system. In a welding environment with strong noise interference, the real-time pose estimation of the welding workpiece is carried out, and the robot's pose is changed in real time. The pose estimation is realized by building point cloud data, constructing a tool coordinate system in real time and obtaining rotation angles. To accurately acquire the point cloud data, efficient convolution operators (ECO) for tracking and the morphological intersection method integrated with a support vector machine (SVM) are adopted to classify the images with strong noise to better suppress the tracking model drift. The offline tracking test shows that compared with the original tracking algorithm, the proposed method can significantly suppress the peak value of pixel error and reduce its mean value. The welding experiment results show that the proposed method can be adapted to various welding conditions and achieve adaptive and real-time robot pose goals, which improves the welding precision and quality.