A Method of Machining Trajectory Planning Based on Multi-View Stereo Vision and Color Regional Growth

To address the challenges of local trajectory planning in robotic manufacturing of curved surface components, this article introduces a method grounded in multi-view stereo reconstruction and color regional growth. Initially, we apply color texture to the surface of a curved component. Subsequently, the improved PatchmatchNet model, which incorporates a high-resolution net and total variation loss, reconstructs high-quality 3D point clouds, achieving 10.42% improvement in accuracy and 10.59% improvement in completeness compared to the baseline. Then, a regional growth point cloud segmentation algorithm based on the hue-saturation-intensity color space and local color distribution similarity is used to segment the color-texture point cloud across three aerospace components with 82.2% Cat.mIoU and 77.1% Ins.mIoU. Following this, point cloud processing and curve fitting are executed on the single-color-texture point cloud, culminating in the derivation of the local processing trajectory feature line. The nonuniform rational B-splines (NURBS) curve fitting generates machining trajectories with R² exceeding 0.94% and 30% lower root mean square error (RMSE) than polynomial and standard B-spline methods. Real machine grinding experiments validate the method's feasibility, demonstrating precise localization and effective material removal along planned trajectories. The proposed approach addresses critical challenges in local grinding of complex curved surfaces, offering significant practical value for aerospace manufacturing.