计算机科学
点云
遥感
迭代最近点
点(几何)
算法
稳健性(进化)
人工智能
计算机视觉
迭代法
图像配准
云计算
大气模式
作者
Xusong Dai,Xinlian Liang
标识
DOI:10.1109/tgrs.2026.3712807
摘要
Laser scanning plays an increasingly important role in forest resource monitoring by enabling accurate and nondestructive 3-D structural acquisition. A key prerequisite for reliable forest interpretation is the precise registration of multiview point clouds into a coherent representation. The registration process typically comprises two stages: coarse and fine registration. Coarse alignment adjusts the relative positions and orientations of input point clouds, while fine registration is essential for achieving high-precision alignment. However, most existing studies on forest point cloud registration have primarily focused on coarse alignment, while fine registration is typically performed using the iterative closest point (ICP) algorithm or its variants. These fine registration methods were originally designed for structured environments and perform poorly in forests due to repetitive geometries, occlusions, noise, and limited overlap. To overcome these challenges, this study proposes a cylinder-constrained iterative closest point (CC-ICP) algorithm specifically designed for the fine registration of forest point clouds. The method employs an adaptive multisource fusion scheme that dynamically combines ground-derived point-to-plane residuals, which provide vertical constraints, and trunk-derived point-to-point residuals, which provide horizontal constraints, into a unified least-squares objective, solved via the Gauss–Newton method. Extensive experiments on 24 plots across two mountainous forests demonstrate that the proposed method consistently outperforms five widely used algorithms—ICP, point-to-plane ICP, generalized ICP (G-ICP), symmetric ICP, and NDT—in terms of robustness, accuracy, and computational efficiency, highlighting its effectiveness in complex forest environments. The source code of the proposed method is available athttps://github.com/MSpace-WHU/CC-ICP
科研通智能强力驱动
Strongly Powered by AbleSci AI