Surface reconstruction of 3D objects using local moving least squares and K-D trees

In computer vision, surface reconstruction from the point cloud of 3D scanners is desirable to create solid models of the scanned objects. However, noise, outliers, holes, and redundant samples are unavoidable from real sampling data. In this research, a combination of methods are proposed to effectively deal with these problems. First K-D trees are used to stored the unorganized data. The k-nearest neighbors algorithm, robust local noise scale estimation, and bilateral filter are applied subsequently to find planes, delete outliers, and reduce noise. Then hole filling is performed by hole detection and the local moving least squares reconstruction method. Finally, parameterization mapping is employed before generating meshes from the Delaunay Triangulation. Experiments were conducted to show satisfactory performance of our method.