关键词: 点云, 配准, 单木分割, 特征提取, 图匹配

Abstract:

Objective: The difficulties and the large registration error of the 3D point cloud registration in forested area present a challenge. Taking different types of forest point cloud data from different perspectives and platforms (ground-based and airborne) as input, and aiming at high-precision fusion of forest point cloud data, an automatic marker-less registration algorithm of point cloud for multi-types of data under forest semantic environment was proposed, focusing on point-pair matching with the single wood object as semantic feature. Method: Firstly, according to the data characteristics of different types of point clouds, the tree positions are extracted respectively. For the side view (ground-based) point cloud, based on the deviation between the main direction dispersion and the vertical angle of the main direction of the point cloud, non-trunk point cloud is eliminated, and the connected component segmentation method of voxel-based partition is used to identify the trunk of individual tree, and the place with the maximum point density of voxel vertical distribution of the single tree trunk point cloud is determined as the planar position of the tree. For the top view (airborne) point cloud, the canopy height model (CHM) is extracted, and then the marker controlled watershed method is used to segment CHM to extract individual trees and the canopy vertex is identified as the planar position of individual trees. Then, taking the extracted positions of individual trees as feature points, the registration matrix is obtained based on Laplace spectral map matching theory, and the 4-DoF (degree of freedom) point cloud coarse registration is completed. Finally, the side view point clouds are precisely registered by using the main trunk matching and the side view and top view point clouds are precisely aligned by using the global iterative closest points (ICP) method. Result: The experimental result show that the registration accuracy between point cloud obtained from ground perspective views is better than that of ground side view and airborne top views point clouds. The average coarse registration error between point clouds from both ground perspective was 0.24 m, and the average error of precision registration was 0.08 m. The average error of coarse registration between side view and top views point cloud was 1.07 m, and the average error of global ICP is 0.44 m. After airborne point cloud and side-looking point cloud are matched through the main trunk, the average error is further improved to 0.36 m. Conclusion: This study followed from coarse to fine registration routine, integrating many kinds of individual tree detection algorithms designed for ground-based and airborne point cloud. We design a framework for multiple source types of point cloud registration, and the experimental result prove the feasibility of the scheme, which is suitable for various forested environment. The marker-less automatic registration algorithm applicable to multi-type point cloud data provides precision and more complete data that is critical to forest resource investigation and evaluation.

Key words: point cloud, registration, tree segmentation, feature extracting, graph matching