Multi-vision Based 3D Reconstruction System for Robotic Grinding

Abstract

High-precision grinding process often requires to plan the grinding paths based on the reconstructed models of the workpieces. This paper presents a multi-vision based 3D reconstruction system which uses four industrial cameras to capture RGB images from multiple angles to reconstruct the 3D models of the workpieces to be ground. Feature extraction, feature matching and loop closure correction were applied for reconstructing the rough outline of the workpiece. Then we used depth map estimation and fusion to expand sparse point clouds into denser point clouds to improve the reconstruction performance. Poisson reconstruction was finally used to repair the voids on the surface to get a smooth model. The experimental results show that this system can obtain a complete and accurate 3D point cloud model of the workpiece in preparation for the path planning of the robotic grinding.