基于几何一致性加权的二部图点云配准方法

A bipartite graph point cloud registration method based on weighted geometric consistency

针对迭代最近点(ICP)算法存在配准时间长、收敛易陷入局部最优、初始位姿较差而导致错误匹配等问题,提出一种基于几何一致性加权的二部图配准算法。首先,计算点云内部形状描述子(ISS)特征点及其3DSC特征与快速点直方图(FPFH)特征;然后,利用FPFH特征获取初始候选对应点对,并利用几何一致性确定高置信度基准点对,同时计算任意点对的几何一致性系数;接着,将对应匹配任务定义为一个代价函数用以模拟几何一致性加权的特征点对的全局相似性;最后,通过Kuhn-Munkres(KM)算法优化代价函数获取全局最优对应关系并利用奇异值分解(SVD)完成点云的粗配准。实验结果表明,相较传统粗配准方法,本文算法能在精度相似的情况下,节省20%的运行时间。

Aiming at the problems of iterative closest point(ICP)algorithm,such as long registration time,easy convergence to local optimum,and poor initial position resulting in incorrect matching,a bipartite graph registration algorithm based on geometric consistency weighting is proposed.First,the point cloud intrinsic shape signatures(ISS)feature points and their 3DSC features are calculated together with the fast point feature histogram(FPFH)features.Then,using the FPFH feature,the initial candidate corresponding point pairs are obtained,and the geometric consistency is used to determine the high-confidence reference point pairs,and the geometric consistency coefficient of any point pair is calculated at the same time.Next,the corresponding matching task is defined as a cost function to simulate the global similarity of feature point pairs weighted by geometric consistency.Finally,the Kuhn-Munkres(KM)algorithm is used to optimize the cost function to obtain the global optimal correspondence,and the singular value decomposition(SVD)is used to complete the rough registration of the point cloud.The experimental results show that,compared with the traditional registration method,the proposed algorithm can improve the running time by 20%under the condition of similar accuracy.