智能视频监控中的算法研究及一种新的背景差分模型
摘要
计算机智能监控识别技术是计算机视觉领域中一个新兴的研究方向,由于其应用广泛得到了越来越多的关注.它综合运用了计算机科学、计算数学、模式识别、图像处理等多个方向学科的知识.智能视频监控就是希望在监控过程中不需要人为干预情况下,利用计算机视觉和视频分析的方法对摄像机拍录的图像序列进行自动分析,实现对动态场景中目标的定位、识别和跟踪,并在此基础上分析和判断目标的行为.计算机智能视频监控技术可以应用于小区安全监控、火情监控、交通违章、流量控制、军事和银行、商场、机场、地铁等公共场所安全防范等。
目前,计算机智能视频监控在理论和应用上都面临着很多难题,在国内外都有大批学者投身于对该领域进行研究和探索,取得比较丰富的成果.但仍有很多问题未能很好的解决,诸如:在基于灰度图像的监控系统中由于存在阴影等虚假部分,而导致监控失误;还有一些监控系统在训练集有运动物体的情况发生时监控会发生失败等等.因此,进一步研究这些问题,对理论分析与实际应用都有必要.本文一方面总结了计算机智能视频监控系统中,特别是变化检测阶段中关于背景模型和前景检测部分,比较经典的算法与思想,另一方面,介绍了作者同导师和同学一起开发的一套智能视频监控系统,给出了一种能够分割出阴影等虚假变化部分的快速模型计算方法,和能够解决在训练集有运动物体存在的背景差分模型算法.
Computer intelligent video surveillance system(CIVSS) is one of the new arising high-tech application fields. It includes knowledge of computer science application mathematics,pattern analysis,image engineering,etc. CIVSS can automatically analysis image sequence with the methods of computer vision and video analysis. The system can real-time detect, recognize,and track moving objects in a special environment. Furthermore,it can also analysis and judge the behavior of objects. Now it is used in monitoring, fire monitoring,traffic flow and peccancy, and the surveillance for bank, shopping, parking lots aerodrome, underground, etc.
As intelligent video monitoring has such broad application , so we go deep into the subject .We have solved the problems caused by the shadow and the obstruction ,so we can accurate segment objects . we expend the premise,and now in our algorithm it is not necessary to demand that there is no active objects in the training time . we also implement our algorithm with computer procedure ,and get a satisfied result .
In chapter 1 we introduce some of the results such as the VSAN system of the United States Military,the ADVISOR system developed by EU,the W~4 systems developed by Haritaoglu,etc .Based on these successful system ,we divide the develop flow into 4 stages : stage 1 Image data acquisition , stage 2 change detection , stage 3 target classification , stage 4 behavior understanding .We introduce the main work we have to do in each stage . And the most important part is the change detection stage .
In the second chapter we summarizes the classical algorithm of the gray image CIVSS.It has two models:one is Non-difference background model,the other is difference background model.One of the non-difference background is Optical Flow method.There is assumptions that the value of the image don't change a lot during a short period of time.Use this assumptions ,we can make a basic equation of optical flow,then compute the field of optical flow.This method can find out the object when the camera is moving .However,this method is complex and sensitive to noise.If it is used in CIVSS,some special equipments are needed in the system .Difference background model is used when the camera is fitted.The main idea is:first we established background model, and then judge changes between the current frame and background model,find out the moving objects.Background subtraction ,frame difference,W~4,mixed Guassian ,kalman filter are always used in the difference background model .The mixed guassian used a number of Guass model to establishes background.The method can be used in multi-modal circumstances of motion detection. However, due to the establishment of a number of Gaussian models, the parameters of the background can't not be updated correctly.So this method can't be used in real-time system.The method of background subtraction and W~4 need a period of time for training .During this period of time,they choose the best background .This kind of algorithm is simple and fast. But if the object exist during the training time, there will be error in the system .
The gray image CIVSS can't find out the false part,like shadow ,which is made by foreground objects. And the same problem is also exists in linear RGB color space . The third chapter presents a new non-linear color model. RGB component will be converted into color components and brightness components.To simplify this new color model,we introduce the concept of Allowed Changing Cone, by the judgement of if pixel is in the Allowed Changing Cone, and determine future goals and greatly reduce the amount of computations. In the third chapter, in the background substraction, W~4 methods, if there exist foreground in the training phase , it will detect ghost phenomenon, the article presents a template through the establishment of two temporary and, in conjunction with frame difference, we determined the changes in the new campaign algorithm. The method combines the two advantages of background and frame difference.Remove the false part through Bool operation of the two templates.The algorithm is simple and easy to realize .We developed the color image intelligent video monitoring system by ourselves .In the third chapter, more details about our system are involved in the changing detection stage process,image pre-processing and post-processing, background updates, and has done introduced.
In the fourth chapter,we summarizes some image features widely used in object classification .For example:area, perimeter,gravity and so on.But these features aren't nature,because they can change with position of foreground.In this condition we prefer the concept of foreground contour , the nature of this feature is more comparison. And also we introduce a method that active contour to find out the foreground contour.
In chapter 5 we showed the experimental results of our system .
目前,计算机智能视频监控在理论和应用上都面临着很多难题,在国内外都有大批学者投身于对该领域进行研究和探索,取得比较丰富的成果.但仍有很多问题未能很好的解决,诸如:在基于灰度图像的监控系统中由于存在阴影等虚假部分,而导致监控失误;还有一些监控系统在训练集有运动物体的情况发生时监控会发生失败等等.因此,进一步研究这些问题,对理论分析与实际应用都有必要.本文一方面总结了计算机智能视频监控系统中,特别是变化检测阶段中关于背景模型和前景检测部分,比较经典的算法与思想,另一方面,介绍了作者同导师和同学一起开发的一套智能视频监控系统,给出了一种能够分割出阴影等虚假变化部分的快速模型计算方法,和能够解决在训练集有运动物体存在的背景差分模型算法.
Computer intelligent video surveillance system(CIVSS) is one of the new arising high-tech application fields. It includes knowledge of computer science application mathematics,pattern analysis,image engineering,etc. CIVSS can automatically analysis image sequence with the methods of computer vision and video analysis. The system can real-time detect, recognize,and track moving objects in a special environment. Furthermore,it can also analysis and judge the behavior of objects. Now it is used in monitoring, fire monitoring,traffic flow and peccancy, and the surveillance for bank, shopping, parking lots aerodrome, underground, etc.
As intelligent video monitoring has such broad application , so we go deep into the subject .We have solved the problems caused by the shadow and the obstruction ,so we can accurate segment objects . we expend the premise,and now in our algorithm it is not necessary to demand that there is no active objects in the training time . we also implement our algorithm with computer procedure ,and get a satisfied result .
In chapter 1 we introduce some of the results such as the VSAN system of the United States Military,the ADVISOR system developed by EU,the W~4 systems developed by Haritaoglu,etc .Based on these successful system ,we divide the develop flow into 4 stages : stage 1 Image data acquisition , stage 2 change detection , stage 3 target classification , stage 4 behavior understanding .We introduce the main work we have to do in each stage . And the most important part is the change detection stage .
In the second chapter we summarizes the classical algorithm of the gray image CIVSS.It has two models:one is Non-difference background model,the other is difference background model.One of the non-difference background is Optical Flow method.There is assumptions that the value of the image don't change a lot during a short period of time.Use this assumptions ,we can make a basic equation of optical flow,then compute the field of optical flow.This method can find out the object when the camera is moving .However,this method is complex and sensitive to noise.If it is used in CIVSS,some special equipments are needed in the system .Difference background model is used when the camera is fitted.The main idea is:first we established background model, and then judge changes between the current frame and background model,find out the moving objects.Background subtraction ,frame difference,W~4,mixed Guassian ,kalman filter are always used in the difference background model .The mixed guassian used a number of Guass model to establishes background.The method can be used in multi-modal circumstances of motion detection. However, due to the establishment of a number of Gaussian models, the parameters of the background can't not be updated correctly.So this method can't be used in real-time system.The method of background subtraction and W~4 need a period of time for training .During this period of time,they choose the best background .This kind of algorithm is simple and fast. But if the object exist during the training time, there will be error in the system .
The gray image CIVSS can't find out the false part,like shadow ,which is made by foreground objects. And the same problem is also exists in linear RGB color space . The third chapter presents a new non-linear color model. RGB component will be converted into color components and brightness components.To simplify this new color model,we introduce the concept of Allowed Changing Cone, by the judgement of if pixel is in the Allowed Changing Cone, and determine future goals and greatly reduce the amount of computations. In the third chapter, in the background substraction, W~4 methods, if there exist foreground in the training phase , it will detect ghost phenomenon, the article presents a template through the establishment of two temporary and, in conjunction with frame difference, we determined the changes in the new campaign algorithm. The method combines the two advantages of background and frame difference.Remove the false part through Bool operation of the two templates.The algorithm is simple and easy to realize .We developed the color image intelligent video monitoring system by ourselves .In the third chapter, more details about our system are involved in the changing detection stage process,image pre-processing and post-processing, background updates, and has done introduced.
In the fourth chapter,we summarizes some image features widely used in object classification .For example:area, perimeter,gravity and so on.But these features aren't nature,because they can change with position of foreground.In this condition we prefer the concept of foreground contour , the nature of this feature is more comparison. And also we introduce a method that active contour to find out the foreground contour.
In chapter 5 we showed the experimental results of our system .
引文
[1] Robert T.Collins etc. A System for Video Surveillance and Monitoring.
[2] Haritaoglu I, Harwood D and Davis L. W 4: W ho, W hen, W here, W hat: A real rime system for detecting and tracking peope(C). Third International Conference on Automatic Face and Gesture, Nara,(1998)
[3] Horn B, Schunch B. Determining optical flow [J].Artificial Intelligence,
[4] Nagel H H. Constrains for the estimation sequences[A].In:Proceeding IJ-CAI' -83 of displacement vector Karlsruhe, Germany, fields from image(1983)
[5] Tretiak O,Pastor L .Velocity estimation from image sequences with differential operators.In:Proceedings Internation Conference on Pattern Recognition.Montreal, Canda.(1984)
[6] Otte M, Nagel H H. Estimation of optical flow based on higher- order spatiotemporal derivatives in interlaced and non-interlaced image sequences.Artifical Intelligence. (1995)
[7] Nils T Sibel.Design and Implementation of People Tracking Algorithms for Visual Survelliance Applications.Submitted in partial fulfilment of the requirements for the degree of Doctor of Philosophy.(2003)
[8] Cutler R ,Davas L.View-based detection.Proceedings of International Conference on Pattern Recognition. (1998).
[9] Lipton A,Fujiyoshi H.Moving target classification and tracking from realtime video.Proceedings of IEEE Workshop on Applications of Computer Vision.
[10]Lisa M.Brown.View Independent Vehicle/Person Classification.Image Processing and Computer Vision.(2004)
[11]Stauffer C,Grimson W.Adaptive background mixture models for real-time tracking.Proceedings of IEEE International Conference on Computer vision and Pattern Recognition.
[12]Stauffer C,Grimson W.Learning pattern of activity using real-time tracking.IEEE Transactions on Pattern Analysis and Machine Intelligence.
[13]Thanarat Horprasert,David Harwood,Larry S,Davis.A Statistical Approach for Real-time Robust Background Subtraction and Shadow Detection.
[14]Ridder C,MunkeltO and Kirchner H.A dap tive background estimation and for background detection using Kalman filtering[C].ICRAM(1995)
[15]Linda G.Shapiro.Computer Vision.PERSON.(2001)
[16]王亮 谭铁牛.人运动的视觉分析综述 计算机学报(2002)
[17]范莹 视频监控中运动图像检测与识别技术研究 大连理工大学硕士论文
[18]章毓晋.图象工程(下册):图象理解与计算机视觉.清华大学出版社(2000)
[19]李刚 邱尚斌等 基于背景差法和帧间差法的运动目标检测方法 仪器仪表学报(2006)
[20]余胜生 肖得贵等 自适应背景抽取法 小型微型计算机系统(2003)
[21]章毓晋.图象工程(中册):图象理解与计算机视觉.清华大学出版社(2005)
[22]代科学等.监控视频运动目标检测减背景技术的研究现状与展望.中国图像图形学报.(2006)
[2] Haritaoglu I, Harwood D and Davis L. W 4: W ho, W hen, W here, W hat: A real rime system for detecting and tracking peope(C). Third International Conference on Automatic Face and Gesture, Nara,(1998)
[3] Horn B, Schunch B. Determining optical flow [J].Artificial Intelligence,
[4] Nagel H H. Constrains for the estimation sequences[A].In:Proceeding IJ-CAI' -83 of displacement vector Karlsruhe, Germany, fields from image(1983)
[5] Tretiak O,Pastor L .Velocity estimation from image sequences with differential operators.In:Proceedings Internation Conference on Pattern Recognition.Montreal, Canda.(1984)
[6] Otte M, Nagel H H. Estimation of optical flow based on higher- order spatiotemporal derivatives in interlaced and non-interlaced image sequences.Artifical Intelligence. (1995)
[7] Nils T Sibel.Design and Implementation of People Tracking Algorithms for Visual Survelliance Applications.Submitted in partial fulfilment of the requirements for the degree of Doctor of Philosophy.(2003)
[8] Cutler R ,Davas L.View-based detection.Proceedings of International Conference on Pattern Recognition. (1998).
[9] Lipton A,Fujiyoshi H.Moving target classification and tracking from realtime video.Proceedings of IEEE Workshop on Applications of Computer Vision.
[10]Lisa M.Brown.View Independent Vehicle/Person Classification.Image Processing and Computer Vision.(2004)
[11]Stauffer C,Grimson W.Adaptive background mixture models for real-time tracking.Proceedings of IEEE International Conference on Computer vision and Pattern Recognition.
[12]Stauffer C,Grimson W.Learning pattern of activity using real-time tracking.IEEE Transactions on Pattern Analysis and Machine Intelligence.
[13]Thanarat Horprasert,David Harwood,Larry S,Davis.A Statistical Approach for Real-time Robust Background Subtraction and Shadow Detection.
[14]Ridder C,MunkeltO and Kirchner H.A dap tive background estimation and for background detection using Kalman filtering[C].ICRAM(1995)
[15]Linda G.Shapiro.Computer Vision.PERSON.(2001)
[16]王亮 谭铁牛.人运动的视觉分析综述 计算机学报(2002)
[17]范莹 视频监控中运动图像检测与识别技术研究 大连理工大学硕士论文
[18]章毓晋.图象工程(下册):图象理解与计算机视觉.清华大学出版社(2000)
[19]李刚 邱尚斌等 基于背景差法和帧间差法的运动目标检测方法 仪器仪表学报(2006)
[20]余胜生 肖得贵等 自适应背景抽取法 小型微型计算机系统(2003)
[21]章毓晋.图象工程(中册):图象理解与计算机视觉.清华大学出版社(2005)
[22]代科学等.监控视频运动目标检测减背景技术的研究现状与展望.中国图像图形学报.(2006)