SocialAttribute-awareForceModel:ExploitingRichnessofInteractionforAbnormalCrowdDetection资源-CSDN文库

101 浏览量 2021-02-07 01:38:29 上传评论收藏 2.23MB PDF 举报

### Social Attribute-aware Force Model: Exploiting Richness of Interaction for Abnormal Crowd Detection #### 概述本研究提出了一种社会属性感知力模型(Social Attribute-aware Force Model)，旨在通过捕捉个体之间的互动特征来提高异常人群检测的准确性。该模型不仅考虑了传统的人群行为分析方法中的物理力作用，还引入了社会属性的概念，如社会混乱度和社会拥堵度等，以更全面地理解人群中个体间的交互。 #### 社会属性感知力模型在传统的社会力模型基础上，本文创新性地引入了社会属性的概念，通过量化社会交互中的关键属性来增强对复杂场景的理解。该模型主要由以下几个部分组成： 1. **场景尺度估计**：在无监督的方式下高效估计场景的尺度，为后续的属性分析奠定基础。 2. **社会属性定义**：定义了两个关键的社会属性——社会混乱度和社会拥堵度。社会混乱度用于衡量人群中交互行为的不确定性或不规律性；而社会拥堵度则反映了人群密度的变化及其对个体移动的影响。 3. **在线融合策略**：利用在线融合策略将社会属性与社会力模型相结合，从而构建出一个能够有效描述互动行为的模型。 4. **异常事件检测**：基于上述社会属性感知力模型进行异常事件检测，并且该模型能够指示潜在异常交互的位置，提高了检测的准确性和鲁棒性。 #### 关键技术与贡献 1. **社会属性的引入**：通过定义社会混乱度和社会拥堵度，模型能够更好地捕捉到个体间的交互特征，尤其是在复杂多变的环境中。 2. **场景尺度估计**：通过无监督学习方法自动估计场景尺度，为后续的社会属性分析提供重要依据。 3. **在线融合策略**：采用在线融合策略将社会属性与社会力模型相结合，增强了模型对动态变化环境的适应能力。 4. **异常检测性能**：在公开数据集上的实验结果表明，该模型相比其他方法具有显著的优势，特别是在处理大规模、高复杂度的人群场景时表现突出。 #### 实验验证为了验证所提出的模型的有效性，研究人员在多个公开可用的数据集上进行了实验评估。这些数据集涵盖了各种不同的场景，包括但不限于正常和异常的人群活动。实验结果表明，社会属性感知力模型在异常检测方面表现出色，相较于现有的方法具有更高的准确率和鲁棒性。此外，该模型还能够有效地定位异常交互的发生位置，进一步提升了其实用价值。 #### 结论社会属性感知力模型是一种创新的方法，它结合了社会属性和社会力模型的优点，能够在复杂的环境下准确地检测异常人群行为。通过对社会属性的精确量化和模型的灵活应用，该方法为视频监控和人群行为分析领域提供了有力的技术支持。未来的研究可以进一步探索更多类型的社会属性以及如何将其更有效地整合到模型中，以应对更加复杂和多样化的应用场景。

资源推荐

资源详情

资源评论

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/TCSVT.2014.2355711, IEEE Transactions on Circuits and Systems for Video Technology

Social Attribute-aware Force Model: Exploiting

Richness of Interaction for Abnormal Crowd

Detection

Yanhao Zhang, Lei Qin, Member, IEEE, Rongrong Ji, Senior Member, IEEE, Hongxun Yao, Member, IEEE,

and Qingming Huang, Senior Member, IEEE

Abstract—Interactions between pedestrians usually play an

important role for understanding crowd behavior. However,

there are great challenges facing accurate analysis of pedestrian

interactions, such as occlusions, motion, appearance variance, etc.

In this paper, we introduce a novel social attribute-aware force

model for detection of abnormal crowd events. The proposed

model incorporates social characteristics of crowd behaviors to

improve the description of interactive behaviors. To this end,

we ﬁrst efﬁciently estimate the scene scale in an unsupervised

manner. Then, we introduce the concepts of social disorder

and congestion attributes to characterize the interaction of

social behaviors, and construct our crowd interaction model on

the basis of social force by an online fusion strategy. These

attributes encode social interaction characteristics and offer

robustness against motion pattern variance. Abnormal event

detection is ﬁnally performed based on the proposed social

attribute-aware force model. In addition, the attribute-aware

interaction force indicates the possible locations of anomalous

interactions. We validate our method on the publicly available

datasets for abnormal detection, and the experimental results

show promising performance compared to alternative and state

of the art methods.

Index Terms—Social Force Model, Social Attributes, Crowd

Behaviors, Abnormal Detection

I. INTRODUCTION

Analyzing crowd behavior has become a salient research

topic in video surveillance and beyond. In contrast to indi-

vidual actions, crowd behaviors are far more challenging to

analyze, due to the possibility for complex interactions among

individuals.

Crowd behavior models aim to describe individuals and

groups in crowded scenes. From a sociological viewpoint,

crowd behaviors usually occur under the constraints of the

sociologically inspired prior knowledge, and hence reﬂect

However, permission to use this material for any other purposes must be

obtained from the IEEE by sending an email to pubs-permissions@ieee.org.

This work was supported in part by National Basic Research Program of

China (973 Program): 2012CB316400, in part by National Natural Science

Foundation of China: 61025011, 61133003, 61332016 and 61035001.

Y. Zhang, Q. Huang, and H. Yao are with School of Computer Science and

Technology, Harbin Institute of Technology, Harbin 150001, China. Q. Huang

is also with University of Chinese Academy of Sciences, Beijing 100190,

China (e-mail: yhzhang@hit.edu.cn; qmhuang@jdl.ac.cn; H.yao@hit.edu.cn).

L. Qin is with Key Laboratory of Intelligent Information Processing of

Chinese Academy of Sciences (CAS), Institute of Computing Technology,

CAS, Beijing 100190, China (e-mail: qinlei@ict.ac.cn).

R. Ji is with Department of Cognitive Science, School of Information

Science and Engineering, Xiamen University, Xiamen 361005, China (e-mail:

rrji@xmu.edu.cn).

high-level semantic interactions. Based on the underlying

motion characteristics, mid-level visual representation has

become increasingly popular, as it can extend the semantic

description of visual content from feature-level to object-level,

focusing on the speciﬁc connections and interactions between

the objects. Therefore, constructing mid-level representations

of crowds for exploiting the richness of interactions can lead

to breakthroughs in a wide range of applications.

To achieve this goal of automatically identifying the in-

teraction of crowd behaviors, one emerging and challenging

task here is the detection of abnormal crowd behaviors. Under

such a circumstance, crowd behavior is usually modeled as

a quantitative result of semantic representation, which is

the basis of analyzing and understanding the abnormality.

Given a video clip, the abnormal detection models motion

consistency among individuals, labeling the ones that are

signiﬁcantly inconsistent with others as “abnormal”. To this

end, extensive work has been proposed towards accurate and

robust abnormality discovery, ranging from motion feature

representation to inconsistency model deﬁnition, such as MP-

PCA [1], low-level statistics [2], Dynamic Texture [3], MR-

F [4], and sparse representation [5]. Rather than classifying

the overall “inconsistency”, work in [5], [3], [6], [7] also

focus on detecting and locating the “local” inconsistency or

the “selective” mechanism of the crowded scene.

A. The Issues

Detecting abnormal crowd behaviors, by analyzing low level

semantics of crowds, is still far from real world applications.

Despite the complex models extensively studied in the litera-

tures, key issues remain in the implementation of robust yet

accurate feature representation speciﬁed for crowd behaviors,

due to:

• The traditional motion features in the existing works [2],

[8], such as optical ﬂow, Space-Time Interest Point

(STIP) [9], as well as spatial-temporal volume [8] are

incapable of characterizing motion inconsistency among

individuals. For instance, basic representation elements

are supposed to have the same status (such as scale and

motion velocity) without considering dynamic changes

and perspective effects of the scene.

• While a few recent works [10], [11], [12] attempt to

model crowd interactions, such modeling still relies on

the low-level visual and/or motion statistics among indi-

viduals, without regard to their higher-level semantics and

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/TCSVT.2014.2355711, IEEE Transactions on Circuits and Systems for Video Technology

structures of crowds, such as whether the group behavior

is actually “inconsistent” from the social or behavior

point of view. In other words, the social and commu-

nity semantics from both microscopic and macroscopic

support [12] are missing in these works, which is in turn

of high importance towards precise abnormity deﬁnition.

• Even through there are some recent works leveraging

social interaction theory [13], the modeling of social

behaviors is still far from satisfactory. The expression-

s of interactions between individuals lack intermediate

features, which ignores the social characteristics of the

crowd working on the individuals. In real-world scenar-

ios, traditional approaches can not efﬁciently give realistic

interactive clues for complex scenes. The principle to

overcome such issues may resort to boosting the feature

representation from low-level to middle-level, for exam-

ple attributes or concepts [14], [15], [16], as recently

emerging trend in object and event detection [17], [18]

in the community.

In this work, we investigate the feasibility of modeling

the social interactions among individuals for abnormal crowd

behavior detection. Our innovation lies in an attribute-level

modeling for social interaction behaviors, by which two groups

of social attributes are proposed, i.e. social disorder attribute

and congestion attribute. Together, they form a novel social

attribute-aware force model to exploit the richness of interac-

tions. Such a model can accurately and robustly express com-

plex interactions among individuals. Given such attributes at

hand, abnormality detection can be achieved at either a global

or local scale. In both cases, the proposed attributes can be

integrated with most existing abnormal detection frameworks

to achieve state-of-the-art performance, as we have extensively

tested in our experiments.

B. Related work

Modeling social interaction plays an important role in de-

scribing group behavior, and contributes to abnormality mod-

eling in crowded scenes. Social interaction modeling considers

the important point that the individuals in crowded scenes

move purposefully, forming interactions with others and affect-

ing the of each other’s trajectories. It also improves traditional

dynamic models for crowd simulation, which considers that

people have their future destinations and take actions to adjust

their trajectories [19], [12]. There are three main perspectives

for modeling and analyzing social interactions: 1) Macroscopic

models, which deal with the group density and velocity instead

of determining the motion of individuals; 2) Microscopic

models, which concern the pedestrians’ motivation in the

movements and deal with the group density/velocity or the

individual motion, for instance, the Social Force Model pro-

posed in [13] that investigates the motion dynamics; 3) Hybrid

methods that take the above two types of models into a

whole account. Both group and individual statistics are fused

to analyze the crowd behaviors, in which external and self

organization factors are exploited [20].

As shown in Figure 1, external organization aims to de-

ﬁne the belief priors such as obstacles, positions or scene

(a) (b) (c) (d)

Fig. 1. External and self organization factors of crowd behavior. The

crowd behaviors are inﬂuenced by the collective and interactive effects. The

collective effects are formed by individuals sharing similar moving patterns

such as behaviors of organisms of school ﬁsh, ﬂocking birds in biology (d).

It usually characterizes crowd behaviors at a macroscopic level which are

regularized by external environment like scene structures, obstacles as (a)(c).

The interactive effects characterizes crowd behaviors at the microscopic level.

The behaviors are caused by interactions among individuals within social

groups which are closely relevant with crowd density (b).

structures. Self organization refers to exploiting the collective

and interactive effects, which have been widely used for

understanding crowd behaviors [21], [22]. In the following,

we give a brief review for modeling the social behavior from

collective and interactive aspects for abnormal crowd behavior

detection.

Collective Effect: Collective effect offers a macroscopic

view in terms of modeling similar motion patterns in crowd be-

havior, which can be treated as regular patterns inﬂuenced by

scene structures and environmental conditions. For instance,

many works on global motion pattern learning are capable

of capturing the collective effect. Ali et al. [23] proposed a

method based on ﬂuid dynamics to simulate crowd behaviors,

which required precise segmentation of motion ﬂow. Lin et

al. [24] proposed extraction of robust ﬂow patterns using

potential geometry migration process and algebraic expression

model. Recent works in modeling global motion patterns

utilize Hierarchical Bayesian models [25], [26] by mining the

co-occurrences of moving pixels to learn behavior patterns.

Bellomo et al. [27] propose a kinetic theory model to show

how the dynamics at the micro-scale is transferred to collective

behaviors. Zhou et al. [4], [28] use Markov Random Field

(MRF) as hypothesis to establish the connection model for

track segments, which is then combined with topic model to

cluster group behavior.

Interactive Effect: Analyzing the interaction between in-

dividuals typically relates to the microscopic-level description

on crowd behavior. It can be used to describe different interac-

tions, social groups or leadership for individuals. Helbing [13]

proposed a social force model to simulate pedestrians which

are forced by surrounding environment and people. This model

is extended by Mehran [10] for anomaly detection in crowd

scenes. The works in [12], [29] utilize social behavior of

individual interaction model to analyze the dynamic behaviors

of pedestrians. Agent models inﬂuenced by personal, social

and environmental factors are constructed in [19], which effec-

tively estimated pedestrians purposes and social relationships.

Compared to the previous works, we encode collective effect

into the local interaction of crowd behavior and integrate social

attributes to reveal the richness of interactions from both the

macroscopic and microscopic viewpoints.

http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI

10.1109/TCSVT.2014.2355711, IEEE Transactions on Circuits and Systems for Video Technology

Abnormal Crowd Detection: Detecting abnormal behav-

iors from the crowd is a longstanding research topic which

has attracted extensive research in the past. Based on the

application scenarios, the task of abnormal crowd behavior de-

tection can be categorized into two-fold, i.e., Local Abnormal

Event (LAE) and Global Abnormal Event (GAE) [5]. LAE

aims at detecting the local behaviors which are different from

the neighborhoods. For instance, Kim et al. [1] modeled the

activity pattern with MP-PCA on the local optical ﬂow and

utilized Markov Random Field to localize anomalies. Adam

et al. [2] employed histograms to measure the probability of

local optical ﬂow pattern. In [8], spatial-temporal gradients

are extracted by Kratz et al. to ﬁt a Gaussian model, and

a coupled Hidden Markov Model (HMM) is used to detect

abnormalities in crowded event. By modeling motion patterns

in local area, dense activity and intrinsic structure of crowd

can be exploited. Mahadevan et al. [3] proposed a dynamic

texture model to jointly model the appearance and dynamics of

the crowded scene. In addition, detecting temporal and spatial

anomalies can be formulated as detecting crowd saliency in

mixtures of dynamic textures [3]. Thida et al. [30] employed

spatiotemporal Laplacian Eigenmap by constructing a pairwise

graph to detect and localize the abnormal regions considering

the visual context of multiple local patches.

The goal of the GAE is to detect whether the whole scene

is abnormal. Mehran et al. [10] adopted the social force model

and particle advection scheme to detect the abnormal crowd

behavior by analyzing the interaction force. Based on the

Lagrangian framework of ﬂuid dynamics, a streakline repre-

sentation [31] is proposed to enhance the representation of

abnormal event using the Helmholtz decomposition theorem.

Cui et al. [11] explored the relationships between the states

of the subjects using interaction energy potentials. The above

approaches are successfully carried out in crowd behavior

modeling for abnormal detection.

C. Our contributions

In this paper, we introduce a social attribute-aware force

model to achieve robust, efﬁcient and interactive crowd be-

havior modeling. Our ﬁrst contribution is to take into account

the scene scale information by proposing a fast novel scene

scale estimation scheme which captures the scene perspective

to better infer crowd characteristics. Given such a scale esti-

mation scheme, it is then feasible to partition the foreground

movements from the background, and extract scale attribute

and density properties of a crowd.

Our second contribution is to revisit the abnormal detection

task from both microscopic and macroscopic viewpoints us-

ing a statistical motion analysis approach. Subsequently, we

introduce two social attributes to formulate the interaction

of crowd behavior. Both the attributes offer the richness of

interaction which reﬂect the social evidence. Our attributes

are constructed by quantitative measurement of the motion

features, which provides socially motivated clues in expressing

the interaction effect.

Our third contribution is to emphasize the social semantic

inﬂuence on the crowd interaction behaviors by using the

proposed social attribute-aware force model, which can also

be extended to an online process via an effective attribute

weight fusion algorithm. This is achieved by representing each

attribute as a grid weighted map with adaptive pooling. We

jointly integrate different weighted maps in an online manner,

which ensures the processing efﬁciency.

Justiﬁcation to Existing Model: In contrast to a recent

Bayesian model [32] which also utilizes different ﬂow ﬁeld

attributes to characterize crowd motion, our model differs in

the following aspects: 1) Social attribute-aware force model

is embedded with completely different scale information from

Bayesian model. We use the combination of local scale and

global perspective, which naturally achieves more robustness

to the scale changes of scene than the regular grid utilized by

Bayesian models. 2) Our model copes with crowd interaction

by social characteristics in an online fusion manner such

that abnormal crowd patterns of various types and density

in this case could be identiﬁed. In comparison, the recent

Bayesian model updates the probability density of optical ﬂow

in a Bayesian framework, which only employs concepts of

divergent centers to detect limited escape events. Additionally,

Bayesian model is sensitive to optical ﬂow estimation and

works on low or medium crowd to identify escape patterns. 3)

The two methods have different objectives i.e., our method

aims at exploring the interaction of crowd motion for all

types of abnormal detection, while the Bayesian model in [32]

focuses on modeling the crowd motion only in the cases of

escape.

The rest of the paper is organized as follows: in Section II,

we describe the social force estimation and propose the

social attribute hypotheses. Section III details social attribute-

aware force model. Experimental results on global and local

abnormal detection are reported in Section IV. We conclude

this paper in Section V.

II. ESTIMATING CROWD INTERACTION FORCE

Figure 2 illustrates the ﬂow chart of the proposed method.

We ﬁrst place a grid of particles in the frame as individuals and

compute the conventional social force by particle advection

scheme. Then, we develop a Social Attribute-aware Force

Model (SAFM), which aims to enrich scene scale and social

attributes on group interactions. We calculate the interaction

force and obtain social attribute-aware force maps by an online

fusion algorithm. Finally, we utilize Bag-of-Words represen-

tation for Global Abnormal Event detection and abnormal

map approach for Local Abnormal Event detection as well

as localization with corresponding classiﬁer. We describe the

social attributes to reﬂect the interaction characteristics, and

to mode attractive and repulsive phenomenon under the social

attribute hypotheses. Our SAFM is thus capable of describing

the self-organized interactive effects of the pedestrian behav-

iors reliably.

A. Particle advection

For the ideal case, it is best to track and analyze the

motion of all the individuals to describe the crowd. However,

tracking all the pedestrians in a dense crowded scene presents

剩余13页未读，继续阅读

评论收藏

内容反馈

weixin_38728464

粉丝: 1
资源: 966

Social Attribute-aware Force Model: Exploiting Richness of Inter...

最新资源

Social Attribute-aware Force Model: Exploiting Richness of Inter...

abnormal crowd behavior detection using social force model

Exploiting relationship between attributes for improved face verification

An estimation model for social relationship strength based on users' profiles, co-occurrence and interaction activities

Laravel开发-entity-attribute-value-model

Apache OFBiz Datamodel 3

Relation-Aware Pedestrian Attribute Recognition with.pdf

颜色分类leetcode-Person-Attribute-Recognition-MarketDuke:在PyTorch中为行人属性识别任务

attribute-depends-calculator：自动计算ActiveRecord的depends属性的集合

DukeMTMC-attribute-master

Python-AttributeAwareAttentionModel用于细粒度表示学习的属性感知注意模型

xslt元素使用说明

twig模板手册

HLS-WWWDC-2017

certbot-dns-aliyun:certbot dns插件，可使用阿里云获取证书

sequelize-attribute-roles:属性白名单和角色列表的黑名单

langchain-chatchat在window上使用cpu运行Qwen-1-8B-Chat时遇到ERROR: object

CAD制图快捷键.pdf

颜色分类leetcode-qgis-raster-attribute-table-plugin:QGIS栅格属性表插件

magento2-customer-attribute-extension:Magento 2客户属性扩展

Python中英文词汇对照表(2020).pdf

eloquent-attribute-value-prediction:使用机器学习预测Laravel Eloquent模型的属性值！

Adversarial Attribute-Text Embedding for Person Search with Natural Language Que

2004-CAD快捷键.pdf

Circle &amp; Search: Attribute-Aware Shoe Retrieval

node-sequelize-starter:节点JS，Sequelize，Express和MySQL

Pedestrian-Attribute-Recognition:基于VesPA，WPAL的一些行人属性识别方法。 PA-100K和RAP进行培训

magento2-attribute-description:Magento 2模块，用于向产品属性添加描述

Ciphertext-Policy Attribute-Based Encryption

最新资源

Circle & Search: Attribute-Aware Shoe Retrieval