摘要
Abstract
In order to overcome the shortcomings that the exiting registration methods are difficult in aligning urban scene point clouds with low overlap and large differences in rotation and/or translation,an accurate registration method for point clouds based on line correspondences matched by a novel line feature descriptor is proposed.First,a novel line feature descriptor based on the FPFH(Fast Point Feature Histogram)and length of the 3D line segment is proposed.Second,the initial corresponding line features are obtained according to the proposed line fea-ture descriptor,using the bidirectional search strategy.Third,the high-quality line correspondences are selected from the initial ones according to the angle constraint between the direction of line features,which are used to cal-culate the 3D rotation matrix.Then,the corresponding point pairs are filtered by the constraint of distance between the midpoints of common perpendicular lines of different lines,to calculate the 3D translation vector.Finally,the points in overlapping region are chosen to refine the transformation matrix.The experiments are conducted on the Bremen datasets,and the results show that,compared with 4PCS(4-Point Congruent Sets)and the 2D line-based registration methods,the proposed algorithm almost improves the accuracy by an order of magnitude,with an aver-age rotation error of 0.33° and an average translation error of 0.13 m,when aligning LiDAR point clouds with low overlap and large differences in rotation and/or translation.关键词
点云配准/同名线元识别/同名点对筛选/线元描述符/几何一致性约束Key words
point clouds registration/line correspondences recognition/corresponding point pairs filtering/line feature descriptor/geometric consistency constraint分类
测绘与仪器
