旋转不变特征描述子的点云自动配准方法

Point cloud registration method based on rotation invariant feature descriptor

针对不同视角下获取的三维点云数据,提出一种基于旋转不变特征描述子的点云自动配准方法。该方法通过引入旋转不变特征描述子,可以在不依赖点云初始空间姿态前提下,计算得到待匹配点云间的初始位置关系,完成粗配准;针对配准过程中可能存在的错误匹配,提出了一种刚性变换一致性检验算法,用于实时监督多视角三维重建过程,有效避免误匹配对于重建结果的影响;精配准阶段采用改进的迭代最近点算法完成二次配准。实验结果表明:该算法具有稳定性强、重建精度高的优点,能够满足多视点云配准的要求。

This paper proposes an automatic point clouds registration algorithm based on rotation-invariant local feature descriptor for three-dimensional point cloud data obtained under different views.This algorithm works first by introducing a rotation-invariant local feature descriptor to determine point correspondences;then by estimating initial transformation without any prior available information of initial position and providing a novel rigid transformation consistency verification method designed for possible registration errors occurring in coarse registration process to supervise multi-view point clouds＇ reconstruction in real-time,thus effectively avoiding mismatch impact;and ultimately by completing the fine registration using a variant of iterative closest point algorithm.The experiments show that our method capable of more accurate and robust automatic point clouds registration is adequate for multi-view registration.