结合分区和改进ICP的三维颅骨自动配准算法

Three Dimensional Skull Registration Automatically Based on Partition and Improved ICP

颅骨配准是统计颅面复原过程的重要步骤之一.在建立颅骨数据库以及后续的相似性颅骨检索中,都需要进行颅骨配准.针对现有的颅骨配准方法准确度和效率不高造成复原效果不理想的问题,本文提出一种结合分区和改进迭代最近点算法(ICP)的三维颅骨自动配准算法.首先根据Voronoi图对颅骨进行区域划分,计算得到每个区域的质心,并作为区域配准的基本单位,根据Euclidean距离和k近邻(KNN)算法实现区域匹配;然后对每对匹配区域运用随机抽样一致算法(RANSAC)选出四对共面的匹配点对,并进行变换矩阵和最大化公共点集(LCP)的求解,根据LCP值得到最优变换矩阵,组合所有区域对的最优变换矩阵求得全局最优变换矩阵,完成初始配准;最后,在ICP算法中设置动态估计(Destimation)来有效剔除误匹配点对,以均方根误差(RMSE)作为配准误差,完成精确配准.实验结果表明,本文算法与基于区域中稀疏ICP算法和基于曲率图中的经典ICP算法对比,迭代收敛性更好,配准准确度有明显的提高,配准的时间复杂度显著降低.

Skull registration is an important technique for statistical craniofacial reconstruction.Skull registration is needed in the establishment of skull database and the subsequent similarity retrieval.Accuracy and efficiency of the existing skull registration methods are not satisfactory,and then influence the facial reconstruction.A three dimensional skull registration automatically based on partition and improved Iterated Closest Point （ICP） is presented in this paper.Firstly,we use the Voronoi diagram to divide the skull into regions,then the centroid of each region is calculated and is regarded as the basic unit of region registration.The initial region registration is taken by computing the Euclidean distance and k-nearest neighbor.Secondly,the Random Sample Consensus （RANSAC） is used to search the four coplanar matching pairs,and also is used to solve the transformation matrix and Largest Common Pointset （LCP）,the global optimal transformation matrix is computed by computing all pairs′ transformation matrix based on the LCP to finish the rough registration.Finally,an improved ICP is presented.A dynamic estimation（Destimation） is introduced to eliminate mismatch points,and the Root Mean Squared Error（RMSE）is regarded as the registration error.The experiments show that interactive convergence is much better,Compared with the sparse ICP algorithm and the traditional ICP algorithm,The accuracy of registration is obviously improved,and the time complexity of registration is significantly decreased.