3D-SIFT Point Cloud Registration Method Integrating Curvature Information

Abstract

Aiming at the issues that point cloud registration is limited by original pose low registration efficiency, and poor robustness, a 3D-SIFT point cloud registration method that integrates curvature information is proposed. First, the Gaussian curvature of the point cloud is calculated, and the scale space of the point cloud is weighted based on the Gaussian curvature. A Gaussian pyramid is constructed, and stable extreme points are detected through the difference of Gaussian (DoG). The main direction of the extreme points is determined to get feature points. Next, the feature points are described using the Fast Point Feature Histogram (FPFH) method improved with adaptive radius. The Random Sample Consensus (RANSAC) algorithm is used to preliminarily estimate the transformation matrix of the point cloud. Finally, the improved Iterative Closest Point (ICP) algorithm is used for fine registration to obtain an accurate transformation matrix. Experiments on public point cloud datasets show that compared with the point cloud registration algorithm based on 3D-SIFT and the point cloud registration method based on the improved ISS algorithm, the proposed method has higher registration accuracy while maintaining algorithmic speed.