摘要
智能视频监控系统具有广阔的应用和开发前景,在该领域研究中,对图像的运动的目标检测和背景分析是当前计算机视觉领域的一个重要研究方向。运动目标检测的目标是在视频序列图像中将运动变化目标区域从背景图像中提取出来,以此跟踪和分析运动目标的存在,为后期的智能预测和视觉处理打下基础。但由于背景图像的动态变化,理想不变的背景模型往往是很少存在的。如何解决对复杂场景下的背景建模,获取实时的背景更新以分离背景和前景,同时对运动目标进行跟踪,这是当前研究的热点,也是本文的主要工作内容。本文主要从以下两个方面进行阐述：
(1)本文介绍了多种常规背景建模方法,并进行了分析和比较。首先介绍了背景减法基本理论和典型的背景建模方法以及详细的推导,非线性贝叶斯预测理论,卡尔曼滤波理论和混合高斯模型的建立,对混合高斯算法作了详细推导,指出参数密度估计理论的不足。简述了密度估计量的基本性质以及密度估计通用表达式,介绍常用的直方图密度估计方法以及核密度估计方法。分析了影响估计结果的因素并给出了不同情况下的密度估计结果,指出了直方图密度估计的优缺点并且分析了核密度估计的渐近收敛特性。同时还简要介绍和分析了其他两种密度估计方法,即K近邻法和基函数展开法。
(2)提出了一种稳定的非参数估计背景建模算法,对影子的检测和抑制错误检测方面进行了实验和数据分析。基于前面的分析和比较,在实验中,我们将核密度估计方法与高斯混合方法进行了简单的比较,重点是论证核密度估计算法的有效性。在跟踪算法中介绍了均值偏移算法的理论知识并对其算法进行了简要的概括性阐述,详细介绍了均值偏移算法中各个关键性的变量的选取原则,讨论了在多维特征空间的概率密度估计中均值偏移算法的理论推导。在选定好关键性变量的基础上描述了均值偏移理论的计算步骤,证明了均值偏移算法的收敛特性,并运用了一个相似性度量函数和统计学中的Bayes误差进行精确地样本估计和评价,以期获取更准确的目标定位,并通过实验验证了均值偏移算法的有效性。
Intelligent video surveillance system is widely application and development prospects, in this research field, the analysis on movement of image and analysis of background is an important research direction. The ultimate purpose of moving target detection in video and image sequences is extract motion area from the background image in image sequences to track and analyze moving targets,which become an basic on intelligent forecasting and visual processing. But because of the dynamic changes of the background image,we search an ideal background model is rarely exist. We pay more attention on how find a background model on complex scenes.Then we update real-time background to separate the prospects,at the same time we track the moving target, This is the focus of research, but also the main contents of this work. This paper mainly discuss from the following two aspects:
(1)Paper introduces various conventional background modeling methods, analyzed and compared. And introduces basic theory of the background subtraction and typical background modeling method and detailed derivation, the nonlinear bayesian theory, kalman filtering theory and gaussian mixture model, detailed the gaussian mixture algorithm,at the same time I pointed out the faults of parameters density estimation. I introduced basic properties of density estimators and discussing the parameter estimation method and the density of histogram density estimation method and the kernel density estimation method.I analyzed the influence factors of the estimated and point out the density estimation under different conditions,then points out the advantages and disadvantages of histogram density estimation and analysis the kernel density estimation of the asymptotic convergence properties. Also I briefly introduced the other two density estimation method based on k-nearest neighbor and basis function expansion method.
(2)We put forward a kind of stability of non-parameter estimation, secondly we discuss the shadow detection and inhibit error detection. Based on the above analysis and comparison,in our experiments, we will compare kernel density estimation method to gaussian mixture method, The most important purpose is demonstrating the effectiveness of the kernel density estimation algorithm. In tracking algorithm,I introduced theoretical knowledge of mean-shift algorithm and described all key variable selection principle and probability density estimation algorithm of mean-shift in the multi-dimensional space. We described mean-shift algorithm calculation steps on the basis of key variables chosen,then use a similarity functions and the bayes error estimation estimated samples accurately in order to get more accurate localization of target. Through the experiments prove the effectiveness of mean-shift algorithm.
引文
[1]明安龙.智能视觉监控中目标跟踪与识别算法研究[D].北京邮电大学,2008.
[2]王彪,王成儒,王芬芬.一种改进的运动目标检测算法[A].计算机技术与应用进展·2007——全国第18届计算机技术与应用(CACIS)学术会议论文集[C],2007.
[3]T.Gevers,A.W.M.Smeulders. Color based object recognition. Pattern recognition. 32(1999)453-464.
[4]王建林,孙孟奎,杨磊,孙永奇,张士国,王志海.一种基于减背景模型的运动目标检测算法[A].2008第四届中国智能交通年会论文集[C],2008.
[5]L.Li,W.M.Huang,I.Y.H.Gu,Q.Tian.Statistical modeling of complex backgrounds for foreground object detection[J]. IEEE Trans. on Image Processing,2004, 13(11):1459-1472.
[6]Piccardi M.Background Subtraction Techniques:A Review [A]. IEEE International Conference on Systems, Man and Cybernetics [C].2004,4:3099-3104.
[7]牛君,李贻斌.基于非参数密度估计点样本分析建模的应用研究[D].山东：山东大学,2007.
[8]Ying Ming,Jingjue Jiang. Background modeling and subtractionusing a local linear dependence based cauchy statistical model[C].Proceedings of the 7th Biennial Australian Pattern Recognition Society Conference DICT.2003,1:469-478.
[9]方帅,薛方正,徐心和.基于背景建模的动态目标检测算法的研究与仿真[J].系统仿真学报,2005,(01).
[10]刘伟,张洪才,赵春晖.复杂场景目标检测技术研究及DSP系统实现[D]西安：西北工业大学,2007.
[11]李华,宋晓虹,张宁.背景模型的建立及保持方法比较[J].电脑与信息技术,2004,(01).
[12]陈柏生,陈锻生.基于归一化RGB彩色模型的运动阴影检测[J].计算机应用,2006,(08).
[13]陈振华,周锐锐,李光伟,毕笃彦.一种改进的高斯混合背景模型算法及仿真[J].计算机仿真,2007,(11).
[14]Isard.M, Blake.A.CONDENSATION-Conditional Density Propagation for Visual Tracking.International Journal of Computer Vision,1998,(29)5-28.
[15]傅军和.基于先验信息的贝叶斯统计检验和经典统计检验的比较[J].统计与决策,2008,(02).
[16]Beadle E R,Djuric P M. A fast weighted Bayesian bootstrap filterfor nonlinear model state estimation[J].IEEE Transactions onAerospace and Electronic Systems,1997,33 (1):338.
[17]Haug A J. A Tutorial on Bayesian Estimation and Tracking Techniques Applicable to Nonlinear and Non-Gaussian Processes[R].MIRTE,2005.
[18]雷明,韩崇昭,肖梅.扩展卡尔曼粒子滤波算法的一种修正方法[J].西安交通 大学学报,2005,(08).
[19]Richard O.Duda Peter E.Hart David G.Storlc著,李宏东姚天翔等译.模式分类.北京：机械工业出版社,2003.
[20]ZIVKOVIC Z. Improved adaptive Gaussian mixture model for back-ground subtraction[C].Proceedings of the International Confer-ence on Pattern Recognition. Amsterdam,Netherlands.2004,:23-26.
[21]向日华,王润生.一种基于高斯混合模型的距离图像分割算法[J].软件学,报,2003,14(07).
[22]孙香云,汪四水.基于EM算法的高斯混合密度参数估计[J].杭州师范学院学报,2005,4(05).
[23]毛燕芬,施鹏飞.高斯核密度估计背景建模及噪声与阴影抑制[J].系统仿真学报,2005,(05).
[24]Zivkovic Z. Improved adaptive Gaussian mixture model for backgroud subtraction[A].Proceedings of the 17th International Conference on Pattern Recognition[C]. Cambridge,United Kingdom.2004,:28-31
[25]徐以美,郭宝龙,陈龙.基于RGB颜色空间的减背景运动目标检测[J].计算机仿真,2008,(03).
[26]Parmeter C,Henderson D J,Kumbhakar S C. Nonparametric estimation of a hedonic price function[J] Journal of Applied Econometrics,2007,22 (03):695-699.
[27]李存华,孙志挥,陈耿,胡云.核密度估计及其在聚类算法构造中的应用[J].计算机研究与发展,2004,(10).
[28]淦文燕,李德毅.基于核密度估计的层次聚类算法[J].系统仿真学报,2004,(02).
[29]毛燕芬,施鹏飞.一种核密度估计动态场景建模算法[J].数据采集与处理,2004,(04).
[30]王典.基于混合高斯的背景建模与阴影抑制算法研究[D].西北工业大学,2006.
[31]Gray A QMoore A W. Nonparametric density estimation:toward computational tractability[C].SIAM International Conference on Data Mining.2003.
[32]李宏东,姚天翔.模式分类[M].北京：机械工业出版社,2003.
[33]Sengupta K,Burman P,Sharma R. A non-parametric approach for independent component analysis using kernel density estimation[C].Proceedings of the IEEE Computer Society Conference on Com-puter Vision and Pattern Recognition. United States,Washington,DC.2004, (2):667-672.
[34]屈文建,熊国经.非参数密度估计法比较分析及应用[J].沈阳农业大学学报,2008,(04).
[35]Prati A, Mikic I, Trived MM,et al. Detecting moving shadows:algorithms and evaluation.IEEE Transactions on Pattern Analysis and Machine Intelligence,2003,25 (7):918-923.
[36]Yaniv G,Boaz L. Rapid spline-based kernel density estimation for bayesian networks[C].Proceedings of thel7th International Con-ference on Pattern Recognition(ICPR'04). United Kingdom,Cam-bridge.2004,:700-703.
[37]Rousson M,Cremers D. Efficient kernel density estimation of shape and intensity priors for level set segmentation[C].8th International Conference on Medical Image Computing and Computer-Assisted Intervention-MICCAI2005. United States,Palm Springs,CA.2005,:757-764.
[38]倪龙强,张旭阳,高社生.核密度估计的随机加权逼近[J].西北大学学报(自然科学版),2008,(02).
[39]贾静平,张飞舟,柴艳妹.基于核密度估计尺度空间的目标跟踪算法[J].清华大学学报(自然科学版),2009,(04).
[40]李存华,孙志挥,陈耿,胡云.核密度估计及其在聚类算法构造中的应用[J].计算机研究与发展,2004,(10).
[41]熊祖光.基于核密度估计的空间聚类算法研究以及改进[D].吉林大学,2008.
[42]李庆忠,何东晓.基于聚类的背景建模与运动目标检测方法[J].计算机工程与应用,2008,(08).
[43]王萍,杨培龙,罗颖听.统计模式识别[M].北京：电子工业出版社,2004.
[44]张学荣.直方图在视频监控系统中的应用[J].电脑知识与技术,2009,(30).
[45]高韬,刘正光,张军.复杂环境下车辆阴影分割算法研究[J].交通运输系统工程与信息,2009,(02).
[46]沈志熙,杨欣,黄席樾.均值漂移算法中的目标模型更新方法研究[J].自动化学报,2009,(05).
[47]汪沁,江淑红,张建秋,胡波.提高Mean-shift跟踪算法性能的方法[J].复旦学报(自然科学版),2007,(01).
[48]王杰,王加银.Mean Shift算法的收敛性讨论[J].北京师范大学学报(自然科学版),2008,(05).
[49]彭宁嵩,杨杰,刘志,张风超.Mean-Shift跟踪算法中核函数窗宽的自动选取[J].软件学报,2005,(09).
[50]李金雷燕胡文广.多运动目标探测标记及跟踪[J].应用科技，2009,36(6).
[51]D COMANICIU,P MEER. Mean shift Arobust approach towandfeature space analysis[J].IEEE Transactions on Pattern Analysisand Machine Intelligence,2002,24 (5):603-609.
[52]明新勇；基于均值漂移和粒子滤波的目标跟踪算法研究[D].南京理工大学，2008.
[53]*士文,卢珞先,王昱.一种快速目标检测跟踪方法研究与实现[J].现代电子技术,2007,(22).
[54]汪亚明,黄文清,周海英.动态图像序列中的运动目标检测[J].计算机测量与控制,2003,(08).
[55]Research of Moving Object Tracking Based on Image Sequence[A]. Proceedings of 6th International Symposium on Test and Measurement(Volume 7)[C],2005.
[56]Li L,Huang W,Gu IYH,et al. Foreground object detection from videos containing complex background[C].Proceedings of ACM Multimedia Conference. Berkeley,CA,USA:ACM Press,2003,2(10).
[57]周敬兵.复杂背景下的目标检测与跟踪技术研究[D].南京理工大学，2007.
[58]罗小兰,刘肖琳.基于Parzen核估计的动态背景建模算法[J].微计算机信 息,2008,27.
[59]王勇.基于统计方法的运动目标检测与跟踪技术研究[D].华中科技大学,2009.