利用曲率过滤特征点改进FPFH+ICP点云配准

Improved FPFH+ICP point cloud registration using curvature filtering feature points

针对传统的点特征直方图(FPFH)+迭代最近点(ICP),在粗配准阶段耗时过大的问题,本文提出一种利用局部曲率来筛选掉部分局部区域较为平坦的特征点,来减少使用采样一致性初始配准(SAC-IA)方法时需要计算的特征点。首先使用体素网格化对原始的点云进行下采样,在对下采样后的点云计算出法线之后,利用法线计算出点云的局部曲率,对经过曲率过滤之后的点计算FPFH特征,随后经过SAC-IA粗配准和KD-tree加速ICP精配准得到最终的变换矩阵,完成整个配准过程。经过实验表明,该方法能够在保证最终配准效果几乎不变的情况下,大大缩短配准所耗时间。

Aiming at the problem that traditional point feature histogram(FPFH)+iterative nearest point(ICP)takes too much time in rough registration,this paper proposes a method that uses local curvature to filter out some feature points which are relatively flat in local regions,so as to reduce the need to calculate feature points when using sac-IA method.Firstly,the original point cloud was down-sampled by voxel meshing.After the normals were calculated for the down-sampled point cloud,the local curvature of the point cloud was calculated by the normals,and the FPFH features of the points filtered by curvature were calculated.Then,the final transformation matrix was obtained through SAC-IA rough registration and KD-Tree accelerated ICP precision registration.Complete the registration process.Experiments show that this method can greatly reduce the time of registration without changing the final registration effect.