Abstract
To address the challenges of redundant point cloud processing and insufficient robustness under complex working conditions in existing teaching-free methods, this study proposes a dual-modal perception framework termed "2D image autonomous recognition and 3D point cloud precise planning", which integrates stereo vision and deep learning. First, an improved U-Net deep learning model is developed, where VGG16 serves as the backbone network and a dual-channel attention module (DAM) is incorporated, achieving robust weld segmentation with a mean intersection over union (mIoU) of 0.887 and an F1-Score of 0.940. Next, the weld centerline is extracted using the Zhang-Suen skeleton refinement algorithm, and weld feature points are obtained through polynomial fitting optimization to establish cross-modal mapping between 2D pixels and 3D point clouds. Finally, a groove feature point extraction algorithm based on improved RANSAC combined with an equal-area weld bead filling strategy is designed to enable multi-layer and multi-bead robot trajectory planning, achieving a mean absolute error (MAE) of 0.238 mm in feature point positioning. Experimental results demonstrate that the method maintains high accuracy under complex working conditions such as noise interference and groove deformation, achieving a system accuracy of 0.208 mm and weld width fluctuation within ±0.15 mm, thereby significantly improving the autonomy and robustness of robot trajectory planning.