Coherent Point Drift derived algorithm enhanced with locality preserving matching for point cloud registration of roll formed parts

Abstract

Due to severe deformation, noise, and occlusion, the registration problem of non-rigid point sets in rolling formed metal workpieces poses challenges, and the demand for real-time data storage and registration during the rolling forming process makes this problem even more prominent. This paper proposes an enhanced nonrigid point set registration algorithm based on the Coherent Point Drift (CPD) framework, introducing novel methods to improve accuracy and efficiency. A refined local distance calculation method combining spatial distance has been proposed to improve matching accuracy. In contrast, an optimized shape context method introduces a new driving force criterion to expedite initial registration and reduce subsequent errors. Leveraging the Expectation-Maximization (EM) algorithm, the approach iteratively solves point correspondences, demonstrating robustness in handling complex scenarios like non-rigid deformation and noise. Experimental validation using real production datasets shows superior accuracy and efficiency over classical algorithms, showcasing a practical solution for non-rigid point set registration challenges in roll forming applications.