基于生成树的多视图激光扫描全局注册方法

175 浏览量更新于2024-08-26 收藏 2.87MB PDF 举报

"这篇研究论文提出了一种使用生成树进行多视图范围扫描局部到全局注册的有效方法，用于3D重建。" 文章深入探讨了3D重建中的一个重要环节——注册问题。注册是构建三维模型的关键步骤，特别是在多视图场景中，确保不同视角的扫描数据能够准确对齐至关重要。该论文提出的方法依赖于生成树的概念，以解决多个激光雷达或深度图像之间的配准问题。首先，论文指出，在给定初始位置的扫描数据后，需要估算每一对扫描之间的重叠百分比。这是确定哪些扫描之间存在关联性以及如何建立生成树的基础。生成树是一种特殊的树结构，其中每个节点代表一个扫描，边则表示扫描之间的关系，特别是它们之间的重叠部分。基于生成树，论文提出了一个序列化的配对注册策略。通过这个策略，可以按照树的结构顺序地进行配对注册。这通常涉及迭代最近点（ICP，Iterative Closest Point）算法，这是一种常用的点云配准方法，它通过迭代优化来寻找两个点集之间的最佳对应关系，从而实现最佳的配准。作者们在文中详细阐述了生成树构建和ICP迭代过程的执行细节，以确保注册的准确性和效率。他们可能还讨论了如何处理遮挡、噪声和不完全覆盖等常见问题，以及如何通过优化算法提高配准的鲁棒性。此外，论文还可能包含了实验结果和分析，对比了所提方法与其他现有技术的性能，以证明其在精度、计算效率和鲁棒性方面的优势。这些实验可能在各种复杂场景下进行，以展示方法在实际应用中的有效性。这篇研究论文为3D重建领域提供了一种创新的多视图范围扫描注册方法，通过生成树的构建和序列化配对注册，为解决大规模3D数据对齐问题提供了新的思路。这种方法对于提升3D重建的质量和效率具有重要意义，尤其是在机器人导航、建筑建模、文化遗产保护等领域有着广泛的应用前景。

J. Zhu et al. / Computers and Electrical Engineering 58 (2017) 477–488 479

Fig. 1. Overview of the coarse-to-ﬁne registration approach. Each dot denotes a scan; the black dot is the reference scan; each dotted line represents an

initial parameter. Scans inside one red circle are integrated into a coarse model.

three steps are included in each iteration: for each point in the data shape, ﬁnd its nearest-neighbor from the model shape;

based on established correspondences, conﬁrm overlapping areas between the model and data shapes; utilize these point

pairs located in overlapping areas to calculate the rigid transformation. Given suitable initial parameters, the optimal solu-

tion can be obtained by the iteration of these three steps until convergence criteria are satisﬁed. For further details, please

refer to [23] .

2.2. Coarse-to-ﬁne registration approach

Given a set of initially posed scans, multi-view registration can be achieved by the coarse-to-ﬁne approach based on the

TrICP algorithm [23] . Fig. 1 displays an overview of this approach, which can be described as follows:

(1) Divide all the involved scans into two parts: the n th scan and the coarse model integrated by all other scans. The

n th scan and the coarse model can be viewed as the data shape and the model shape, respectively. In this case,

parameters of the n th scan can be reﬁned by pairwise registration. Because the reference frame is attached to the

ﬁrst scan, it is always in its correct position. Therefore, there is no requirement to reﬁne its registration parameters.

(2) Reﬁne registration parameters of the n th scan. In general, the n th scan has a high overlap percentage to the coarse

model, which is integrated by all other scans. Given initial parameters of the n th scan, the TrICP algorithm can be

applied to obtain more accurate parameters, which can be returned to reﬁne the coarse model for the registration of

other scans.

(3) Process each scan sequentially by the implementation of the above two steps.

(4) To obtain the desired registration result, several update cycles are required. After one update cycle, registration pa-

rameters of all scans become more accurate. However, during the update of the i th scan, the other scans are not in

their correct positions; hence, it is not possible to obtain the desired registration with only one update cycle.

(5) Repeat steps (1) ∼ (4) until the iteration number reaches the threshold or other stop conditions are satisﬁed.

Given suitable initial parameters, the coarse-to-ﬁne approach can obtain accurate results for multi-view registration.

However, its eﬃciency and robustness must be further improved especially when multi-view scans are initially posed with

less than suitable parameters.

3. Methodology

This section presents the proposed approach for multi-view registration of initially posed scans using a spanning tree,

including the construction of the spanning tree, local, and global multi-view registration.

3.1. Overview

Given a set of initially posed scans, the goal of multi-view registration is to reﬁne these rigid transformations of the

other scans to the reference scan. Without loss of generality, the coordinate of the ﬁrst scan can be viewed as the reference

frame. Because there are many parameters required to be estimated, multi-view registration is time-consuming and accurate

results are diﬃcult to obtain. Therefore, an effective approach must be proposed to solve this problem.

For the registration problem, three facts can be reviewed: the more suitable the initial parameters provided for multi-

view registration, the more easily it will be solved; multi-view registration is more diﬃcult than pairwise registration, how-

ever, its diﬃculty is decreased with a decrease of the number of involved scans; for pairwise registration, its diﬃculty is

decreased with an increase of the overlap percentage between two scans. Motivated by these three facts, the multi-view

registration problem can be decomposed into sub-problems: pairwise registration, local, and global multi-view registration.

Fig. 2 displays an overview of the proposed approach, which solves these sub-problems by three steps:

(1) Construction of a spanning tree. Based on initial parameters, estimate the overlap percentage for each scan pair and

then construct a spanning tree according to the estimation of overlap percentages.

(2) Local multi-view registration. For these scans located in the second level of the spanning tree, local multi-view regis-

tration can be applied to reﬁne their registration parameters. After local registration, these scans in the second level

剩余11页未读，继续阅读

weixin_38746515

粉丝: 15
资源: 945

基于生成树的多视图激光扫描全局注册方法

mysql高频面试题及答案

[21天学通Oracle].张朝明.扫描版.pdf

Oracle11gR2中的物化视图和查询优化

Go依赖项树状视图：可视化你的go.mod文件

Linux多用户、多任务环境下的安全策略：确保环境安全的6个技巧

揭秘Oracle数据库表名查询：从基础到高级

【排序的艺术】：从字典到有序字典的转换技巧

Oracle数据库性能调优实战：从理论到实践的进阶之路

Oracle数据库性能调优实战指南：从监控到优化，提升系统效率

揭秘Oracle数据库性能优化秘籍：从理论到实战，全面提升数据库性能

最新资源