Anomaly Detection in Crowd Scene—中文整理

Anomaly Detection in Crowd Scene—中文整理
Anomaly Detection in Crowd Scene—中文整理

Anomaly Detection in Crowd Scene Abstract—Anomaly detection in crowd scene is very important

because of more concern with people safety in public place. This

paper presents an approach to automatically detect abnormal

behavior in crowd scene. For this purpose, instead of tracking

every person, KLT corners are extracted as feature points to

represent moving objects and tracked by optical flow technique

to generate motion vectors, which are used to describe motion.

We divide whole frame into small blocks, and motion pattern

in each block is encoded by the distribution of motion vectors

in it. Similar motion patterns are clustered into pattern model

in an unsupervised way, and we classify motion pattern into

normal or abnormal group according to the deviation between

motion pattern and trained model. The results on abnormal

events detection in real video demonstrate the effectiveness of

the approach.

人群检测异常的是非常重要的,因为更多的人在公共场所的安全关切。这提出了一种方法来自动检测异常行为人群。为了这个目的,而不是跟踪每个人,KLT角点作为特征点代表移动的物体,利用光流技术追踪产生运动矢量,来描述运动。我们的整体帧划分成小块,每一块运动模式用运动矢量分布进行编码块(To encode motion pattern, we model the distribution of all motion vectors in one block as Gaussian distribution, and define parameters of Gaussian distribution as our motion pattern descriptor.)。在无监督的方式下,类似的运动模式聚集到一个模式(on-line method),并且我们根据运动模式和训练模型的偏差把运动模式归类为正常或异常组。在真实视频检测证明在异常事件检测事件的办法有效性。

II. RELATED WORK

The basis of whether analysis every individual’s activity, there are two categories: tracking stage and holistic stage.

第二种(整体的)

[1] Carolina Garate, Piotr Bilinski, Francois Bremond, ”Crowd Event Recognition Using HOG Tracker, ”IEEE International Workshop on Performance

Evaluation of Tracking and Surveillance, 2009.

[3] N. Ihaddadene and C. Djeraba, ”Real-time crowd motion analysis,”IEEE

Intl Conf. on Pattern Recognition, Dec. 2008, pp. 14.

[5] Louis Krate, Ko Nishino, ”Anomaly Detection in Extremely Crowded

Scenes Using Spatio-Temporal Motion Pattern Models,”Computer Vision

and Pattern Recognition, 2009.

[7] Jaechul Kim and Kristen Grauman, ”Observe Locally, Infer Globally :

a Space-Time MRF for Detecting Abnormal Activities with Incremental Updates,”Computer Vision and Pattern Recognition, 2009.{They generate activity

pattern using MPPCA,then do Bayesian inference on MRF.}

[8] Ramin Mehran, Alexis Oyama, Mubarak Shah, ”Abnormal Crowd Behavior Detection using Social Force Model,”Computer Vision and Pattern Recognition, 2009.

[11] Ernesto L.Andrade, Scott Blunsden and Robert B.Fisher, ”Hidden

Markov Models for Optical Flow Analysis in Crowd,”International

Conference on Pattern Recognition, 2006.

On the basis of behavior definition, there are also two categories: explicit definition and deviation definition.

The first one

[1] Carolina Garate, Piotr Bilinski, Francois Bremond, ”Crowd Event Recognition Using HOG Tracker, ”IEEE International Workshop on Performance Evaluation of Tracking and Surveillance, 2009.

[12] S.Saxena, F.Brmond, M.Thonnat, and R.Ma, ”Crowd behavior recognition

for video surveillance,”In 10th International Conference on Advanced

Concepts for Intelligent Vision Systems, 2008

[13] N.Johnson and D.Hoggr, ”Learning the Distribution of Object Trajectories

for Event Recognition,”In Proc. of British Machine Vision Conf,

pages 583592, 1995.

The second one

[3] N. Ihaddadene and C. Djeraba, ”Real-time crowd motion analysis,”IEEE

Intl Conf. on Pattern Recognition, Dec. 2008, pp. 14.

[5] Louis Krate, Ko Nishino, ”Anomaly Detection in Extremely Crowded

Scenes Using Spatio-Temporal Motion Pattern Models,”Computer Vision

and Pattern Recognition, 2009.

[8] Ramin Mehran, Alexis Oyama, Mubarak Shah, ”Abnormal Crowd Behavior Detection using Social Force Model,”Computer Vision and Pattern Recognition, 2009.

[11] Ernesto L.Andrade, Scott Blunsden and Robert B.Fisher, ”Hidden

Markov Models for Optical Flow Analysis in Crowd,”International

Conference on Pattern Recognition, 2006.

III. MOTION DESCRIPTOR

常用来表现运动信息的技术主要有:光流(option flow)和时空梯度(Spatial-Time Gradient)

光流技术特点:表现好但计算太大,在场景系统中实时性是基本要求,所以计算每个像素的光流不现实。

时空梯度技术特点:时空梯度技术计算量小,但包含边缘信息,在人群中,

人与人重叠并且相对位置变动比较大,这样使边缘信息变得很复杂,就不适合应用于人群。

考虑两者的优缺点:我们用KLT角点作为特征点代表移动的物体,利用光流技术追踪产生运动矢量,来描述运动。(KLT角点跟踪[10] J. Shi, C. Tomasi, ”Good Features to Track,” IEEE Conference on Computer Vision and Pattern Recognition,Seattle, USA, 1994, 593-600.)

我们用大小320*240的帧,并分割成16*16的块,块大小有摄像机距离和需要获得信息多少有关。因为即使在同一个块中人群运动模式也不是单一的可以简单的用平均速度和方向来表示。为了使我们运动表示算子表现更好,我们用高斯分布对运动向量建立模型分布。(one motion pattern can be described as P(U,O), where U = (μv, μr) is mean vector: μv is mean velocity, μr is mean direction , and O = (σ2v, σ2r ) is variance vector: σ2v is variance velocity, σ2r is variance direction.

摄像头角度处理

IV. CLUSTER MOTION PATTERNS

So we clustered similar motion patterns as one pattern model M(U,O). In crowd scene, we cannot know how many kinds of movement there will be, so we do not know the number of cluster. We used an on-line cluster method like [5], which does not need the number of cluster fi.rst.

我们把相似的运动模式作为一种模式M(U,O)。在人群中,我们不能知道有多少种运动,也不知道聚类的数量。我们用一种on-line 聚类方式([5] Louis Krate,Ko Nishino, ”Anomaly Detection in Extremely Crowded Scenes Using Spatio-Temporal Motion Pattern Models,” Computer Vision and Pattern Recognition, 2009.),他不需要在开始知道聚类数量。

把接近的模式归为同一模式模型,用公式(1)更新参数:

Nk表示属于模式模型Mk的模式模型数量,Pl是新获得的模式模型。

如何测量相似度?一般用对称KL(symmetric Kullback-Leibler (KL) divergence)计算两个分布的偏差([14] S. Kullback and R. A. Leibler, ”On Information and

Suf.ciency,”)。

在一些应用中可以加入些权重系数,比如有是运动方向比运动速度更重要,我们就加大运动方向的权重。为解决这个问题,我们定义偏差测量公式为:

如何设定见文献。

相关主题
相关文档
最新文档