智能交通监控中基于轨迹的异常行为检测方法
91 浏览量
更新于2024-08-27
收藏 710KB PDF 举报
“Trajectory-based Anomalous Behaviour Detection for Intelligent Traffic Surveillance”
这篇研究论文探讨了基于轨迹的异常行为检测方法在智能交通监控中的应用。作者包括Yingfeng Cai、Hai Wang、Xiaobo Chen和Haobin Jiang，分别来自中国江苏大学的汽车工程研究所和汽车与交通工程学院。
在智能交通系统中，异常行为检测是关键任务之一，旨在识别可能违反规则或可能引发事故的行为。这篇论文提出了一种高效的方法，该方法利用车辆轨迹分析来实现这一目标。该框架由两个主要部分组成：轨迹模式学习模块和在线异常检测模块。
在轨迹模式学习模块中，研究者采用了粗到细的聚类策略。首先，根据主要流动方向（MFD）向量将车辆轨迹粗略地分组到连贯的聚类中。这有助于捕捉交通流的基本模式。然后，通过一个三阶段过滤算法进一步细化这些群体，去除不符合正常模式的轨迹。
接下来，研究者在每个粗聚类内部应用鲁棒的K-means聚类算法，以实现更精细的分类。这种方法能够识别并隔离那些与大多数车辆轨迹显著不同的“离群点”，这些离群点很可能代表异常行为。
在线异常检测模块则负责实时分析新产生的轨迹数据。一旦识别出异常行为，系统可以立即发出警报，以便于交通管理者采取相应的措施。这种实时分析能力对于提高交通安全性、减少事故发生以及优化交通流量至关重要。
论文指出，这种方法的优势在于其对大规模轨迹数据的处理效率和准确性。通过结合模式学习和在线检测，它能够在复杂的城市交通环境中有效检测异常行为，而不仅仅是依赖静态规则。
这项工作为智能交通监控提供了一种创新的解决方案，通过深入分析车辆运动轨迹，可以提升城市交通系统的智能化水平，有助于预防交通事故，提升道路安全，并为未来的智能交通管理系统设计提供了有价值的理论依据。
Trajectory-based anomalous behaviour
detection for intelligent traffic surveillance
ISSN 1751-956X
Received on 18th September 2014
Revised on 4th February 2015
Accepted on 21st February 2015
doi: 10.1049/iet-its.2014.0238
www.ietdl.org
Yingfeng Cai
1
, Hai Wang
2
✉
, Xiaobo Chen
1
, Haobin Jiang
2
1
Automotive Engineering Research Institute, Jiangsu University, Zhenjiang 212013, People’s Republic of China
2
School of Automotive and Traffic Engineering, Jiangsu University, Zhenjiang 212013, People’s Republic of China
✉ E-mail: wanghai1019@163.com
Abstract: This study proposes an efficient anomalous behaviour detection framework using trajectory analysis. Such
framework includes the trajectory pattern learning module and the online abnormal detection module. In the pattern
learning module, a coarse-to-fine clustering strategy is utilised. Vehicle trajectories are coarsely grouped into coherent
clusters according to the main flow direction (MFD) vectors followed by a three-stage filtering algorithm. Then a robust
K-means clustering algorithm is used in each c oarse cluster to get fine cl assification by which the outliers are
distinguished. Finally, the hidden Markov model (HMM) is used to establish the path pattern within eac h cluster. In the
online detection module, the new vehicle trajectory is compared against a ll the MFD distributions and the HMMs so
that the coherence with common motion patterns can be evaluated. Besides t hat, a real-time abnormal detection
method is proposed . The abnormal behaviour can be d etected when happening. Experimental results illustrate that the
detection rate of the proposed algorithm is close to t he state-of-the-art abnormal event detection systems. In addition,
the proposed system provides the lowest false detection rate among selected methods. It i s suitable for in telligent
surveillance applications.
1 Introduction
Intelligent visual surveillance is an important facet of machine vision
research. One of the most important goals of visual surveillance
systems is to track objects and further analyse their behaviours in
order to detect anomalies or predict potential anomalous
behaviours before occur [1, 2]. In traffic applications, the detection
of vehicle anomalous behaviours (such as wrong turn and illegal
parking), which can improve the road safety, is one main concern
of accident management department [3–5]. Nowadays, the
increasing number of cameras generates large amounts of video
data, making it infeasible for a human to accurately process.
Behaviour analysis systems can be employed to filter out relevant
data, focusing attention where it is needed most. By these reasons,
behaviour analysis research in the field of intelligent transportation
systems became very significant.
In most cases, vehicles in the traffic scene do not move randomly.
Instead, they usually follow specific motion patterns. So the
anomalous behaviours can be detected based on the learning of
normal trajectory patterns [6–8]. The general overview of the
trajectory-based abnormality detection system is given in Fig. 1.
The system is divided into two modules.
In the first module, trajectories captured from video sources are
filtered to remove the noise caused by wrong tracking. Then
trajectories are clustered to determine the common route models,
where clustering serves as a pre-processing step for further
modelling. By modelling of the common routes, normal vehicle
behaviours can be obtained.
In the second module, the new trajectories are extracted by vehicle
detection and tracking. Those that do not belong to any normal route,
can then be considered as anomalies.
1.1 Related work
The key tasks of trajectory analysis include trajectory clustering and
modelling. Jung et al. [3] proposed a method for event detection
based on trajectory clustering. In the training period, captured
trajectories were grouped into coherent clusters for which
four-dimensional (4D) motion histograms were built. In the test
period, each new trajectory was compared with the 4D histograms
of all the clusters. Piciarelli et al. [9] proposed a support vector
machine (SVM)-based anomalous event detection method. They
pointed out that the outliers in the training samples were the
anomalous events. On the basis of that, the trajectories were
clustered with a single-class SVM and the outliers were detected
by the geometric considerations in SVM feature space. Morris and
Trivedi [10] proposed a three-stage hierarchical framework for
unsupervised visual scene description and live analysis. Such
framework indicated important image points, connects them
through spatial routes, and probabilistically modelled
spatio-temporal dynamics with a hidden Markov model (HMM).
Once these path definitions were established, abnormal trajectories
were detected and future intent was predicted. A survey about
vision-based trajectory learning was also done by Morris and
Trivedi [11]. It presented common learning techniques in the
literature to extract automated scene modelling, automated activity
analysis and object interactions. Karavasilis et al. [12] presented a
model-based clustering method by trajectories’ Harris corners.
Clustering is achieved through a sparse regression mixture model
that embodies efficient characteristics in order to handle
trajectories of variable length, and to be translated in measurement
space. Hu et al. [13] used a hierarchic clustering method based on
spatial and temporal information of trajectories. In the work, each
motion pattern is represented with a chain of Gaussian
distributions. On the basis of the learned statistical motion
patterns, statistical methods are used to detect anomalies and
predict behaviours. Besides that, Hu et al. [14] also proposed a
Dirichlet process mixture model (DPMM) for trajectory clustering
and modelling. Each cluster of trajectories was learned by a
time-sensitive DPMM, and then, a parameterised index is
constructed for each cluster. Makris and Ellis [15] proposed a
method for labelling the scene with topological information, which
is then used within a Bayesian approach to detect anomalous
trajectories. Moreover, Bashir et al. [16], Hanwell and Mirmehdi
[17], Cohen et al. [18], Li et al. [19], Singha et al. [20] also made
IET Intelligent Transport Systems
Research Article
IET Intell. Transp. Syst., 2015, Vol. 9, Iss. 8, pp. 810–816
810
&
The Institution of Engineering and Technology 2015
下载后可阅读完整内容，剩余6页未读，立即下载
2019-08-05 上传
2020-08-15 上传
2021-02-22 上传
2021-02-22 上传
2011-06-07 上传
2013-09-22 上传
2021-05-26 上传
2021-05-23 上传
2021-06-04 上传
weixin_38723192
- 粉丝: 8
- 资源: 870
我的内容管理
展开
