摘要
Abstract
In response to the difficulty in constructing continuous correspondence relationships for point clouds at multiple time points during continuous three-dimensional(3D)laser point cloud scene flow registration,this paper first used learnable basis functions to describe point cloud flows in different scenes,establishing paired correspondences for target points with different poses at two time points.Next,a nearest-neighbor search was used to match features,and a robust kernel function and weighted iteration method were introduced to constrain the transformation relationships of corresponding features,improving both the accuracy and robustness of the registration.Finally,for point clouds of the same target at multiple time points,the cyclic consistency of the target point cloud flow was considered,and overall constraints were applied to the transformation matrix,achieving multi-target point cloud registration.The algorithm's performance was validated using various registration datasets.The results show that the proposed method can effectively handle partial occlusions and noise.The optimal accuracy on the human automatic dressing(CAPE)dataset reaches 99.38%,while the optimal accuracy on the noisy Stanford dataset is 98.09%,and the optimized version achieves an accuracy of 96.47%,with a 30%efficiency improvement.This method not only ensures accuracy but also enhances the efficiency of point cloud scene flow registration,showing promising application prospects.关键词
点云配准/点云场景流/非刚性点云/基函数映射Key words
point cloud registration/point cloud scene flow/non-rigid point clouds/basis function mapping分类
天文与地球科学
