| The technology of three-dimensional laser-scanning gives a quick, precise, and untouched solution to reconstruct the surface of complex objects. Because of its special advantage, this technology is widely used in cad reverse engineering, computer vision, pattern analysis, three-dimensional animation, virtual reality, digital medicine, computer game and other fields. To reconstruct the surface of a three-dimensional object, one must get range images from the surface of the object. Confined by eye direction and the shape of an object, range images can not be acquired to describe the object by one scanning. To acquire a complete surface model of an object, one must scan the object from different views firstly, register range images of different views then and finally merge all range images in a unified coordinate. The key in the above process is the registration of range images, which influences the final result and the precision hugely.Now the widely-used registration algorithm is ICP(Iterative Closest Point) algorithm and some improvements based on ICP. The slow convergence and the time-cost computing on the corresponding points are the main disadvantages. ICP algorithm based on the squared distance improves the convergence greatly; but the time-cost computing still exists. Up till now, no complete analysis are given on the convergence and the converging speed of ICP algorithms. The situation now is not satisfying; mostly heuristic strategy must be adopted.This thesis focuses on the registration of range images, give theoretical analysis on the convergence and converging speed of ICP algorithms, points out that the classical ICP algorithm is in fact a gradient-descent method with linear convergence; the improved ICP algorithm proposed by Chen and Medioni is a Gauss-Newton method with square convergence but it is not stable, may diverge and the improved ICP algorithm by Mitra et al. is a stable quasi-Newton method with square convergence. We propose a new algorithm which is not based on ICP. The... |
