Weld identification and robot path planning algorithm based on stereo vision and YOLO deep learning framework

In order to realize the teaching-free path planning of robot welding, an efficient and stable weld intelligent identification algorithm was developed by combining deep learning and point cloud processing technology. Firstly, an industrial-grade 3D camera with an eye-to-hand identification was used to obtain the two-dimensional image and 3D point cloud model around the weldment, and the pre-trained YOLOv8 target detection model was used to identify the ROI(Region of Interest) area where the weldment is located. The model identification accuracy is 99.5%, thereby realizing the rapid removal of background point clouds. Based on point cloud processing algorithms such as RANSAC plane fitting and Euclidean clustering, the three-dimensional point cloud in the ROI area was finely identified for the spatial position of the weld seam. Finally, the trajectory was converted into welding trajectories in the robotic user coordinate system with the coordinate transformation matrix. The verification experimental results show that the algorithm developed in this paper can realize the automatic identification of randomly placed weld seams and welding robot path planning, and the generated trajectory has a similar effect to the manual teaching trajectory, with a deviation of within 0.5 mm.