<p>Accurate localization of weld seams from workpiece surface point clouds using segmentation networks often faces challenges due to the difficulty in preparing adequate training datasets. To mitigate this limitation, a physics-model-driven approach for generating weld seam point clouds that replicate the geometric and noise characteristics of laser structured-light measurement is proposed. This approach utilizes weld bevel geometry and spatial trajectories of feature points as prior knowledge. It employs NURBS curve fitting to uniformly discretize feature point trajectories, while incorporating Gaussian noise based on laser vision imaging principles to generate high-fidelity, automatically annotatable weld point cloud datasets. Building upon this foundation, a feature point extraction method integrating local oriented bounding boxes (LOBB) with concavity-convexity analysis is proposed. This enables accurate identification of critical edge points on welds while effectively mitigating misidentification caused by surface protrusions on lap welds. Experimental results demonstrate that the model’s performance degradation in real-world scenarios does not exceed 14%, with average path error for three weld types remaining below 0.2 mm. This validates the method’s effectiveness and engineering feasibility.</p>