An automatic feature point extraction method based on laser vision for robotic multi-layer multi-pass weld seam tracking

Laser vision-based seam tracking has been an important research hotspot in modern welding manufacturing. However, severe noise interference during welding and the complex contour curves of filling welds hinder the development of high-precision seam tracking in multi-layer multi-pass (MLMP) welding. To solve this problem, a point distribution model (PDM) has been implemented to express the laser stripe pattern of MLMP welds. Then, an end-to-end feature point extraction algorithm is proposed. The “coarse-to-fine” positioning strategy achieves global correlation and local constraints. The low-resolution heatmap regression and coordinate offset regression balance the efficiency and precision, where the backbone is improved with attention mechanisms. Furthermore, the soft coordinate loss and the Gaussian mixture model were combined to improve the generalization performance. Based on the model, an automatic ROI extraction method and output points filtering are implemented to complete the whole tracking process. In experiments, the proposed method achieved good tracking performance even under strong noise, with the mean absolute error (MAE) being controlled within 0.3 mm. The feature point extraction method shows advantages in both precision and stability, laying a foundation for advanced robotic MLMP welding production.