Real-time Crowd Motion Analysis
Nacim Ihaddadene and Chabane Djeraba
Computer Science Laboratory of Lille(LIFL) University of Sciences and Technologies of Lille-France nacim.ihaddadene@lifl.fr,chabane.djeraba@lifl.fr
Abstract
Video-surveillance systems are becoming more and more autonomous in the detection and the reporting of abnormal events.In this context,this paper presents an approach to detect abnormal situations in crowded scenes by analyzing the motion aspect instead of track-ing subjects one by one.The proposed approach es-timates sudden changes and abnormal motion varia-tions of a set of points of interest(POI).The number of tracked POIs is reduced using a mask that corre-sponds to hot areas of the built motion heat map.The approach detects events where local motion variation is important compared to previous events.Opticalflow techniques are used to extract information such as den-sity,direction and velocity.To demonstrate the interest of the approach,we present the results on the detection of collapsing events in real videos of airport escalator exits.
1.Introduction
Public safety is increasingly a major problem in public areas such as airports,malls,subway stations, etc.Processing the recorded videos can be exploited to present informative data to the security team who needs to take prompt actions in a critical situation and react in case of unusual events.In the last years,most of surveillance systems integrated computer vision algo-rithms to deal with problems like motion detection and tracking.However,a few have treated the problems in-volving crowded scenes due to the problematic com-plexity.
Recently,the computer vision community has started taking interest in addressing different research problems related to the scenarios involving large crowds of peo-ple.The focus so far has been on the tasks of crowd detection and tracking of individuals in the crowd.
Our approach is based on motion variation on re-gions of interests.First,a motion intensity heat map is computed during a certain period of time.The mo-tion heat map represents hot and cold areas on the basis of motion intensities.The hot areas are the zone of the scene where the motion is high.The cold areas are re-gions of the scene where the motion intensities are low. More generally,we consider several levels of motion intensity.The points of interest are extracted in the se-lected regions of the scene.Then,these points of inter-est are tracked using opticalflow techniques.Finally, variations of motions are estimated to discriminate po-tential abnormal events.The main advantage of the ap-proac
h is the fact that it doesn’t require a huge amount of data to enable supervised/unsupervised learning.It detects all events where the variations are important.
The structure of this paper will start by presenting some related works in the next section.Sections3and 4will discuss the different steps to estimate the abnor-mality of a crowdflow.In section5,a case of process-ing algorithm application is presented with an example of an escalator camera.Finally,section6presents the conclusion and some future extensions of the work.
2.Related Works
There are two categories of related works.Thefirst category is related to crowdflow analysis,and the sec-ond category is related to abnormal event detection in crowdflows.The works of thefirst category,estimate crowd density[12,13,9,11].These methods are based on textures and motion area ratio and make an inter-esting static analysis for crowd surveillance,but do not detect abnormal situations.There are also some op-ticalflow based techniques[4,6]to detect station-ary crowds,or tracking techniques by using multiple cameras[5].
reacttomotion翻译The works in the second category detects abnormal events in crowdflows.The general approach consists of modelling normal behaviors,and then estimating the deviations between the normal behavi
or model and the
978-1-4244-2175-6/08/$25.00 ©2008 IEEE
observed behaviors.These deviations are labelled as abnormal.
The principle of the general approach is to exploit the fact that data of normal behaviors are generally avail-able,and data of abnormal behaviors are generally less available.That is why,the deviations from examples of normal behavior are used to characterize abnormal-ity.In this category,[2,3]combines HMM,spectral clustering and principal component for detecting crowd emergency scenarios.The method was experimented in simulated data.[1]uses Lagrangian Particle Dynam-ics for the detection offlow instabilities,this method is efficient for segmentation of high density crowdflows (marathons,political events,...).
Our approach contributes to the detection of abnor-mal events in crowdflows by the fact that it doesn’t need learning process and training data,it also consid-ers simultaneously density,direction and velocity and focuses analysis on specific regions where the density of motions is high.
3Algorithm steps
The proposed algorithm is composed of several steps:motion heat map and direction map building,fea-tures extraction,opticalflow calculation andfinally the estimation of abnormality description measure.
3.1Motion heat map
Motion heat map is a2D histogram indicating the main regions of motion activity.This histogram is built from the accumulation of binary blobs of moving ob-jects,which were extracted following background sub-traction method[8].
The obtained map is used as a mask to define the Re-gion of Interest(RoI)for the next step of the process-ing algorithm such as“Features detection”(described later).The use of heat map image improves the qual-ity of the results and reduces processing time which is an important factor for real-time applications.Fig-ure1shows an example of the obtained heat map from an escalator camera view.The results are more sig-nificant when the video duration is long.In prac-tice,even for the same place,the properties of unusual events may vary depending on the context(day/night, indoor/outdoor,normal/peak time,...).We built a mo-tion heat map for each set of conditions.
3.2Features detection and tracking
In this step,we start by extracting a set of points of interest from each input frame.We use a mask to
define
Figure  1.Example of motion heat map
generation.(a)camera view,(b)gener-
ated motion heat map and(c)masked
view.
a region of interest.This mask is obtained from the hot areas of the heat map image.In our approach,we con-sider Harris corner as a point of interest[7].We con-sider that in video surveillance scenes,camera positions and lighting conditions allow to get a large number of corner features that can be easily captured and tracked.
Once we define the points of interest,we track these points over the next frames using opticalflow tech-niques.For this,we used Kanade-Lucas-Tomasi fea-ture tracker[10,14].After matching features between frames,the result is a set of vectors:
V={V1...V N|V i=(X i,Y i,A i,M i)} where X i and Y i are the coordinates of the feature i, A i is the motion direction of feature i and M i is the distance between the feature i in the frame t and its matched feature in frame t+1.
This step also allows removal of static and noise fea-tures.Static features are the features that moves less than two pixels.Noise features are the isolated features that have a big angle and magnitude difference with their near neighbors due to tracking calculation errors.
Images infigure2show the set of vectors obtained by opticalflow feature tracking in two different situa-tions.The left image shows an organized vectorflow. The right one shows a cluttered vectorflow due to
the collapsing
situation.
Figure2.Example of vectorflows.
4Direction map
Our actual approach considers also the“Direction Map”which indicates the average motion direction for each region of the camerafield.The camera view is di-vided into small blocks and each block is represented by the mean motion vector.The distance between average direction histogram and instant
histogram correspond-ing to the current frame increases in case of collapsing situations.
The right image of Figure3shows the mean direc-tion in each block of the viewfield of an escalator cam-era.Some tendencies can be seen.In the blue region, the motion is from top to bottom.In the yellow region, the motion is from right to
left.
Figure3.Example of block direction his-
togram:top to bottom for the left escala-
tor,bottom to top for the right one.
4.1Measuring entropy
In this step,we define a statistic measure that will de-scribe how much the opticalflow vectors are organized or cluttered in the frame.We studied a set of statistical and optical measures(variance,entropy,heterogeneity and saliency).The result is a measure M which is the scalar product of the normalized values of the following factors:
Motion area ratio:In crowded scenes the area cov-ered by the moving blobs is important compared to un-crowded scenes.This measure is also used in density estimation techniques.
Direction variance:After calculating the mean di-rection of the opticalflow vectors in a frame,we calcu-late the direction variance of these vectors.
Motion magnitude variance:Observation shows that this variance increases in abnormal situations.With one or many people walking even in different directions, they tend to have the same speed;which means a small value of the motion magnitude variance.It’s not the case in collapsing situations and panic behaviors that often engender a big value for the motion magnitude variance.
Direction histogram peaks:The calculation of vec-tors direction and magnitude variances is not sufficient. We build a direction histogram in which each column indicates the number of vectors in a given angle.The result is a histogram that indicates the direction tenden-cies,and the number of peaks in this histogram repre-sents the different directions.
4.2Deciding:Normal/Abnormal
This decision is taken in a“Static way”by compar-ing the calculated and normalized measure with a spe-cific threshold.Configuration has been necessary to es-timate the Normal/Abnormal threshold because it varies depending on camera position,escalator type and posi-tion.The decision may also be taken in a“Dynamic way“by detecting considerable sudden changes of the cluttering measure through time.
5Experimental Results
In our experiments,we used a set of real videos provided by cameras installed in an airport to monitor the situation of escalator exits.Videos are exploited to present informative data to the security team who needs to take prompt actions in the critical situation of col-lapsing.
The data set is divided into two kind of situations: normal situations and abnormal situations,with a sub-set of20videos for each type.The normal situations correspond to crowdflows without collapsing in the es-calator exits.Generally,in the videos,we have two es-calators,corresponding to two traffic ways,in opposite directions.Abnormal situations correspond to videos that contain collapsing events in escalator exists.
The original video frame size is640x480pixels and each video sequence for the different situations has more than4000frames.For the features detection and tracking we extract about1500features per frame(in-cluding static and noise features).
Figure4shows an example of a collapsing situation in an escalator exit.The variation of the cluttering mea-sure M through time is shown in the graph.The red part of the curve represents the time interval where the collapsing event happened.
From the video data set,our approach detected all collapsing events.The collapsing events,detected by the system,have been compared with collapsing events annotated manually.Till now,the result is very satisfac-tory.However,it is necessary to select the appropriate threshold and the regions of interest carefully.
Figure4.Measure variation.
6Conclusion
In this paper,we proposed a method that estimates the abnormality of a crowdflow.We defined a measure that is sensitive to crowd density,velocity and direc-tion.The method doesn’t require prior specific training processes and it has been applied to detect collapsing events in airport escalator exits.The results till now are promising on it’s robustness.
The work is still in progress,it is expected to ex-tend the estimation of the motion variations with factors such as acceleration by tracking the POIs over multi-ple frames.The contextualization of the system is also important.Introducing context information to optimize the system configuration allows to use the same under-lying detection algorithms in different locations and in an efficient way.The understanding of the system con-figuration is also important for the security team to as-sess the current situation.
Acknowledgements.This work has been performed by members of MIAUCE project,which is a6th Frame-work Research Programme of the European Union (IST-2005-5-033715),The authors would like to thank the EU for thefinancial support and the partners within the project for a fruitful collaboration.For more information about the MIAUCE project,visit:
References
[1]S.Ali and M.Shah.A lagrangian particle dynamics ap-
proach for crowdflow segmentation and stability anal-ysis.Computer Vision and Pattern Recognition,2007.
CVPR’07.IEEE Conference on,pages1–6,17-22June 2007.
[2]  E.Andrade,S.Blunsden,and R.Fisher.Hidden markov
models for opticalflow analysis in crowds.Pattern Recognition,2006.ICPR2006.18th International Con-ference,vol.1:460–463,2006.
[3]  E.L.Andrade,S.Blunsden,and R.B.Fisher.Mod-
elling crowd scenes for event detection.In ICPR’06: Proceedings of the18th International Conference on Pattern Recognition,pages175–178,Washington,DC, USA,2006.IEEE Computer Society.
[4]  B.Boghossian and S.Velastin.Motion-based machine
vision techniques for the management of large crowds.
Electronics,Circuits and Systems,1999.Proceedings of ICECS’99.The6th IEEE International Conference, vol.2:961–964,1999.
[5]  F.Cupillard,A.Avanzi,F.Bremond,and M.Thon-
nat.Video understanding for metro surveillance.Net-working,Sensing and Control,2004IEEE International Conference,vol.1:186–191,2004.
[6]  A.Davies,J.H.Yin,and S.Velastin.Crowd monitoring
using image processing.Electronics&Communication Engineering Journal,vol.7(1):37–47,1995.
[7]  C.Harris and M.Stephens.A combined corner and
edge detector.In Alvey Vision Conference,page147152, 1988.
[8]L.Li,W.Huang,I.Y.H.Gu,and Q.Tian.Foreground
object detection from videos containing complex back-ground.In MULTIMEDIA’03:Proceedings of the eleventh ACM international conference on Multimedia, pages2–10,New York,NY,USA,2003.ACM.
[9]S.-F.Lin,J.-Y.Chen,and H.-X.Chao.Estimation of
number of people in crowded scenes using perspective transformation.Systems,Man and Cybernetics,Part A, IEEE Transactions,31(6):645–654,2001.
[10]  B.Lucas and T.Kanade.An iterative image registration
technique with an application to stereo vision.Proceed-ings of the International Joint Conference on Artificial Intelligence,(1):pp674–679,1981.
[11]R.Ma,L.Li,W.Huang,and Q.Tian.On pixel count
based crowd density estimation for visual surveillance.
Cybernetics and Intelligent Systems,2004IEEE Con-ference,vol.1:170–173,2004.
[12]  A.Marana,S.Velastin,L.Costa,and R.Lotufo.Esti-
mation of crowd density using image processing.Im-age Processing for Security Applications(Digest No.: 1997/074),IEE Colloquium,pages11/1–11/8,1997. [13]H.Rahmalan,M.S.Nixon,and J.N.Carter.On crowd
density estimation for surveillance.In International Conference on Crime Detection and Prevention,2006.
[14]J.Shi and C.Tomasi.Good features to track.In IEEE
Conference on Computer Vision and Pattern Recogni-tion,pages593–600,1994.

版权声明：本站内容均来自互联网，仅供演示用，请勿用于商业和其他非法用途。如果侵犯了您的权益请与我们联系QQ:729038198，我们将在24小时内删除。