.png){kind=icon}
.png){kind=icon}
Return
GPSCO: Global Planar Structural Constraint Optimal-Based Point Cloud Registration Algorithm for Repetitive Structures
.png){kind=icon}
PRE
AI
DOI：10.1109/TGRS.2024.3421607.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
3-D point cloud registration is a prerequisite for scene reconstruction and 3-D object recognition in computer vision and remote sensing. Numerous previous studies have presented a series of point cloud registration algorithms with diverse efficiencies and accuracies. However, registering point clouds with repetitive structures is still challenging. In this study, we propose the global planar structural constraint optimal (GPSCO)-based algorithm, which is specifically designed to handle the registration of repetitive structures. Its novelty lies in establishing the geometric constraint of multiple planes and registering based on the global optimal geometric constraint. The specific algorithm involves clustering the parallel planes into plane groups, estimating matching scores between plane groups, and selecting three corresponding pairs of nonparallel plane groups to form the plane structural constraint. The transformation matrix determined in the case of the optimal structural constraint is taken as the final result. The two terrestrial LiDAR datasets (HS1 and HS2) of real scenes with repetitive structures were collected to evaluate the GPSCO algorithm. Additionally, the GPSCO algorithm is validated on four public benchmark datasets, such as Whu-Park, Whu-Campus, ETH-Hauptgebaude, and ETH-Stairs. The registration results demonstrate that the GPSCO algorithm achieves 95.65%, 86.36%, 100%, 100%, 100%, and 88.89% successful registration rates (SRRs) on the six datasets, respectively, and significantly outperforms the existing methods on HS1 and HS2 with repetitive structures. The corresponding datasets and code are available at [https://github.com/fog223/GPSCO].
Keywords:
Point cloud compression
Feature extraction
Clustering algorithms
Vectors
Three-dimensional displays
Transmission line matrix methods
Benchmark testing
3-D point cloud
datasets
plane constraint
registration
repetitive structures
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W