基于模糊积分和轮廓特征的步态识别
|
摘要
步态识别技术是通过研究步态图像序列中人的行走方式来识别个体的一种技术,因其可以远距离识别,具有不可侵犯性,被广泛应用在视频监控领域,近年来成为生物特征识别领域的研究热点之一。
本文在步态轮廓检测,步态特征提取和步态识别方面做了以下工作：
在步态检测方面,利用自适应高斯混合模型进行背景建模,利用背景减除法提取出步态轮廓后进行滤波处理,并与帧差法提取运动目标进行对比,实验结果表明利用背景建模法建立背景后,背景减除法提取的运动目标轮廓较准确。
在阴影检测方面,利用阴影的视觉特性来判断是否存在阴影,对存在的阴影利用多梯度分析和二值快速聚类法检测,将其与一般的检测方法进行对比,可知二值快速聚类法的检测效果较好。
在步态特征提取方面,分别提取步态的动态特征和静态特征,基于步态轮廓提取一个步态周期内的步态宽度特征和步态轮廓直方图特征,再利用傅里叶变换和小波变换的基础理论,提取步态轮廓的傅里叶描述子特征和小波矩特征,根据特征数据分布的不同,对高维特征向量分别利用不同的降维算法,例如PCA+LDA算法,LPP算法等,将高维特征降为低维特征,减少后续识别的运算量。
在步态识别方面,分别利用隐式马尔科夫模型进行步态识别和利用模糊积分对分类器进行融合识别,并将识别结果进行比较。模糊积分融合方法是根据不同特征的识别率来确定分类器的隶属度,根据隶属矩阵来确定分类器的重要性,能够更好的使分类器之间互补,从而能将步态特征在决策层进行有效的融合。
最后,利用VC调用MATLAB引擎进行步态运动目标的提取,使用中科院步态数据库20个人3个视角的步态视频进行仿真实验,证明本文所采用的方法具有优越性。
Gait recognition is a technology which detected the gait through a series of video gait images sequences. Since it can detect the object from a long distance and could not be offensive, so it is widely used in the area of video surveillance.
In this thesis, the following work about gait detection、feature extraction and gait recognition is done:
In the area of gait recognition, the adaptive Gaussian mixture model is used to construct the background model. Background reduction division is also used to extract the gait contour. The experiment results show the method in this thesis is much better compared to the frame difference method.
In the process of shadow detection the existence of the shadow is judged based on the visual characteristics of shadows. If the shadow exists, gradient analysis and binary fast clustering method are used to detect it. At last the two methods are compared with each other and the result can prove the binary fast clustering method is better.
In gait characteristics extraction, dynamic characteristic gait and static characteristics were extracted based on the Gait contour. The gait width characteristics and the gait outline histogram, Fourier descriptor features and characteristics of wavelet using the Fourier transform and basic theory of wavelet transform in a gait circle are extracted.And PCA together with LDA is used to reduce the feature's dimension. LLE and LPP are also introduced. The work can reduce the complex of the following recognition.
In gait feature recognition, fuzzy integral is used in classifiers fusion, the classifier's importance is determined according to the characteristics of the subordinate matrix after the membership is determined. Thus make classifier is complementary, can be effectively in decision-making compared to the HMM model method.
Lastly, use VC calls MATLAB engine to get the gait moving targets, use Gait database of 20 individual with three perspectives of video gait sequence to construct the simulation experiment, the experiment result in this thesis proves the method has advantages. Key words:Fuzzy Integral; Moving Object Detection; Moving Shadow Detection; Background Modeling; Gait Recognition
本文在步态轮廓检测,步态特征提取和步态识别方面做了以下工作：
在步态检测方面,利用自适应高斯混合模型进行背景建模,利用背景减除法提取出步态轮廓后进行滤波处理,并与帧差法提取运动目标进行对比,实验结果表明利用背景建模法建立背景后,背景减除法提取的运动目标轮廓较准确。
在阴影检测方面,利用阴影的视觉特性来判断是否存在阴影,对存在的阴影利用多梯度分析和二值快速聚类法检测,将其与一般的检测方法进行对比,可知二值快速聚类法的检测效果较好。
在步态特征提取方面,分别提取步态的动态特征和静态特征,基于步态轮廓提取一个步态周期内的步态宽度特征和步态轮廓直方图特征,再利用傅里叶变换和小波变换的基础理论,提取步态轮廓的傅里叶描述子特征和小波矩特征,根据特征数据分布的不同,对高维特征向量分别利用不同的降维算法,例如PCA+LDA算法,LPP算法等,将高维特征降为低维特征,减少后续识别的运算量。
在步态识别方面,分别利用隐式马尔科夫模型进行步态识别和利用模糊积分对分类器进行融合识别,并将识别结果进行比较。模糊积分融合方法是根据不同特征的识别率来确定分类器的隶属度,根据隶属矩阵来确定分类器的重要性,能够更好的使分类器之间互补,从而能将步态特征在决策层进行有效的融合。
最后,利用VC调用MATLAB引擎进行步态运动目标的提取,使用中科院步态数据库20个人3个视角的步态视频进行仿真实验,证明本文所采用的方法具有优越性。
Gait recognition is a technology which detected the gait through a series of video gait images sequences. Since it can detect the object from a long distance and could not be offensive, so it is widely used in the area of video surveillance.
In this thesis, the following work about gait detection、feature extraction and gait recognition is done:
In the area of gait recognition, the adaptive Gaussian mixture model is used to construct the background model. Background reduction division is also used to extract the gait contour. The experiment results show the method in this thesis is much better compared to the frame difference method.
In the process of shadow detection the existence of the shadow is judged based on the visual characteristics of shadows. If the shadow exists, gradient analysis and binary fast clustering method are used to detect it. At last the two methods are compared with each other and the result can prove the binary fast clustering method is better.
In gait characteristics extraction, dynamic characteristic gait and static characteristics were extracted based on the Gait contour. The gait width characteristics and the gait outline histogram, Fourier descriptor features and characteristics of wavelet using the Fourier transform and basic theory of wavelet transform in a gait circle are extracted.And PCA together with LDA is used to reduce the feature's dimension. LLE and LPP are also introduced. The work can reduce the complex of the following recognition.
In gait feature recognition, fuzzy integral is used in classifiers fusion, the classifier's importance is determined according to the characteristics of the subordinate matrix after the membership is determined. Thus make classifier is complementary, can be effectively in decision-making compared to the HMM model method.
Lastly, use VC calls MATLAB engine to get the gait moving targets, use Gait database of 20 individual with three perspectives of video gait sequence to construct the simulation experiment, the experiment result in this thesis proves the method has advantages. Key words:Fuzzy Integral; Moving Object Detection; Moving Shadow Detection; Background Modeling; Gait Recognition
引文
[1]黄红梅,李广林,王海潼.远距离人体步态识别算法[J].计算机工程,2007,33(21)：181-182,185.[2]薛召军,明东.步态识别研究现状与进展[J].生物医学工程学,2008,5(25)：1218-1222[3]赵喜玲,李其申,卢致天,李俊峰,一种改进的步态识别方法[J].计算机工程,2009,2(35)：177-179[4]王丽.基于帧间变化向量的步态识别[M].山东大学1-92[5] Kan Chao, Srinath MD.Combined features of cubic B-spline wavelet moments and Zernike moments for invariant character recognition[J].Information Technology,2001,4(2):511-51[6]王宏群.视频监控系统中的运动目标检测算法研究[D].合肥工业大学,2008：15-17[7]汪慧兰,陈思宝,罗斌.基于t混合模型和Greedy EM算法的彩色图像分割[J].中国图象图形学报,2007，12(5)：882-887[8]罗斌,涂铮铮,郭玉堂.基于运动信息和相位一致性的运动目标提取算法[J].系统仿真学报,2008,20(1)：94-98[9]李敏.视频监控系统中运动目标检测方法研究[D].西安：西北大学,2008：7-8[10]许文芳.基于人体宽度特征的步态识别[D].华侨大学硕士学位论文.2007[11]AiHua Wang, JiWei Liu.A Gait Recognition Method Based On Positioning Human Body Joints[C] Proceedings of the 2007 International Conference on Wavelet Analysis and Pattern Recognition, Beijing, China,2-4 Nov.2007 [12]Daoliang Tan, Kaiqi Huang, Shiqi Yu, Tieniu Tan.Efficient Night GaitRecognition Based on Template Matching[C]. The 18th International Conference on Pattern Recognition (ICPR'06)[13]李春明,李玉山，张大朴,刘洋.运动人物的检测、跟踪与识别综述.http://www.paper.edu.cn,
[14]M.A.T. Figueiredo, A.K.Jain. Unsupervised learning of finite mixture models[J].IEEE Trans Pattern Analysis and Machine Intelligence.2002,Vol.24, No.3:381-396.
[15]WangLiang,HuWei Ming.Tan Tie Niu.Gait-based human identification[J].Chinese Journal of Computer,2003,26(3):353-360
[16]叶波,文玉梅.基于轮廓宽度特征的步态识别[J].计算机应用,2005,25(8)：1792-1794
[17]薛召军,张帆,明东,万柏坤.基于人体轮廓面积特征的步态识别[J].计算机工程,2008,1(34)：195-197
[18]Yang Wang, Member, IEEE, Kia-Fock Loe, Jian-Kang Wu.A Dynamic Conditional Random Field Modelfor Foreground and Shadow Segmentation[C]IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL.28, NO.2, FEBRUARY 2006
[19]刘宏,李锦涛,刘群,钱跃良,李豪杰.融合颜色和梯度特征的运动阴影消除方法[J].计算机辅助设计与图形学学报,2007,10(19)：1281-1282
[20]高立志,方勇,林志航.图像的小波矩[J].西安交通大学学报,1999,5(33)：105-109
[21]苏菡,黄凤岗.基于周期序列宽度图的步态识别方法[J].中国图像图形学报,2007，8(12)：1383-1387
[22]M urat Ekinci,M urat AykutHuman. Gait Recognition Based on Kernel PCA Using Projections[J]. JOURNAL OF COMPUTERSCIENCE AND TECHNOLOGY 22(6):867-876 Nov.2007
[23]吴彦文,艾学轶.运用模糊积分进行信息融合的人脸识别方法[J].小型微型计算机系统,2008(6)：1154-1157
[24]李娟.模糊积分的近似计算[J].数学教学研究,2009，28(6)
[25]Meng X.L., Rubin D.B.,. Recent extensions to the EM algorithm,Bayesian Statistics 4[MS], Oxford University Press,1992:307-320.
[26]王勇,谭毅华。田金文.基于阴影消除和混合高斯模型的视频分
割算法[J]
[27]Shuichi Enokidat Ryo Shimomotot Tomohito Wadat Toshiaki Ejimat.A PREDICTIVE MODEL FOR GAIT RECOGNITION[J], National Institute of Fitness and Sports in Kanoya,2006 Biometrics Symposium
[28]E.Hayman,J.Eklundh.Statistical Background Subtraction for a Mobile Observer[J].In Proceedings ICCV,2003
[29]张虹,陈文楷.一种基于图像小波矩识别的方法[J]北方工业大学学报,2004.3(14)：428-430
[30]陈实,高有行.一种轮廓变化图像小波矩的步态识别[J].西安交通大学学报,2009,43(1)：90-93
[31]付耀文,黎湘,庄钊文.一种自适应模糊密度赋值的决策层融合目标识别算法[J].电子学报,2004(9)：1433-1435
[32]汪慧兰,陈思宝,罗斌.基于t混合模型和Greedy EM算法的彩色图像分割[J].中国图象图形学报,2007，12(5)：882-887
[33]颜根廷,李传江,马广富.支持向量分类器的模糊积分集成方法[J]哈尔滨工程大学学报,2008,40(7)：1016-1018
[34]曾莹,杨路明,曾庆冬.基于特征融合的步态识别算法[J].计算机工程与应用,2008,44(3)：220-222
[35]杨高波,亓波.精通MATLAB7.0混合编程[M].北京：电子工业出版社,2006：68-87
[36]胡学刚,吴勇.基于轮廓结构元素和阈值分割的形态学去噪[J].Computer Engineering V.35 No.23,2009
[37]章毓晋.图象分割[M].北京：科学出版社,2001：17-19
[38]王科俊,贲晛烨,赵玥.步态识别中的步态检测与序列预处理[J].模式识别与仿真,2009，28(8)：69-79
[39]Xu Han, JiWei Liu, Lei Li, Zhiliang Wang.Gait Recognition Considering Directions of Walking[J].2006 IEEE
[40]张前进,齐美彬,蒋建国,王倩,李莉.基于人体静态和动态特征融合的步态识别[J].系统仿真学报,2009，21(5)：1320-1323
[41]纪筱鹏,魏志强,冯业伟，王鹏.基于多特征融：合的运动物体识
别方法研究[J].哈尔滨工业大学学报,2009，3(41)：170-235
[42]Ye Bo,Wen Yu Mei.recognition based on the width of tht body contour[J].Computer Application,2005,25(8):1792-1794
[43]韩鸿哲,李彬,王志良,冀伟.基于傅立叶描述子的步态识别[J],2005,31(2)：48-16
[44]范晓峰,施泽生.基于小波矩的新型图形识别算法[J],计算机工程与应用,2001,7
[45]王琳,李勇平,王成波,张鸿洲.基于二维Gabor小波矩阵表征人脸的识别算法[J].计算机工程,2008,2(34)：8-10
[46]马瑞,宋亦旭.基于局部线性嵌入(LLE)非线性降维的多流形学习[J].清华大学学报(自然科学版),2008，48(4)
[47]谢永华,陈伏兵,张生亮,杨静宇.融合WMD矩阵与2DPCA的人脸特征抽取与识别[J].计算机工程,2007，16(33)：23-25
[48]路远.基于模糊支持向量机的步态识别[J].计算机工程,2009,21(35)：189-191
[49]阳爱民,胡运发,周咏梅.一种模糊分类器集成的方法[J].计算机工程,2006,17(32)：44-47
[50]Samaria F, Young S. HMM-based architecture for face identification.Image and Vision Computing,1994,12(8).
[51]黄凤岗,韩雪花.基于Radon变换的特征提取在步态识别中的应用[J].哈尔滨工程大学学报,2007，28(3)
[52]陈俊芬,何强,范铁刚.多分类器融合中模糊测度的确定[J].2007,21(3)：107-110
[53]付耀文,黎湘,庄钊文.一种自适应模糊密度赋值的决策层融合目标识别算法[J].电子学报,2004(9)：1433-1435
[14]M.A.T. Figueiredo, A.K.Jain. Unsupervised learning of finite mixture models[J].IEEE Trans Pattern Analysis and Machine Intelligence.2002,Vol.24, No.3:381-396.
[15]WangLiang,HuWei Ming.Tan Tie Niu.Gait-based human identification[J].Chinese Journal of Computer,2003,26(3):353-360
[16]叶波,文玉梅.基于轮廓宽度特征的步态识别[J].计算机应用,2005,25(8)：1792-1794
[17]薛召军,张帆,明东,万柏坤.基于人体轮廓面积特征的步态识别[J].计算机工程,2008,1(34)：195-197
[18]Yang Wang, Member, IEEE, Kia-Fock Loe, Jian-Kang Wu.A Dynamic Conditional Random Field Modelfor Foreground and Shadow Segmentation[C]IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL.28, NO.2, FEBRUARY 2006
[19]刘宏,李锦涛,刘群,钱跃良,李豪杰.融合颜色和梯度特征的运动阴影消除方法[J].计算机辅助设计与图形学学报,2007,10(19)：1281-1282
[20]高立志,方勇,林志航.图像的小波矩[J].西安交通大学学报,1999,5(33)：105-109
[21]苏菡,黄凤岗.基于周期序列宽度图的步态识别方法[J].中国图像图形学报,2007，8(12)：1383-1387
[22]M urat Ekinci,M urat AykutHuman. Gait Recognition Based on Kernel PCA Using Projections[J]. JOURNAL OF COMPUTERSCIENCE AND TECHNOLOGY 22(6):867-876 Nov.2007
[23]吴彦文,艾学轶.运用模糊积分进行信息融合的人脸识别方法[J].小型微型计算机系统,2008(6)：1154-1157
[24]李娟.模糊积分的近似计算[J].数学教学研究,2009，28(6)
[25]Meng X.L., Rubin D.B.,. Recent extensions to the EM algorithm,Bayesian Statistics 4[MS], Oxford University Press,1992:307-320.
[26]王勇,谭毅华。田金文.基于阴影消除和混合高斯模型的视频分
割算法[J]
[27]Shuichi Enokidat Ryo Shimomotot Tomohito Wadat Toshiaki Ejimat.A PREDICTIVE MODEL FOR GAIT RECOGNITION[J], National Institute of Fitness and Sports in Kanoya,2006 Biometrics Symposium
[28]E.Hayman,J.Eklundh.Statistical Background Subtraction for a Mobile Observer[J].In Proceedings ICCV,2003
[29]张虹,陈文楷.一种基于图像小波矩识别的方法[J]北方工业大学学报,2004.3(14)：428-430
[30]陈实,高有行.一种轮廓变化图像小波矩的步态识别[J].西安交通大学学报,2009,43(1)：90-93
[31]付耀文,黎湘,庄钊文.一种自适应模糊密度赋值的决策层融合目标识别算法[J].电子学报,2004(9)：1433-1435
[32]汪慧兰,陈思宝,罗斌.基于t混合模型和Greedy EM算法的彩色图像分割[J].中国图象图形学报,2007，12(5)：882-887
[33]颜根廷,李传江,马广富.支持向量分类器的模糊积分集成方法[J]哈尔滨工程大学学报,2008,40(7)：1016-1018
[34]曾莹,杨路明,曾庆冬.基于特征融合的步态识别算法[J].计算机工程与应用,2008,44(3)：220-222
[35]杨高波,亓波.精通MATLAB7.0混合编程[M].北京：电子工业出版社,2006：68-87
[36]胡学刚,吴勇.基于轮廓结构元素和阈值分割的形态学去噪[J].Computer Engineering V.35 No.23,2009
[37]章毓晋.图象分割[M].北京：科学出版社,2001：17-19
[38]王科俊,贲晛烨,赵玥.步态识别中的步态检测与序列预处理[J].模式识别与仿真,2009，28(8)：69-79
[39]Xu Han, JiWei Liu, Lei Li, Zhiliang Wang.Gait Recognition Considering Directions of Walking[J].2006 IEEE
[40]张前进,齐美彬,蒋建国,王倩,李莉.基于人体静态和动态特征融合的步态识别[J].系统仿真学报,2009，21(5)：1320-1323
[41]纪筱鹏,魏志强,冯业伟，王鹏.基于多特征融：合的运动物体识
别方法研究[J].哈尔滨工业大学学报,2009，3(41)：170-235
[42]Ye Bo,Wen Yu Mei.recognition based on the width of tht body contour[J].Computer Application,2005,25(8):1792-1794
[43]韩鸿哲,李彬,王志良,冀伟.基于傅立叶描述子的步态识别[J],2005,31(2)：48-16
[44]范晓峰,施泽生.基于小波矩的新型图形识别算法[J],计算机工程与应用,2001,7
[45]王琳,李勇平,王成波,张鸿洲.基于二维Gabor小波矩阵表征人脸的识别算法[J].计算机工程,2008,2(34)：8-10
[46]马瑞,宋亦旭.基于局部线性嵌入(LLE)非线性降维的多流形学习[J].清华大学学报(自然科学版),2008，48(4)
[47]谢永华,陈伏兵,张生亮,杨静宇.融合WMD矩阵与2DPCA的人脸特征抽取与识别[J].计算机工程,2007，16(33)：23-25
[48]路远.基于模糊支持向量机的步态识别[J].计算机工程,2009,21(35)：189-191
[49]阳爱民,胡运发,周咏梅.一种模糊分类器集成的方法[J].计算机工程,2006,17(32)：44-47
[50]Samaria F, Young S. HMM-based architecture for face identification.Image and Vision Computing,1994,12(8).
[51]黄凤岗,韩雪花.基于Radon变换的特征提取在步态识别中的应用[J].哈尔滨工程大学学报,2007，28(3)
[52]陈俊芬,何强,范铁刚.多分类器融合中模糊测度的确定[J].2007,21(3)：107-110
[53]付耀文,黎湘,庄钊文.一种自适应模糊密度赋值的决策层融合目标识别算法[J].电子学报,2004(9)：1433-1435