《目标跟踪基础：Subhash Challa与Darko Musicki指南》

需积分: 7 24 浏览量更新于2024-07-19 收藏 2.95MB PDF 举报

《目标跟踪基础：Subhash Challa、Mark Morelande、Darko Musicki与Robin Evans合著》是一本深度探讨目标跟踪领域的经典著作，由National ICT Australia (NICTA) 和 University of Melbourne 的学者们共同编写，以及来自Hanyang University的Darko Musicki共同完成。本书是针对研究者和专业人员的详尽指南，旨在提供一个全面理解目标跟踪原理和技术的框架。书中涵盖了目标跟踪的各个方面，从基本概念到最新研究成果，包括但不限于追踪算法的设计、特征选择与描述、多目标跟踪、视觉传感器数据处理、深度学习在目标跟踪中的应用、以及实际场景中的挑战与解决方案。作者们结合丰富的理论知识和实践经验，将复杂的理论分解成易于理解的模块，使读者能够系统地掌握这一领域的核心原理。对于那些想要深入理解计算机视觉、人工智能和机器人技术的读者来说，本书是一本宝贵的参考资料。它不仅阐述了传统的跟踪方法，如卡尔曼滤波和粒子滤波，还介绍了现代深度学习技术如何推动了目标跟踪技术的革新，如卷积神经网络（CNN）在追踪任务中的应用。此外，书中特别强调了跨学科的视角，讨论了如何将目标跟踪应用于不同领域，如自动驾驶、视频监控、无人机控制等。书中还包含了案例研究和实践指导，帮助读者将理论知识转化为实际操作能力。值得注意的是，版权方面，所有复制或再版行为需遵守版权法规定及集体许可协议，未经Cambridge University Press的书面许可不得进行。该书于2011年首次出版，并在全球范围内由Cambridge University Press发行。感兴趣的读者可以通过其官方网站获取更多关于图书的信息和购买链接。《目标跟踪基础》是一本既实用又具有学术价值的书籍，对于任何从事目标跟踪研究、开发或应用的人来说，都是一个不可或缺的学习和参考资源。通过阅读这本书，读者可以提升自己的专业知识，跟上该领域的发展动态。

6 Introduction to object tracking

Figure 1.4 Balloon for weather monitoring (NASA/courtesy of nasaimages.org).

height from which the ground-based radar can track it automatically. A second

approach is to use GPS devices in the balloon. But as the balloon is unrecoverable,

the cost of the device adds to the cost of the whole process. Moreover, as this

process is repeated throughout the day, using a GPS device becomes extremely

costly.

A clever solution is to build a system that tracks the initial stages of the balloon

trajectory and then hands it over to the ground radar once the balloon reaches a

speciﬁed height. In order to do so, a radiosonde device is attached to the balloon.

This device sends a synchronization pulse to three base stations. The location of

the balloon is estimated by a ﬁlter that works on the principle of triangulation

using the time-of-arrival of the synchronization pulse in three base stations. Once

the balloon crosses a pre-speciﬁed height, a ground-based radar is locked onto the

balloon and tracks it further into the atmosphere.

1.1.4 Cell biology

In studies of humans, animals, plants and insects, medical researchers and pathol-

ogists routinely study birth and death rates and the movement of biological cells.

In immunology, the organism’s immune response is correlated with the life cycle

1.2 Bayesian reasoning 7

Figure 1.5 Tracking of lymphocyte cells.

of lymphocytes. The parameters of interest are the division/birth and death time of

each generation of cells. In fertility studies, the interesting parameter is the velocity

or shape of sperm cells. In the case of anti-inﬂammatory diseases, researchers

look for the speed and acceleration of lymphocyte cells (Figure 1.5). Medical

researchers carefully prepare the cells and take sequential images at regular inter-

vals. In some cases, the images can be collected over several days. These images

are then examined manually. The process of going through the set of images and

ﬁnding out the required parameters, i.e., speed, division and death, of each cell is

both time-consuming and prone to error and in some cases is almost impractical.

The problem can be formulated as the problem of tracking cells over a sequence

of images. The events of cell division and death are observed by looking at the

track initiation and termination probabilities. The process is not only faster and

reliable but also enables the discovery of previously unobserved phenomena. The

key concept is to deﬁne the “data association” parameters for cell identiﬁcation and

tracking and then using the domain knowledge of cell immunology to model the

“division/birth” and “death” of cells and integrate those into the recursive Bayesian

solutions.

1.2 Bayesian reasoning with application to object tracking

The Bayesian approach is a well-developed probabilistic and statistical theory that

can be applied to the modeling and solution of problems encountered in object

tracking. Bayesian methods are widely used in the statistical and engineering com-

munities, and engineers have long made use of them in solving a large variety of

problems.

In most object tracking problems, like the ones introduced in Section 1.1, mea-

surements from sensors are received in a sequential manner over time. At each

measurement stage, new estimates of the object state are made by combining new

information with current estimates of the object state. These latest estimates are

8 Introduction to object tracking

updated in light of new information received at the next stage and so on. The recur-

sive form of Bayes’ theorem is an appropriate framework for handling the sequen-

tial nature of the measurement data and the associated uncertainties. The Bayesian

recursive approach is based on Bayes’ theorem, the mathematical and conceptual

basis upon which the Bayesian paradigm is built. The Bayesian paradigm is a the-

oretical framework for reasoning under uncertainty. The history of Bayes’ theorem

goes back to the eighteenth and early nineteenth centuries with the work of Thomas

Bayes (1764) and Pierre-Simon de Laplace (1812). The theorem was ignored for

a long time. In the latter half of the twentieth century interest in the Bayesian

approach grew rapidly ﬁnding application in many areas of science, engineering

and statistical inference. Stigler (1986) provides an historical account of Bayesian

thinking.

1.2.1 Bayes’ theorem

Bayes’ theorem is the encapsulation of a philosophy to consistently reconcile past

and current information by exploiting the conditional probability concepts of prob-

ability theory. It may be stated as follows: given two related events x and y,the

conditional probability of event x given observation of event y is

p(x|y) =

p(x, y)

p(y)

. (1.1)

It is equal to the joint probability of events x and y, p(x, y), normalized by the

unconditional probability of event y, p(y). Using (1.1) twice, Bayes’ theorem can

be rewritten as

p(x|y) =

p(y|x)p(x)

p(y)

. (1.2)

Various aspects of the object tracking problem, e.g., the number of objects and

their states, can be modeled as events x and many types of sensor outputs, e.g.,

radar returns or infrared images, as events y. Bayes’ theorem in (1.2) can then be

applied to obtain the conditional probabilities p(x|y) as a probabilistic answer to

the object tracking problem.

Interpretation of Bayes’ theorem

Let x be a random variable of interest. In object tracking, x is often the state of

the object under consideration. Let y be a measurement related to x. Given knowl-

edge of y, we want to update our existing knowledge about x.Sincex is a random

variable, and using a probabilistic approach in dealing with uncertainty, we repre-

sent our knowledge about x with a probability distribution or a probability density

1.2 Bayesian reasoning 9

function (pdf) p(x), depending on whether x takes discrete values or is continu-

ous. p(x) is a function of x that attributes probability values to x. The problem

at the core of most tracking problems is how to update p(x) to take into account

the new information y. The answer is provided in (1.2) by p(x|y), the conditional

distribution of x given y.Viewingx as a random variable that can take a range of

values, we can see that for a ﬁxed y, p(x) is altered by p(y|x) to become p(x|y).

We can ignore the effect of p(y), since it is the same for all x values, once y is

known. It is this, theoretically sound, alteration of p(x) for each possible value

x that makes us change our probability representation of x, from p(x) to p(x|y).

It is done with a re-weighting of the initial probability distribution with the term

p(y|x

) for each x value. This deceptively simple operation is at the heart of many

complex algorithms for estimating parameters of dynamic processes. The term

L(x) = p(y|x) is known as the likelihood function when viewed as a function of x.

We write

p(x|y) ∝ p(y|x)p(x)

to mean that p(x|y) is proportional to the product of p(y|x) and p(x). The term

p(y) is a constant when viewed as a function of x and is called the normalizing

constant or normalization factor which ensures that p(x|y) as a function of x sums

up or integrates to 1. The initial distribution p(x) is called the prior distribution.

The new distribution of x, p(x|y), is called the posterior distribution.

Consider the single object whose state can be represented by the one-

dimensional vector x. Suppose that given all past history, our current knowledge

about x is contained in the probability density function

p(x) =



√

2π

−

(x−x

)

2

This means that we think that x is near x

, with an estimation error quantiﬁed

by 

. We may have arrived at this normal distribution (Gaussian distribution)

through some previous calculations. We write p(x) = N(x;x

, 

). The bigger



the less certain we are about x. Assume that a sensor produces an observation

y about x. Knowing that the sensor output has measurement error, we write the

following equation:

y = x + w, p(w) = N (w;0, 

The measurement error w is modeled probabilistically using a Gaussian distribu-

tion with zero mean and variance 

. Using the basic properties of the Gaussian

distribution (see Hogg and Craig, 1995), the likelihood function L(x) = p(y|x)

剩余384页未读，继续阅读

space_006

粉丝: 17
资源: 53

《目标跟踪基础：Subhash Challa与Darko Musicki指南》

kalman-filter-in-single-object-tracking-main.zip

Evaluating Multiple Object Tracking Performance - The Clear MOT Metrics

分别详细介绍以下的GCC编译选项的功能原理： -fno-var-tracking-assignments-toggle -fno-var-tracking-uninit -fvariable-expansion-in-unroller -fno-tree-partial-pre -funconstrained-commons -fno-unroll-all-loops -funroll-loops -funsafe-math-optimizations -fno-vpt

详细介绍一下GCC的编译选项 -fno-var-tracking-assignments-toggle

h5 tracking-min.js

github.com/easinal/target-recognition-and-tracking-of-vehicle-system

请解决以下代码的问题ui_mainwindow.h:614:10: note: variable tracking size limit exceeded with -fvar-tracking-assignments, retrying without void setupUi(QMainWindow *MainWindow)

tracking-by-detection

详细分析一下GCC编译选项-fno-var-tracking-assignments-toggle 的作用原理

tracking-min.js下载

最新资源