A Fast Feature Extraction-Based Method for Globally Accurate Registration of Multi-View Point Clouds

Abstract

In this work, a new multi-view point cloud registration method is proposed, which has both pairwise registration and global registration accuracy, and solves the problem of large cumulative errors in traditional multi-view registration methods. It begins with the application of an enhanced RANSAC (Random Sample Consensus) algorithm for coarse registration, followed by the design of a new ICP algorithm that utilizes KD-trees (K-Dimensional-trees) to further refine the registration. The improved RANSAC algorithm leverages the SHOT (Signatures of Histograms) feature descriptor to enhance the removal of corresponding point pairs during the coarse registration, thereby improving the accuracy of point cloud data positions. Additionally, SDP (semi-positive definite planning) is incorporated to optimize the global point cloud post-ICP (Iterative Closest Point) registration, thereby enhancing the overall robustness and accuracy of the registration process. This method aims to overcome the challenges associated with traditional ICP algorithms, such as high registration position requirements and slow convergence speed, while also offering optimization for multi-view point cloud registration. Furthermore, it provides a comprehensive solution for aligning source and target point clouds, which holds potential applications in reverse engineering, 3D reconstruction, and industrial inspection.