实时流形正则化上下文感知追踪算法优化

需积分: 5 79 浏览量更新于2024-07-10 收藏 2.71MB PDF 举报

本文探讨了实时流形正则化上下文感知相关跟踪（Real-time Manifold-Regularized Context-Aware Correlation Tracking）这一主题，发表于2020年的《Frontiers in Computational Science》期刊，卷14，第2期，334-348页。该研究论文的DOI为<https://doi.org/10.1007/s11704-018-8104-y>。作者是Jiaqing FAN、Huihui SONG、Kaihua ZHANG、Qingshan LIU、Fei YAN和Wei LIAN，分别来自南京信息科技大学的大数据分析技术江苏省重点实验室以及大气环境与装备技术协同创新中心，以及长治大学的计算机科学系。传统的基于相关滤波器（Correlation Filter，CF）的跟踪方法在实际应用中取得了显著的成功，然而它们假设样本具有循环结构，这导致了学习有效分类器时存在显著的冗余。论文的核心创新在于，提出了一种新的实时跟踪算法，它考虑了不同类型样本的局部流形结构信息，从而克服了循环结构带来的局限性。该算法首先区别于传统的CF跟踪方法，后者仅依赖单一的“模板”来构建特征。新算法引入了流形正则化，这是一种对数据分布进行建模的技术，有助于捕捉样本之间的非线性关系和复杂结构。通过这种方法，算法能够在跟踪过程中更好地理解和适应目标物体随时间和环境变化的动态特性。此外，上下文感知的引入进一步增强了跟踪性能。通过结合周围环境信息，算法能够更准确地评估目标在不同场景中的相似性，避免因单一模板变化而引起的误跟踪。这在处理复杂背景干扰和目标遮挡等问题时表现出优势。在实施上，该算法实现了快速迭代，确保了实时性，这对于许多实时视频监控和机器人视觉等应用场景至关重要。论文还可能包括实验结果，展示了与传统CF方法相比，流形正则化上下文感知相关跟踪在精度、稳定性和速度方面的改进。这篇研究论文为计算机视觉领域的跟踪问题提供了一个新颖且有效的解决方案，它通过结合流形理论和上下文信息，优化了基于相关滤波的跟踪算法，提高了性能，并为实时应用带来了显著的提升。

336 Front. Comput. Sci., 2020, 14(2): 334–348

a part-based CF tracking approach that models the target with

multiple parts using several CFs. Lukezic et al. [26] model the

part-based CF responses and their constellation constraints

jointly as an equivalent spring system, and derive a highly ef-

ﬁcient optimization approach to infer the most probable tar-

get deformation.

The DCF based trackers usually suﬀer from boundary ef-

fects, thereby limiting the discrimination capabilities of the

learned CFs. To address this issue, recently, Danelljan et al.

[27] reformulate the CF objective by introducing a spatially

Gaussian weight function to penalize non-zero ﬁlter values

outside the object bounding box. Mueller et al. [14] present

a framework that allows to explicitly incorporate surround-

ing context information into the CF learning. Diﬀerent from

the above-mentioned methods that construct lots of virtually

circulant samples to train a CF, recently, Galoogahi et al. [15]

leverage the whole frame to get a set of real negative samples,

which facilitate learning a better classiﬁer.

2.2 Manifold regularized tracking

Manifold regularization is usually applied to semi-supervised

learning with both labeled and unlabeled samples [28–30],

which constructs a Laplacian graph to leverage the samples to

exploit the hidden geometrical structure of the feature space.

For example, in feature space analysis, Chang and Yang [29]

exploit both labeled and unlabeled training data for a more

reliable feature space selection algorithm. Moreover, in vi-

sual tracking, Yu et al. [30] leverage the manifold structure

in the appearance space with spatio-temporal constraints to

perform robust person localization and tracking in real world

surveillance scenarios. Bai and Tang [31] employ an online

Laplacian regularized ranking support vector machine to es-

timate the object location for visual tracking. To make better

use of the unlabeled data and the manifold structure of the

sample space, Hu et al. [32] propose a manifold regularized

DCF based tracker with augmented circularly shifted sam-

ples and leverage a block optimization strategy that can be

eﬃciently computed via FFTs. Zhuang et al. [33] construct a

discriminative sparse similarity map for visual tracking based

on a Laplacian regularized multitask reverse sparse represen-

tation.

3 Manifold regularized context-aware corre-

lation tracking

We ﬁrst review the context-aware CF tracking approach [14]

that is most related to our MRCT, and then introduce the prin-

ciple of our MRCT in detail.

3.1 Context-aware correlation tracking

In [14], a set of k contextual patches x

∈ R

around the

tracked object x

∈ R

are extracted, whose corresponding

circulant matrices are X

∈ R

s×s

and X

∈ R

s×s

. These con-

textual patches are served as negative samples with zero la-

bels. The aim is to learn a ﬁlter w ∈ R

that gives the target a

high score while the surrounding area a low one. As a result,

the objective function is

min

X

w − y

+ λ w

+ λ

X

w

, (1)

where y is a vectorized regression target of a 2D Gaussian

and λ, λ

> 0 are regularization parameters.

To solve the convex optimization Eq. (1) eﬃciently, the

terms therein can be rewritten by stacking the contextual

patches and regression target into the special forms as fol-

lows

B =

⎡

⎢

⎣

√

⎤

⎥

⎦

y =

⎡

⎢

⎣

⎤

⎥

⎦

. (2)

Then Eq. (1) can be rewritten as

min

Bw −

y

+ λ w

. (3)

Similar to the traditional CF learning [7], a closed-form

solution of Eq. (3) can be eﬃciently calculated in the Fourier

domain, among which the following characteristic of the cir-

culant matrix is the key component for solving the problem

eﬃciently

X = Fdiag

(ˆ

)

, X

= Fdiag

(ˆ

∗

)

, (4)

where

x denotes the Fourier transform F

x, x

∗

represents the

conjugate of x. Then, the solution can be eﬃciently achieved

w =

∗



∗



+ λ + λ



i=1

∗



. (5)

The detection procedure is the same as the traditional CF

based tracking, and when a new image patch z is coming, the

tracking result is determined as the location of the maximum

response r, which can be simply calculated in the Fourier do-

main by

(

w, z

)

z 

w. (6)

3.2 Manifold regularization for augmented samples

As shown in Fig. 1, given the bounding box of the tracked

target, we ﬁrst expand it to include some surrounding re-

gions, and then partition the expanded region into several

剩余14页未读，继续阅读

weixin_38686231

粉丝: 10

实时流形正则化上下文感知追踪算法优化

流形正则化非光滑非负矩阵分解新方法

矩阵三分解与流形正则化的零镜头学习算法

流形正则化视角下的降维理论解析

流形正则化matlab代码-LapEMR:论文代码：用于可扩展流形正则化的Laplacian嵌入式回归

流形正则化相关对象跟踪

歧视感知流形正则化提升半监督分类

流形正则化转移距离度量学习

降维：从流形正则化角度的解释

稀疏诱导流形正则化凸非负矩阵分解算法

重叠社区检测的流形正则化对称联合链接模型

最新资源