目标跟踪的变结构多模型估计工程师指南

需积分: 11 17 浏览量更新于2024-07-26 收藏 625KB PDF 举报

“Engineer’s Guide to Variable-Structure Multiple-Model Estimation for Tracking” by X.R. Li from the Department of Electrical Engineering, University of New Orleans. 在目标跟踪领域，工程师们面临着一个混合估计问题，该问题涉及到连续性和离散性的不确定性。传统上，目标跟踪的建模包括目标运动/动力学以及传感器系统的建模。通常的做法是利用连续值的系统噪声和测量噪声来掩盖未知的建模误差或模型与真实系统之间的偏差。然而，目标跟踪的主要挑战源自两种离散性不确定性：测量源不确定性和目标运动不确定性。测量源不确定性指的是传感器系统为目标跟踪提供的测量可能源于外部源，这包括杂波、假警报、邻近目标，以及跟踪的目标本身。此外，它也可能来自目标采取的反制措施。这种不确定性使得区分有效测量和干扰变得复杂，对跟踪算法提出了挑战。另一方面，目标运动不确定性是指目标运动状态的不精确性，这可能是由于目标行为的不可预测性、环境影响或建模简化。例如，目标可能突然改变速度或方向，或者存在未知的动力学特性，这些都可能导致传统的单模型估计方法失效。为了应对这些挑战，变量结构多模型估计（Variable-Structure Multiple-Model，简称MM）方法被引入。这种方法允许系统动态地切换或选择不同的模型，以适应目标状态的变化和不确定性。变量结构系统的特点是其结构能够根据输入或状态信息的变化而改变，从而提供更灵活且适应性强的估计框架。在变量结构多模型估计中，通常会构建一组模型集合，每个模型对应目标可能的不同行为模式。随着观测数据的更新，这些模型的权重或概率会被动态调整，以反映哪个模型最能解释当前的测量。这种“模型跳变”策略使得系统能够在不确定性环境中优化跟踪性能。具体实现上，可以采用诸如“自适应滤波”（如扩展卡尔曼滤波器EKF或无迹卡尔曼滤波器UKF）和“模型预测控制”等技术，结合粒子滤波等非线性估计方法，以处理复杂的系统动态和不确定性。此外，还可能涉及贝叶斯决策理论和统计学习方法，用于概率模型的选择和更新。 “Engineer’s Guide to Variable-Structure Multiple-Model Estimation for Tracking”探讨了如何利用变量结构多模型方法来有效地处理目标跟踪中的离散不确定性问题，这对于设计高效、鲁棒的跟踪系统至关重要。通过理解和应用这些概念和技术，工程师可以开发出更适应复杂环境和行为变化的目标跟踪解决方案。

510 Multitarget-Multisensor Tracking: Applications and Advances

structure MM (FSMM) algorithm is one with a ﬁxed set of models while a variable-

structure MM (VSMM) algorithm is one with a variable set of models.

As the outcome of the advances during the past three decades, the state-of-the-

art FSMM estimators usually perform quite well for problems that can be handled

by the use of a small set of models. Consequently, they have found a great success in

solving many state estimation problems compounded with structural or parametric

uncertainty, particularly in target tracking. However, when they are applied to solve

many real-world problems (e.g., many practical target-tracking problems), it is often

the case that the use of only a few models is not good enough. Although further

development is certainly possible, the FSMM estimation techniques have arrived at

such a stage that great improvement can no longer be expected. This perception is

based on an understanding of the fundamental limitations of the FSMM approach.

These limitations stem from the following facts:

• It assumes fundamentally that the system mode at any time can be represented

(with a sufﬁcient accuracy) by one of a ﬁxed set of models that can be deter-

mined in advance.

• The set of possible system modes is not ﬁxed. It depends on the hybrid state of

the system at the previous time.

• As shown in [34, 38], use of more models in an FSMM estimator does not

necessarily improve performance; in fact, the performance will deteriorate if

too many models are used.

• It cannot incorporate certain types of a priori information.

• Clearly, the amount of computational resourcerequired by an FSMM estimator

increases dramatically with the number of models used.

These facts are explained next.

10.3.1 Limitations of Fixed Structures

Let S

be the set of possible system modes at time k and let S be the mode space

(i.e., the set of all modes at all times), which is the union of all S

’s.

Fact 1 below is based on the theoretical result of [34,38] that for a static prob-

lem, the MM estimator is optimal only if the model set used is equal to the mode

space.

Fact 1: The “optimal” FSMM estimator (i.e., the one that uses optimally all

possible model sequences formed by the elements of M) is not optimal if the model

set M used does not match exactly the mode set in effect at any time. Speciﬁcally, the

“optimal” (in the sense of having the minimum mean-squareerror) FSMM estimator

is given by

ˆx

k|k

= E[x

, S

= M, S

= M, . . . , S

= M] (10.14)

Some models may have zero probabilities at a given time, though.

Engineer’s Guide to Variable-Structure Multiple-Model Estimation for Tracking 511

which in general differs from the optimal estimate E[x

]. It is thus clear that

the “optimal” FSMM estimator is not optimal if the model set M is not equal to the

mode space S (i.e., M 6= S) or the set S

of possible modes is actually not ﬁxed.

There are many practical situations in which M 6= S, such as the following:

• The mode space S is too large and a smaller M is used because of limited

resources for processing or computation. This is almost always the case when

unknown parameters are involved.

• The system modes are too complex and their simpliﬁed models are used in

M. For instance, actual target maneuvers are usually far more complex than

the commonly used coordinated-turn or constant-acceleration models, and the

like.

• The mode space S is not known completely.

These situations also reveal the importance of and the difﬁculty involved in

model-set design for MM estimation.

Fact 2: Use of more models in an FSMM estimator optimally may still degrade

performance. This is shown theoretically in [34, 38]. This fact is closely related

with fact 1: The optimal FSMM ﬁlter is the one that uses M = S

for all time

k. It follows that adding more models into M may actually increase the mismatch

between M and S

and thus lead to performance deterioration.

Fact 3: The set S

k+1

of possible system modes at any time k + 1 depends in

general on the current hybrid state ξ

= [x

, s

]

of the system. This state depen-

dency of the mode set arises from the fact that a particular system mode may only

jump to the system modes for which the corresponding transition probabilities are

not zero. In reality, the mode transition probabilities are quite often dependenton the

base state of the system. For example, a car (as a ground target) on a closed highway

may move only along the highway, while a car at a four-way intersection may go

straight or take a left or right turn. Similarly, an aircraft approaching a runway from

an angle is most likely to take a turn toward the runway.

The FSMM estimators cannot make use of the above state dependency of the

mode set because the model set M is determined in advance relying only on a priori

information, that is, before any measurement is received in real time, without online

information of the state.

When tracking a car on a closed highway, the inclusionof any maneuver models

in the MM estimator will degrade the performance, while some maneuver models

should be present when the car is at an intersection. Note that intersections of a

different type in general require the use of different model sets for best performance

as well as computational complexity. Similarly, to save the resources for process-

ing and computation and to enhance performance, it is better not to include many

maneuver models in an MM estimator for tracking a civilian aircraft in an en-route

ﬂight. However, relatively more maneuver models should be present in the MM es-

512 Multitarget-Multisensor Tracking: Applications and Advances

timator when the same aircraft is in a terminal area. These situations exemplify the

need to use a variable model set (i.e., a variable structure) for better performance.

Fact 4: It is difﬁcult or impossible for an FSMM estimator to use many types

of a priori information concerning the system mode. One type of such a priori infor-

mation is the knowledge that some system modes are unlikely but not impossible.

For example, certain emergency actions of a target (e.g., a quick evasive maneuver

of a civilian aircraft) may only be triggered by a combination of certain adverse con-

ditions. It would be impossible for an FSMM estimator to include such information

unless it has a single model representing the combination of the adverse conditions.

The presence of such a model, however, would degrade the estimator’s performance

when the target is not in this emergency, let alone the potentially dramatic increase

in computation because of the need to cover all such unlikely situations. Another

example of such a priori information involves the transition time and/or jumping

magnitude between two consecutive modes; that is, the time elapse and/or modal

distance between two consecutive modes. For instance, a priori knowledge may

be available about the time period for a target of a particular type to reach certain

velocity change and/or the statistical distribution of the modal jump magnitude.

The limitations of the FSMM estimation have been more or less perceived by

those in the research community. Ad hoc remedies have been proposed for particular

applications, but few theoretical attempts were made to break away from the ﬁxed

structure prior to [37]. The investigation of the adaptive-grid PDA ﬁlter [22], the

moving-bank SMM estimators [49], and the simplex-directed SMM estimator [30]

within the SMM architecture are the only earlier meaningful efforts along this line

known to the author. Serious theoretical work was initiated in [37] and continued

in [32,34,38] to lay a theoretical foundation for VSMM estimation.

10.3.2 When Should a Variable Structure Be Used?

In general, a variable structure shouldbe considered if any of the followingsituations

applies.

• The mode space is large for the given resources for processing or computation.

• The set of (most) likely system modes is highly time variant or state dependent.

• The mode space is not known.

• Useful but complex a priori information and knowledge about the possible

modes is available.

In general, the more complex the problem is, the more superior the variable

structure is to the ﬁxed structure. A good example of the above situations is the

problem of tracking ground targets [28], described in Chapter 6. Two other examples

are given in Sections 10.7 and 10.8.

剩余68页未读，继续阅读

lvbaohong

粉丝: 0
资源: 2

目标跟踪的变结构多模型估计工程师指南

Design Engineer’s Guide to OpenVPX

轻量级人体姿态估计lightweight-human-pose-estimation.rar

lightweight-human-pose-estimation.pytorch-master

few-shot object detection and viewpoint estimation for objects in the wild

新版yolov8中，我找到了这样几个yaml文件，yolov8.yaml，yolov8-cls.yaml，yolov8-p2.yaml，yolov8-p6.yaml，yolov8-pose.yaml，yolov8-pose-p6.yaml，yolov8-seg.yaml

write DCC-GARCH model code

最新资源