基于核贝叶斯框架的多线索融合鲁棒头部跟踪

5 浏览量更新于2024-08-27 收藏 423KB PDF 举报

"这篇论文提出了一种基于核贝叶斯框架的多线索融合的稳健头部追踪算法。该算法结合了高斯混合模型(GMM)的空间约束外观模型和多通道 chamfer 匹配形状模型，两者互补，提高了目标与背景的区分度。同时，采用了选择性更新技术适应外观和光照变化，并将核方法的 mean-shift 算法融入贝叶斯框架，以在假设生成过程中提供启发式预测，减轻了传统贝叶斯追踪器的计算负担。实验结果证明了该算法的有效性和鲁棒性。" 本文关注的是计算机视觉领域中的目标追踪问题，特别是针对头部追踪的应用。作者提出的方法结合了多种技术，以实现对头部目标的精确和稳健追踪。首先，论文采用了一个基于高斯混合模型(MoG)的外观模型，利用空间约束来描述目标的外观特征。高斯混合模型是一种概率模型，可以有效地表示复杂的目标分布，通过多个高斯分量来捕捉目标的不同外观状态。空间约束则有助于确保模型在跟踪过程中的定位准确性。其次，引入了多通道 chamfer 匹配形状模型，这是一种计算几何方法，用于衡量目标形状与模板之间的相似度。通过这种方式，算法可以更准确地识别和匹配目标的形状，即使在部分遮挡或变形的情况下也能保持追踪性能。为了适应目标外观的变化（如光照变化），论文提出了选择性更新技术。这种技术只更新那些对跟踪最有影响的模型参数，避免了全局更新导致的过拟合或追踪漂移问题，提高了算法的鲁棒性。此外，论文将核方法的 mean-shift 算法集成到贝叶斯框架中。mean-shift 是一种非参数聚类算法，能够在特征空间中寻找目标的局部模式。在贝叶斯框架下，它被用来生成假设，帮助预测目标的下一个位置，减少了传统方法中生成大量无效假设的计算需求。实验结果显示，结合了这些技术的头部追踪算法在处理真实场景中的头部追踪任务时，不仅能够有效应对复杂的环境变化，而且具有较高的计算效率，提升了追踪的稳定性和精度。这篇论文提供的是一种综合的、基于多线索融合的头部追踪解决方案，对于理解和改进现有的目标追踪算法，特别是在实时和复杂环境下的应用，具有重要的参考价值。

ZHANG et al.: ROBUST HEAD TRACKING BASED ON MULTIPLE CUES FUSION IN KERNEL-BAYESIAN FRAMEWORK 1199

and illumination changes, and to prevent the model from

drifting away.

The arrangement of this paper is as follows. A brief review

of kernel-based and Bayesian-based tracking frameworks is

given in Section II. The kernel-Bayesian framework is de-

scribed in detail in Section III. The multiple cues fusion-based

similarity measure and its application in the kernel-Bayesian

framework are discussed in Section IV. Experimental results

are presented in Section V, and Section VI is devoted to a

conclusion.

II. Review of Kernel-Based and Bayesian-Based

Frameworks

In this section, we brieﬂy review the two typical track-

ing frameworks: kernel-based framework and Bayesian-based

framework.

A. Kernel-Based Framework

The most famous kernel-based framework, namely the mean

shift algorithm, ﬁrst appeared in [23] as a method for estimat-

ing the gradient of a density function. It was applied for visual

tracking by Comaniciu et al. [3] in 2000.

Mean shift is a nonparametric mode seeking technique that

shifts each data point to the average of the data points in

its neighborhood [23]. Let R be a ﬁnite set embedded in

n-dimensional space, the mean shift vector ms of x is deﬁned

as follows:

ms =



K(a − x)w(a)a



K(a − x)w(a)

− x, a ∈ R (1)

where K is a kernel function and w is a weight function. The

mean shift algorithm works by iteratively shifting the data in

the direction of mean shift vector until convergence.

B. Bayesian-Based Framework

Another popular approach is to view tracking as an online

Bayesian inference process for estimating the unknown state

from sequential observations o

1:t

perturbed by noise. A

dynamic state-space form employed in Bayesian inference

framework is shown as follows [27]:

state transition model : s

= f

t−1

,

)(2)

observation model : o

= h

,ν

) (3)

where s

are system state and observation, 

,ν

are the

system noise and observation noise, respectively, f

(., .) char-

acterizes the kinematics of the object, and h

(., .) models the

observations. The key idea of Bayesian inference is to approx-

imate the posterior probability distribution by a weighted sam-

ple set {(s

(n)

)|n =1,... ,N}. Each sample consists of an

element s

(n)

that represents the hypothetical state of an object

and a corresponding discrete sampling probability w

(n)

, where



n=1

(n)

= 1. First, the sample set is resampled to avoid

the degeneracy problem, and the new samples are propagated

according to the state transition model. Then, each element of

the set is weighted with probability w

(n)

= p(o

(n)

), which

is calculated from the observation model. Finally, the state

estimate

can be either be the minimum mean square error

estimation or the maximum a posteriori (MAP) estimation.

III. Kernel-Bayesian-Based Framework

The kernel-based framework has a low-computational com-

plexity, but it is often trapped in local optima, while Bayesian-

based framework can improve the robustness of the tracking

process, but it suffers a large computational load by generating

a huge number of hypotheses to cover the global optimum.

Thus, in this section, we propose a kernel-Bayesian tracking

framework that combines the merits of both frameworks.

A. Kernel-Bayesian Framework

The state transition model is an important component of

the Bayesian tracking framework. Most of the existing models

use a naive random walk around previous system states [28]

or learn through prelabeled video sequences [2]. The random

walk approaches do not use information about the object

motion, and thus involve a quite large computational load since

a large number of hypotheses need to be randomly generated to

cover the object. The learning-based approaches often suffer

from over ﬁtting, so they are only effective for the training

sequences.

The kernel-based mean shift algorithm provides an estimate

of the object motion, which motivates us to embed the kernel

method into a Bayesain framework to provide a heuristic prior.

In detail, the mean shift algorithm is ﬁrst applied to the current

frame to obtain the direction of the object motion and the offset

of the object state, which are then incorporated into the state

transition model as prior information. In this way, the kernel-

based method and the Bayesian-based method are combined

into a uniﬁed framework.

B. Optimization View

A reinterpretation of the kernel-Bayesian framework from

an optimization point of view is presented to show why this

framework can combine the merits of both the kernel method

and the Bayesian method.

An input image with three templates superimposed, cor-

responding to the initialization, the local maximum and the

global maximum are illustrated in the left column of Fig.

1, and its likelihood function based on the spatial-constraint

MOG-based appearance model is shown in the right column of

Fig. 1. As shown in Fig. 1, starting from the initial position, the

kernel method converges to a local maximum that is near to the

global maximum. It is clear that a few hypotheses generated

around the local maximum are enough to guide the algorithm

to the global maximum. If the tracker starts from the initial

position, more hypotheses need to be generated in order to

reach the object.

IV. Proposed Tracking Algorithm

In our work, the motion state of a tracked object between

two consecutive frames is approximated by a set of afﬁne

剩余11页未读，继续阅读

weixin_38610012

粉丝: 3
资源: 866

基于核贝叶斯框架的多线索融合鲁棒头部跟踪

Robust and Precise Vehicle Localization based on Multi-sensor Fusion in...中文翻译

Robust Object Tracking via Sparsity-based Collaborative Model

Robust MKKM (Multiple Kernel k-Means) using Min-Max Optimization 多核聚类算法中，通过核权重系数来最大化簇间差异有什么优点

Robust MKKM (Multiple Kernel k-Means) using Min-Max Optimization 多核聚类算法中，max不是使聚类结果更差吗

springboot考勤打卡精确定位

springboot vue

Robust MKKM (Multiple Kernel k-Means) using Min-Max Optimization 多核聚类算法中，通过核权重系数来max什么

Recovering Realistic Texture in Image Super-resolution by Deep Spatial Feature Transform

最新资源