Tensor Voting理论解析与代码示例

4星 · 超过85%的资源需积分: 42 134 浏览量更新于2024-07-19 收藏 11.38MB PDF 举报

本篇论文深入探讨了Tensor Voting理论在自由空间建模中的应用，特别是在变化检测和地形提取领域的创新方法。作者 Bradford James King 在 Rensselaer Polytechnic Institute 提交的博士论文中，旨在解决计算机科学中的关键问题，如三维数据分析和物体识别。论文的核心目标包括： 1. **自由空间建模**：论文首先介绍了自由空间模型，这是一种利用多源传感器数据，如激光雷达、结构光或者时间飞行法（TOF）生成的三维环境表示，它有助于理解场景的物理特性。 2. **变化检测**：张量投票理论在此背景下被用来检测和识别场景中的变化，这对于动态环境监控和实时更新地图至关重要。通过将不同传感器的数据融合，算法可以区分静态和动态元素，增强系统的鲁棒性。 3. **Tensor Voting**：这是一种基于多视角信息融合的方法，通过计算和聚合多个观测数据的局部一致性，形成全局决策，从而提高位置估计的精度和可靠性。这种技术特别适合处理非结构化和复杂的数据，如激光雷达点云。 4. **地形提取**：论文的重点应用之一是利用张量投票来提取和重构地形信息，这在地理信息系统（GIS）和机器人导航中具有实际价值，能够帮助机器理解和适应其周围环境。 5. **挑战与解决方案**：论文还探讨了在实施张量投票过程中可能遇到的挑战，如数据噪声、传感器精度差异以及计算复杂性。作者针对这些问题提出了一系列有效的策略和优化算法。 6. **论文结构**：整个论文按照逻辑顺序组织，从背景介绍到理论阐述，再到实验验证和结论，为读者提供了一个全面且深入的理解框架。 7. **版权信息**：论文在2008年由Bradford James King撰写，并保留所有权利，对于学术交流和后续研究具有重要意义。这篇论文提供了对Tensor Voting理论在自由空间建模中如何应用于变化检测和地形提取的详尽解析，对于从事计算机视觉、机器人技术或GIS领域的研究人员来说，是一份宝贵的参考资料。

xvi

environments present new problems because they are typically uncontrolled. Scene

changes may occur during or between scans due to transient objects such as doors,

cars, and people, or due to structural changes such as new construction or the re-

moval of a tree. Some objects may be missed if they do not reﬂect suﬃcient light

towards the scanner. Furthermore, viewpoint selection is often limited. An observer

may not have the capability, permission, or time to acquire range scans from all de-

sired viewpoints. These problems result in range data of substantially lower quality

than that acquired in controlled environments, as shown in Figure

1.1. We require

new advances in analysis techniques to deal with it.

1.1 Objectives

The overall objective of this thesis is to produce range data analysis tech-

niques capable of dealing with scans acquired in outdoor, uncontrolled environments.

Speciﬁcally, we attack the following problems.

1.1.1 Free-Space Model ing

Laser-range scans observe more information than just the surface sample points

reported as range data. They also implicitly observe free-space, the empty volume

through which a scanning laser passes before it reﬂects oﬀ a surface and while it

returns to the scanner. A free-space model describes the geometry of the free-space

associated with given range data. Free-space modeling is a powerful tool suited to

applications such as those we illustrate in Figure 1.2.

• Change Detection (Figure 1.2a): Changes in a scene may occur during the

time between two diﬀerent scans observing it. Range samples that appear in

one scan at a locations inside free-space observed by another scan imply that

a change occurred. This application is the focus of our free-space modeling

work. Section 1.1.2 discusses it in more detail.

• Surface Topology Inference (Figure 1.2b): Surface reconstruction from

range data and multi-view imagery is an ill-deﬁned problem — many diﬀerent

scenes may produce a given set of range data or intensity images. Ambiguities

reconstructed from range s cans.

Our objective is to infer the free-space observed by range scans of a scene and

construct a free-space model. In particular, we focus on single-viewpoint, time-of-

ﬂight scanners. These scanners measure depth along each direction by the time laser

light takes to travel to and from a surface. They imply a line-segment of free-space

between each range sample an the viewpoint, since the laser is able to pass through

the space. The goal is to combine this information from all range samples to produce

a coherent free-space volume and boundary.

1.1.2 Change Detection

In order to distinguish the static and dynamic structure in a scene, one must

detect changes that occur by comparing multiple observations made at diﬀerent

times. Change detection in intensity images has been widely studied due to its

variety of applications [181]. To our knowledge, no one has studied change detection

using observations made by a range scanner in uncontrolled, outdoor environments.

Our objective is to detect evidence of changes in a scene given two or more

laser-range scans acquired at diﬀerent times and from varying viewpoints. Figure

1.3 shows an example of the desired result. Speciﬁcally, our goal is to identify a

set of points in the scans that lie on objects added, moved, or removed over the

period of time during which scans were taken. Identiﬁcation of the points is useful

to ﬁlter data before surface reconstruction or as a primitive for object-level change

detection. Our work focuses on the data structures and algorithms to support point-

level change detection.

1.1.3 Tensor Voting

The Tensor Voting Framework [154] is a computational method for inf erence

of structure among unorganized sample points in N-D. Its theoretical foundation

lies in perceptual organization and is based on the intuition that the human visual

system prefers connections with the lowest total curvature [

104] [105]. This makes it

particularly suitable for computer vision applications, including surface reconstruc-

tion [219], stereo analysis [142] [221], motion estimation and segmentation [128]

剩余188页未读，继续阅读

yanyikun

粉丝: 0
资源: 5

Tensor Voting理论解析与代码示例

voting:投票程序

voting

voting-app:一个投票应用程序，我们可以在许多不同类别上进行投票

PyTorch中torch.tensor与torch.Tensor的区别详解

tensor转dataframe的代码

如何用代码实现tensor降维

torch::Tensor out_tensor = module.forward({ input_tensor }).toTensor();代码解释

PyTorch Tensor和MLP创建代码

如何用tensor画图，展示代码

两个tensor矩阵相减代码

最新资源