基于改进型3D SIFT正态分布变换算法的点云配准

摘要

Aiming at the low computational efficiency problem of point cloud registration by using three-Dimensional Scale Invariant Feature Transform (3D SIFT) algorithm (based on three-dimensional point cloud),a new three-Dimensional Normal Distribution Transform (3D-NDT) algorithm was proposed based on the improved 3D SIFT algorithm.Firstly,spherical structure was used to substitute the original square structure to construct the new SIFT feature point descriptor,then the original dimensions of 3D SIFT algorithm characteristics description are reduced from 128 to 48.Secondly,3D SIFT feature extraction for three-dimensional points cloud was carried out,the final point cloud matching points and the transformation matrix were obtained by using bidirectional fast approximate nearest neighbor search and random sample consensus algorithm,thus finishing the initial registration.Finally,the NDT algorithm was used to voxel grids point cloud and the probability distribution function was used to accurately register the point cloud.The experimental results show that the improved algorithm is suitable for the registration of point cloud data acquired by KinectV2.0 with different viewpoints,and its accuracy of point cloud registration is good.%针对点云配准中采用三维尺度不变特征转换(3D SIFT)算法(基于三维点云)计算效率低的问题,提出一种基于改进型3D SIFT的三维正态分布变换(3D-NDT)配准算法.该算法首先利用球形代替方形来构造SIFT特征点描述符,并将3D SIFT算法特征描述向量128维降低到了48维;其次,对三维点云进行3D SIFT特征提取并估算特征描述符,使用双向快速近似最近邻搜索和随机采样一致性算法确定最终点云匹配点,求解出变换矩阵,完成初始配准;最后,使用NDT算法体素化点云,并使用概率分布函数对点云精确配准.实验结果表明,改进后的算法适用于由KinectV2.0获取的不同角度的点云数据配准,且其配准精度较高.