We present an algorithm for creating a CAD model of an existing physical object from a scanned point cloud containing badly scanned regions. The algorithm has a linear complexity and can be applied to processing big clouds. For the triangulation of well-scanned regions, a method from the group of greedy triangulation algorithms is used. The possibility and efficiency of use of very simple and fast tests for this method are shown. Triangulation of badly scanned regions is done using a projection onto 2D surface. The projection surface is defined locally for each region on the basis of already triangulated well-scanned regions. The presented algorithm can reconstruct a surface when well-scanned regions are represented by isolated "islands".