High-precision point cloud registration system of multi-view industrial self-similar workpiece based on super-point space guidance

The demand for 3D information in intelligent manufacturing makes complete point cloud of large workpiece increasingly important in the industrial field. However, due to the limited measurement range, the existing 3D reconstruction methods are diffcult to measure the large workpiece. The self-similar structure of workpiece also results in the low performance of existing 3D registration methods. To address the above problems, the point cloud registration system based on super-point space guidance is proposed by combining fringe projection profilometry (FPP) and point cloud registration technology to register multi-view point clouds of large workpiece. Specifically, to reduce the impact of self-similar structure on registration, we utilize spatial compatibility to partition the point clouds into local super-point space pairs, based on which to guide the multi-scale feature extraction network (MFENet) to mine effective super-point features, then the super-point features with high confidence is selected to estimate the optimal pose matrix. Experimental results show our registration error measured by standard ball is 0.024 mm, and the point cloud of large workpiece we measured reach the accuracy level of laser tracker. In addition, the registration recall of our system at higher accuracy thresholds is 95%, which demonstrates the high reliability of the method for accuracy-critical applications.