A 3D Point Cloud Reconstruction Method

Abstract

In order to quickly obtain a precise point cloud model of the target, this paper proposes a 3D point cloud reconstruction method. First, a point cloud acquisition system, which consists of Kinect, turntable, green background frame, and computer, is constructed to acquire the point cloud model of the target at 16 shooting angles. After that, when the coordinate system conversion is finished, the original point cloud model are processed by both a pass-through filter and a voxel filter: the green background and other obstacles in the original point cloud model are removed by the Straight-through filter; and cloud models are further simplified by the voxel filter in order to alleviate the computer's computational pressure, improve the registration speed, and reduce the interference of noise on the registration process. At this point, the number of point clouds is simplified to an acceptable level and the basic features of the target point cloud model are not missing. Finally, this paper proposes a improved ICP algorithm with curvature as the feature descriptor and the KD-Tree as the search mechanism to register the processed point cloud models. In this improved method, the curvature can be calculated relatively simply and not susceptible to noise interference, as well as, the KD-Tree can solve the problem of the inefficiency of the classical ICP algorithm search mechanism and boost the search speed. And, a layer-by-layer registration strategy is adopted in the register to reduce the cumulative error. Several experiments have been carried out on several targets in this paper using the above method, and satisfactory experimental results are obtained.