| With the rapid development of scanning technology, in recent years, vivid 3d digital model can be obtained by various scanning devices, and it has become one of the main means of obtaining object’s surface information. This technique has been widely used in Computer Graphics, Computer Vision, Metrology, Robotics, Archaeology, etc. The raw date acquired by acquisition device is usually represented by Point Clouds, because of the simplicity and flexibility, it has gradually become common processing object in all kinds of research and engineering application. Analysis and process for point clouds is the key issue of 3D digital visualization technology. The information processing technology of 3D point clouds has become a bottleneck in the current application, the development of processing theory and method for 3D point clouds have become the focus of the current academic research. This thesis has studied the key techniques of point clouds process, including de-noising, registration and grid reconstruction. In summary, this thesis makes the following contributions.Firstly, a de-noising smoothing algorithm based on normal mollification for point clouds is proposed to preserve the sharp features, while eliminating the defects in point clouds, including noise and outliers. The proposed method removes outliers using statistical analysis techniques, estimates the normal and curvature by weighted covariance analysis method, limits the local neighborhood in the similar region of the normal by using the adaptive dynamic neighborhood, on the basis of the trilateral filtering for nomal, adjusts the location of the noise points. Experimental results show that the proposed method not only can effectively remove outliers, but also can preserve the sharp and edge feature while smoothing the small noise, de-noising effect is good.Secondly, a RANSAC initial registration method based on key points description is presented in order to deal with the defects of the repeated calculation. The ideal of 3DSIFT feature detection is applied in the proposed method to narrow the scope of points. A robust FPFH description is adopted to generate effective corresponding point pairs. To find the optimal transformation efficiency, the proposed method is by reducing sample set and setting the minimum distance threshold. Experimental results show that this method can effectively improve the accuracy and stability of the registration, provide a good initial value of accurate registration for the next step.Thirdly, an ICP precise registration method based on neighborhood characteristics is proposed to solve the time-consuming problems for the traditional ICP. The combination of uniform sampling and feature space sampling is adopted to select the feature points. The search method of corresponding points is improved to improve the accuracy of registration. The neighborhood features are applied to reduce the possibility of error matching, and speed up the convergence of iteration closest point. Experimental results show that this method improves the registration precision and speed, overcomes the problem of huge calculation for ICP algorithm, and has higher reliability and stability.Finally, in order to quickly reconstruct triangular mesh surface from point clouds with redundancy and noise, a fast surface reconstruction algorithm is proposed for high noise and overlapping point cloud. The proposed method removes redundancy by using improved down-sampling method to get the uniform sampling surface, by moving least square method(MLS) to obtain a more robust surface normal, at the same time, get the manifold surface, on this basis, a multi-criteria local projection triangulation algorithm is proposed to quickly complete the grid reconstruction of point cloud. Experimental results show that this algorithm can obviously improve the speed in the process of surface reconstruction, in addition, the down-sampling and MLS smoothing strategy can significantly improve the processing ability of point cloud with noise.
|
PDF Full Text Request