Point cloud registration algorithm for enhanced feature constraints on the body-in-white sheet metal parts

Abstract To address the low detection accuracy and high inspection-mold cost of the traditional sheet-metal part detection methods, a fast and accurate non-contact 3D measurement system was designed. For the registration difficulty due to unclear regional features of sheet-metal parts, an ICP algorithm with local boundary neighborhood feature constraints was proposed to fix data slippage and weak convergence during registration. First, the internal point cloud centroid was extracted via bounding-box discretization for downsampling to reduce data stratification. Next, the surface index of downsampled point cloud data was calculated, and the original point cloud boundary-adjacent point set was mapped and extracted. Then, two homologous and heterogeneous point clouds were combined as the input for the point cloud registration algorithm. Considering the distribution and local features of source and target point clouds, the objective function was modified to converge and enhance the algorithm’s focus on feature boundaries. Finally, point cloud data from public datasets and real-world collections were used for verification. The system’s final spatial-position detection accuracy was within ± 0.2 mm, with a mean value of less than 0.05 mm and a standard deviation of less than 0.15 mm, fully meeting the measurement requirements for the body-in-white sheet-metal parts.