Feature Detection On Point Clouds Via Local Reconstruction

In the problem of feature point extraction in the point cloud, due to the lack of connection relationship between the point data, and the data is often subject to noise and missing data problem, how to quickly and effectively extract the characteristic information of the geometric feature will contribute to the solution to the point cloud denoising and simplified, mesh reconstruction and point cloud data matching, etc.In recent years, scholars have proposed different point cloud feature algorithm, such as multi-scale analysis algorithm through local structures Riemann tree, minimum spanning tree feature line extraction algorithm, the least square fitting surface curvature analysis algorithms and Gaussian normal clustering algorithm. However, these algorithms can not effectively extract sharp features and smoothing feature, In order to solve the problem of existing algorithm, this paper presents a reconstruction feature point extraction algorithm via local reconstruction. First, for each point, a weight measuring the likelihood of a point to be feature is assigned according to covariance analysis on its local neighborhood. By threshold filtering, the initial feature points are detected. Then, in the local neighborhood of each initial feature point, a triangle set is constructed, which has no triangle crossing feature region and effectively reflects the local feature structure.Due to the presence of noisy data, l1normal reconstruction is adopted to obtain valid information about implicit surface. After that, applying the shared nearest neighbor clustering algorithm on the triangles’normals, we can obtain the clusters of points in the local neighborhood. Finally, for points of each cluster, one plane is fitted. The initial feature point is further identified as a true feature point, if it is nearly locating on the intersection of multiple fitting planes. Experimental results show that our method is simple, stable, and can effectively extract the sharp features while preserving some weak features. Our method is insensitive to the size of selected neighborhood and robust to noise to certain extend.