CN102663374B - Multi-class Bagging gait recognition method based on multi-characteristic attribute - Google Patents

Multi-class Bagging gait recognition method based on multi-characteristic attribute Download PDF

Info

Publication number
CN102663374B
CN102663374B CN 201210134185 CN201210134185A CN102663374B CN 102663374 B CN102663374 B CN 102663374B CN 201210134185 CN201210134185 CN 201210134185 CN 201210134185 A CN201210134185 A CN 201210134185A CN 102663374 B CN102663374 B CN 102663374B
Authority
CN
China
Prior art keywords
gait
multi
image
attribute
algorithm
Prior art date
Application number
CN 201210134185
Other languages
Chinese (zh)
Other versions
CN102663374A (en
Inventor
杨新武
翟飞
杨跃伟
Original Assignee
北京工业大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京工业大学 filed Critical 北京工业大学
Priority to CN 201210134185 priority Critical patent/CN102663374B/en
Publication of CN102663374A publication Critical patent/CN102663374A/en
Application granted granted Critical
Publication of CN102663374B publication Critical patent/CN102663374B/en

Links

Abstract

基于多步态特征属性的多类别Bagging步态识别方法,属于模式识别技术领域。 Bagging gait multi-class recognition method based on multi-attribute Gait, belonging to the field of pattern recognition technology. 该方法用最近邻分类器作为弱分类器,在基于小波包分解和完全主成分分析后的20个步态属性特征集合上,通过将两类属性Bagging方法扩展到多类别来构建集成分类器,进行步态身份鉴别。 The method for nearest neighbor classifier as a weak classifier in the feature attribute after 20 gait analysis and wavelet packet decomposition based on the main component is completely set by the extended attribute Bagging two methods to construct the multi-class classifier integrated, gait authentication. 该方法的步骤包括:预处理、特征提取,最后用最近邻分类原理与MCAB算法相结合的方法对测试样本进行归类。 The method comprises the step of: preprocessing, feature extraction, and finally classified by the test samples nearest neighbor method and the principle of combining MCAB algorithm. 本发明首次采用融合WPD和(2D)2PCA方法来提取并选择步态特征,解决已有基于小波变换的步态识别方法中高频分量丢失或简单采用全部数据所致维数过大问题,具有更高的识别率和视角变化鲁棒性。 Fusion of the present invention for the first time and WPD (2D) 2PCA method to extract features of the gait and select, based on solving major problems have been lost, or simply a high frequency component data using all wavelet transform gait recognition method due to over dimension, a more a high recognition rate and robustness of viewing angle.

Description

基于多特征属性的多类别Bagging步态识别方法 Bagging gait multi-class recognition method based on feature properties

技术领域 FIELD

[0001] 本发明属于模式识别技术领域,具体涉及一种基于多步态特征属性的多类别Bagging步态识别新方法,是一种利用计算机技术、数字图象处理技术、模式识别等实现人的步态的自动分析与判别的方法,是生物特征识别领域中关于步态特征提取与识别的算法。 [0001] The present invention belongs to the technical field of pattern recognition, in particular, to multi-class based on multi Bagging attribute Gait gait recognition method, a use of computer technology, digital image processing and pattern recognition to achieve al. analysis and automatic determination of the gait, a biometric art on the gait feature extraction and recognition algorithm.

背景技术 Background technique

[0002] 生物特征识别技术是指利用人类本身所拥有的、能够标识其身份的生理特征或行为特征进行身份验证的技术。 [0002] Biometrics is the use of humanity itself have, capable of identifying the identity of the physiological characteristics or behavioral characteristics authentication technology. 与传统的身份验证技术相比,生物特征识别技术从根本上杜绝了伪造和窃取,具有更高的可靠性、安全性,已经越来越广泛的应用于一些安全系统的身份认证。 Compared with traditional authentication technology, biometric identification technology to eliminate the fundamental forgery and theft, greater reliability, security has become increasingly widely used in some security authentication system.

[0003] 步态识别技术作为一种新型的生物特征识别技术,它是根据视频序列中人走路的姿势进行身份识别的生物识别技术。 [0003] gait recognition technology as a new type of biometric technology, which is biometrics identification according to a video sequence of human gait. 和其它生物特征识别技术相比,步态识别技术以其非侵犯性、远距离识别性以及难以隐藏等优点受到了人民的青睐,在国家公共安全、金融安全、身份认证、视频监控等领域有着广泛的应用前景。 And compared to other biometric identification technology, gait recognition technology for its non-invasive, long-distance and the difficulty of identifying hidden advantages favored by the people, with the national public security, financial security, authentication, video surveillance and other fields wide range of applications.

[0004] 关于步态的特征提取技术,有文献采用小波包分解较好地解决了这个问题,但小波包分解后的图像特征维数较高,且其采用经典的PCA算法对进行特征提取,即采用奇异值分解的方法来求相关矩阵的特征值和特征向量时,计算耗费大。 [0004] For the gait feature extraction, wavelet packet decomposition literature solve this problem, but the higher the image feature dimension wavelet packet decomposition, and the classical PCA algorithm for feature extraction, i.e., the method of singular value decomposition to find the eigenvalues ​​and eigenvectors of the correlation matrix, a large cost is calculated. 二维主成分分析(2DPCA)可直接对图像数据矩阵进行计算,计算量相对少很多,但2DPCA后需要n*k(其中,η为图像分辨率,k为变换后选取特征列向量个数,且k < η)个数据来表示图像,特征向量的维数仍较高。 Dimensional principal component analysis (2DPCA) the image data matrix can be calculated directly, relatively less amount of calculation, but need 2DPCA n * k (where, [eta] is the image resolution, k is the number of columns of the transformed feature vectors selected, and k <η) of data to represent the image dimension feature vector is still high. 完全主成分分析((2D)2PCA)可以进一步降低特征向量的维数,从而降低识别耗费,且其识别性能上与2DPCA相当,甚至优于2DPCA。 Principal Component Analysis completely ((2D) 2PCA) can be further reduced dimension feature vector, thereby reducing the cost to identify, and on which identification 2DPCA performance equivalent or superior to 2DPCA.

[0005]目前已经有人提出一种关于两类问题的属性Bagging算法AB,本发明是在两类问题的属性Bagging算法AB的基础上,提出一种基于多步态特征属性的多类别Bagging步态识别新方法-MCAB(Mult1-class Attribute Bagging)算法。 [0005] has now been suggested that an attribute Bagging algorithm AB about two types of problems, the present invention is based on the properties of the two types of AB Bagging algorithm problem, we propose a gait characteristics based on multi-attribute multi-class Bagging gait recognition method -MCAB (Mult1-class Attribute Bagging) algorithm.

发明内容 SUMMARY

[0006] 本发明的内容是提出了一个基于多步态特征属性的多类别Bagging(MCAB)步态识别新方法。 [0006] The present invention is a new method based on multi-multi-class attribute Gait Bagging (MCAB) proposed gait recognition. 该方法用INN作为弱分类器,通过将两类属性Bagging方法扩展到多类别来构建集成分类器MCAB(Mult1-class Attribute Bagging)。 The method as INN weak classifiers to build an integrated classifier MCAB (Mult1-class Attribute Bagging) extended by two methods Bagging attributes to multiple categories. 我们在NLPR步态数据库上对该方法进行了评测,结果表明,与单纯采用小波包和(2D)2PCA等识别方法相比,本方法具有更高的识别率和视角变化鲁棒性。 We NLPR gait database on the method of evaluation, the results show that, and (2D) 2PCA the like as compared to only using the method of identification packet, the method has a higher recognition rate and robustness of viewing angle.

[0007] 本发明的技术内容如下: [0007] The teachings of the present invention is as follows:

[0008] 为利用MCAB算法进行实验,我们需首先通过对预处理后的归一化步态图像序列进行周期检测,来提取步态能量图,以克服步态数据量过大问题;再对步态能量图进行小波包分解和完全主成分分析,得到的结果图像分别表示了步态图像不同侧面的特征;最后依据前述不同特征的分类性能,通过将各特征视为步态的不同属性并用MCAB算法进行分类。 [0008] The experiment was conducted using MCAB algorithm, we need to hold by normalizing the image sequence gait period detection preprocessed to extract energy FIG gait, the gait data is too large to overcome the problem; again step FIG state energy full wavelet packet decomposition and principal component analysis, the results obtained represent the image of different characteristics of the image side surface of gait; Finally, according to the classification performance of the different features, each feature considered by gait and with different attributes MCAB classification algorithm.

[0009] —个基于多步态特征属性的多类别Bagging(MCAB)步态识别新方法,该方法的步骤包括:人体步态序列的预处理、特征提取,最后利用MCAB算法把测试样品归到相应的类中,并对识别效果给予评价,其具体步骤如下: [0009] - more than two categories on Bagging (MCAB) based on multi-attribute Gait gait recognition method, the method comprising the steps of: pre-sequence human gait, feature extraction, and finally the test sample using the algorithm MCAB attributed to the corresponding class identification and evaluation of the effect of administration, the specific steps are as follows:

[0010] 步骤一、预处理 [0010] Step a pretreatment

[0011] (I)形态学处理 [0011] (I) morphological

[0012] 对已经背景分离的人体运动目标图像进行形态学处理,以去除二值化图像存在的空洞,获得更优的分割效果; [0012] Background of the isolated human motion has certain morphological image processing to remove voids present in the binarized image, to obtain a better segmentation results;

[0013] (2)目标提取 [0013] (2) object extraction

[0014] 利用8连通分量分析的方法来提取一个单连通的运动目标,即人的侧影去除残余噪声,从而获得更优的二值轮廓图; [0014] to extract a moving object communication utilizing a single connected component analysis method of 8, i.e., the human silhouette remove residual noise, so as to obtain better binary profile;

[0015] (3)图像归一化 [0015] (3) image normalization

[0016] 根据人体轮廓坐标裁剪出标准的步态图像,得到尺寸归一化图像,其中,图像的大小统一为64*64像素。 [0016] The cut-out according to the standard outline coordinate human gait image size to obtain a normalized image, wherein the image of uniform size is 64 * 64 pixels.

[0017] 步骤二、特征提取 [0017] Step two, the feature extraction

[0018] (I)步态周期的检测 [0018] Detection (I) of the gait cycle

[0019] 利用人体的轮廓宽度随时间发生同步周期性改变的特性,通过人体轮廓的宽度变化信号来划分步态周期; [0019] synchronization using periodically changing profile width characteristics of the human body occurs over time, changes in the width of the signal contour of the body to divide the gait cycle;

[0020] (2)建立步态能量图(GEI),将步态能量图作为不同步态序列的代表; [0020] (2) establishing gait energy image (GEI), as the gait energy image sequences representing different gait;

[0021] 步态能量图:在进行步态周期检测之后,通过对一个周期内的步态序列图像处理生成GEI为:1 N1 [0021] FIG Gait Energy: After the detection of the gait cycle is performed, an image processing sequence of the gait cycle of a generation of GEI: 1 N1

[0022] [0022]

Figure CN102663374BD00041

[0023] 式中,G(x, y)是这个周期序列的步态能量图,N1是完整步态周期序列的长度,Bt (X,y)是一个周期中的第t个步态图像,X, y代表二维图像平面坐标。 [0023] In the formula, G (x, y) is the periodic sequence gait energy image-on, N1 is the length of the complete sequence of the gait cycle, Bt (X, y) is the image of a t-th gait cycle, X, y coordinates of the representative two-dimensional image plane.

[0024] (3)融合WPD+ (2D) 2PCA 特征选择 [0024] (3) fusion WPD + (2D) 2PCA feature selection

[0025] 首先,采用二级小波包分解(WPD)步态能量图,经分解后,每幅步态能量图得到20个分解后的图像;然后,分别采用完全主成分分析((2D)2PCA)(即行、列方向的二维主成分分析)进一步提取有效特征;最后用最近邻分类器分别进行识别,分别对每个特征进行识别率的计算。 [0025] First, two wavelet packet decomposition (the WPD) gait energy image, after the decomposition, each web gait energy image obtained after decomposition of the image 20; then, were analyzed ((2D) 2PCA main component a fully ) (i.e. row and column direction of the two-dimensional principal component analysis) is further effective feature extraction; nearest neighbor classifier finally identified separately for each feature are calculated by the recognition rate.

[0026] 步骤三、集成分类识别 [0026] Step three, integrated classification

[0027] (I)确定重抽样的原始属性集合 [0027] (I) determined resampling the original set of attributes

[0028] 在MCAB算法中,首先需要对原始训练实例的属性集合进行η次有放回的重抽样,η为经过WPD+(2D) 2PCA后识别率大于等于50%的属性个数,由这η个属性构成新的训练实例,再由这些新的训练实例构造弱分类器。 [0028] In MCAB algorithm, first needs to properties of the original collection of training examples of [eta] re-sampling with replacement times, [eta] is after WPD + (2D) after 2PCA recognition rate more than 50% is equal to the number of attributes by which [eta] attributes constitute a new training examples, then the weak classifiers configured these new training examples. 由于对原始属性集进行有放回的重抽样,因此在新训练例描述中有的属性可能出现多次,有的属性可能一次也不出现。 Since the original attribute set of re-sampling with replacement, so some properties may appear several times in the new training described, some properties may not appear once. 我们选择那些识别率大于等于50%的属性构成属性集合AttributeSet。 We chose those recognition rate of 50% or more attributes constitute a set of attributes AttributeSet.

[0029] (2)利用MCAB算法进行分类识别 [0029] (2) using a classification algorithm MCAB

[0030] 有文献提出了一种关于两类问题的属性Bagging算法AB,设有两个类别,类别空间为Y = {1,-1},实例空间为(2,训练实例集为s = {(Xl,yi), (x2,y2),...,(xm,ym)},其中Xi e <2 ,Yi e Y,实例Xi用η个属性表示,属性集A为{a1; a2..., an}。算法AB共进行T轮,每一轮中从原始属性集中有放回的重抽样η个属性,由这η个新属性描述的训练例构成新的训练例集,再将新训练实例用基分类算法训练出基分类器ht (x) — {-1,1},分配给ht (x)的权值为at ;最后对未知的实例用T个基分类器的结果进行投票表决,简单投票函数为 [0030] has been proposed an attribute AB Bagging algorithm on two issues, there are two categories, category space Y = {1, -1}, of instance space (2, examples of the training set as s = { (Xl, yi), (x2, y2), ..., (xm, ym)}, where Xi e <2, Yi e Y, Xi is represented by η instance attributes, attribute set A is {a1; a2. .., an}. algorithm AB total of T rounds, each round from the original attribute set of re-sampling with replacement attribute η, η training cases by these new attributes described embodiments constitute a new set of training, and then the new group classification algorithm with training examples to train the classifier based ht (x) - {-1,1}, the weight assigned to ht (x) is the AT; final instance with unknown results of the T groups classifiers vote, simple poll function

[0031] H(x) = sign(f (x)) (I) [0031] H (x) = sign (f (x)) (I)

Figure CN102663374BD00051

[0033] 但步态识别问题是一个多类别的分类问题,为此,我们引入Ml算法将上述算法扩展为多步态多类别属性Bagging算法MCAB。 [0033] However, gait recognition problem is a multi-class classification problem, for which we introduce Ml algorithm will extend the above algorithm is a multi-multi-class property gait Bagging algorithm MCAB. 设多类别空间为Y= {1,2,...,m},yi e Y。 Provided multi-class space Y = {1,2, ..., m}, yi e Y. 对于多类别Ml算法,设每轮训练使用的训练集为St,其中t= 1,…,T,其中T为迭代次数,在St上用弱学习算法训练生成弱分类器——预测函数ht(x) — Y,分配给ht(x)的权值为at,最后的强预测函数为 For multi-class algorithm Ml, provided training set used is trained round St, where t = 1, ..., T, where T is the iteration number, an algorithm for generating a weak classifier trained on the weak learning St - prediction function HT ( x) - Y, assigned ht (x) is the weight at, the final strength prediction function

[0034] [0034]

Figure CN102663374BD00052

[0035] 其中,当基分类器分到的类别标号与y相同时,Mht(X) = y|为1,否则为O; [0035] wherein, when the group classification assigned class label y with the same, Mht (X) = y | is 1, or is O;

Figure CN102663374BD00053

为ht(x)分类正确的个数减去分分类错误的个数,t为迭代次数,t= 1,...,T。 Is ht (x) the number of correctly classified by subtracting the number of sub-classification errors, t is the number of iterations, t = 1, ..., T.

[0036] 上式可以理解为测试样本X在每个基分类器上会分到一个类别ye Y,设y有权值W,若基分类器ht分到了y上,则对应的权值w为该基分类器ht的权值at,最终权值之和最大的那类为测试样本分到的类。 [0036] formula X may be understood in each of the test sample will be assigned to a group classification category ye Y, provided the right y value W, when the group assigned to the classifier ht y, the corresponding weight values ​​w is the right of base classifier ht value at, final value and the largest kind of test samples assigned to the class.

[0037] MCAB执行T轮,一般取10 = < T < = 100,每一轮从前述的原始属性集合AttributeSet中有放回的重采样η个属性,由这η个新属性描述的训练例构成新的训练例集;在新训练例集上用INN算法训练出不同的基分类器ht(x);最后由上述多个基分类器按式(2)构成集成分类器。 [0037] MCAB performed wheel T, and generally 10 = <T <= 100, each in an AttributeSet η resampling with replacement of attributes from the original set of properties, which the training cases η new configuration attributes described Example new training set; training on new training set cases with different algorithms INN base classifier ht (x); and finally by the plurality of base classifiers by formula (2) constituting an integrated classifier.

[0038] 本发明的有益效果在于:1.通过对步态周期计算步态能量图,这幅步态能量图包含了步态周期中各步态图像的轮廓、频率、相位等步态信息,可保证不抛弃步态特征的同时减少待处理的步态图像数据量,以降低计算消耗;2.利用基于WPD+(2D)2PCA方法来提取特征,解决已有基于小波变换的步态识别方法中高频分量丢失或简单采用全部数据所致维数过大问题,此方法能更为准确地提取反应运动人体行走特征的有效信息和减少用于识别步态的特征维数;3.利用MCAB算法进行分类识别以提高步态识别正确率,且对视角变化具有很好的鲁棒性。 [0038] Advantageous effects of the present invention: FIG. 1 by the energy calculation gait gait cycle, this piece comprising a gait gait energy image information of each gait cycle gait contour image, frequency, phase, etc., gait reducing the amount of image data to be processed is not guaranteed, while disposable gait characteristics, to reduce computational overhead;. 2 extracted using WPD + (2D) 2PCA based on feature to solve the existing gait recognition method based wavelet transform high frequency component lost due to all of the data or simply using dimensionality problem is too large, this method can more accurately extract valid characteristics information human walking motion of the reaction and reduce the feature dimension for identifying gait; 3 using the algorithm MCAB classification in order to improve gait recognition accuracy, and has good robustness to changes in perspective.

附图说明 BRIEF DESCRIPTION

[0039] 图1是本发明算法的流程图。 [0039] FIG. 1 is a flowchart showing an algorithm of the present invention.

[0040] 图2是本发明算法中预处理的流程图 [0040] FIG 2 is a flowchart showing an algorithm of the present invention is pretreated

[0041] 图3是本发明算法特征提取的流程图 [0041] FIG. 3 is a flowchart showing an algorithm of the present invention, the extracted feature

[0042] 图4是本发明算法中二级WPD+ (2D) 2PCA特征提取流程图 [0042] FIG. 4 is an algorithm of the present invention, two WPD + (2D) 2PCA flowchart feature extraction

[0043] 图5是本发明算法中小波包分解图具体实施方式 [0043] FIG. 5 is an algorithm of the present invention, wavelet packet decomposition DETAILED DESCRIPTION FIG.

[0044] 下面详细给出该发明技术方案中所涉及的各个细节问题的说明: [0044] The following detailed description will be given of details of each aspect of the invention relates to:

[0045] 步骤一、预处理过程如下: [0045] Step a pretreatment process is as follows:

[0046] 我们采用的数据库是中国科学院自动化研究所的一个步态数据库,该数据库已经将背景分离,本发明要做的工作是在这基础上进行预处理,从而进行周期检测及计算步态能量图等操作。 [0046] We use the database is a database gait Chinese Academy of Sciences, Institute of Automation, the database has been separated from the background, the present invention is to pre-treatment work to be done on this basis, thereby performing periodic gait detection and calculation of energy FIG other operations.

[0047] (I)形态学处理 [0047] (I) morphological

[0048] 由于天气、光照、影子等其它外界因素的影响,背景分离后的图像中难免会存在噪声,因此还需要对图像做进一步处理,以获得最佳的分割效果。 [0048] Since the influence of other external factors, weather, illumination, shadow, etc., the images will inevitably separated background noise is present, it is also necessary to further image processing, in order to obtain optimum segmentation results. 本发明使用形态学滤波来消除二值图像中的噪声并填补运动目标的缺失。 The present invention, morphological filtering is used to eliminate the noise in the binary image and to fill in missing the moving object. 作为一种常用的图像滤噪方法,形态学用于图像滤波的最基本运算是膨胀与腐蚀,由膨胀与腐蚀的相互结合又派生出另外两种运算:开运算和闭运算。 As a conventional method of image noise filtering, morphological filtering for the basic operation is the image dilation and erosion, and corrosion by the expansion sent birth bonded to each other two operations: opening operation and closing operation. 开运算可平滑对象的凸轮廓,断开狭窄的连接,去掉细小的突起部分;闭运算可平滑对象的凹轮廓,将狭长的缺口连接成细小的弯口。 Opening operation can be smoothly convex contour of the object, the narrow connecting disconnect, remove fine projections; closing operation of the object can be smoothly concave profile, the narrow gap into small curved connecting port. 利用这个性质可以实现滤波和填充空洞的目的。 This property can be implemented using filtering and filling empty object.

[0049] ⑵目标提取 [0049] ⑵ object extraction

[0050] 经形态学处理后,仍可能存在部分杂散噪声形成大小不一的块,而真正的运动目标往往是这些块中最大的。 [0050] After morphological processing, still may be some spurious noise block sizes is formed, and the real moving object is often the largest of these blocks. 因此对图像进一步进行连通域分析,即利用8连通分量分析的方法来提取一个单连通的运动目标,目的在于仅保留图像中的运动目标,从而获得更优的二值轮廓图。 Thus the image is further connected component analysis, i.e., connected component analysis using the method of 8 to extract a single moving object communication, retaining only moving object in the target image to obtain a binary better profile.

[0051] (3)图像归一化 [0051] (3) image normalization

[0052] 为了消除图像大小对识别的影响,应首先使人体居中,然后将图像的大小统一为64*64像素。 [0052] In order to eliminate the influence of the size of the image recognition, so that the body should first middle, and then the image of uniform size is 64 * 64 pixels.

[0053] 步骤二、特征提取过程如下: [0053] Step two, the feature extraction process is as follows:

[0054] (I)步态周期的检测 [0054] Detection (I) of the gait cycle

[0055] 人的行走是一个周期性的行为,定义步态周期为:从足跟着地到同侧腿足跟再次着地所经历的时间,包括两个站立期和两个摆动期。 [0055] people walking is a cyclical behavior, gait cycle is defined as: from heel strike to the ipsilateral legs and feet touch the ground again with time experienced, including the stance phase two and two swing phase. 为了提高效率,本发明利用人体的轮廓宽度随时间发生同步周期性改变的特性,通过人体轮廓的宽度变化信号来划分步态周期。 To improve efficiency, the present invention utilizes the characteristics of the synchronization periodically changing profile width of the body occur over time, divided by the width of the gait cycle signal variation profile body.

[0056] (2)步态能量图 [0056] (2) gait energy image

[0057] 在进行步态的周期检测之后,直接采用步态周期进行步态识别存在数据量庞大,加剧了步态特征提取的难度和计算消耗,为保证不抛弃步态特征同时减少待处理的步态图像数据量,本发明采用了步态能量图的方法,即将一周期中的若干步态图像经过加权平均的方法合成为一幅图像,这幅图像包含了步态周期中各步态图像的轮廓、频率、相位等步态信息。 [0057] After detecting the gait cycle is performed, the direct use of the gait cycle for identifying the presence of large amounts of data gait, the gait feature extraction increased the difficulty and computational cost, in order to ensure not abandon gait to be treated while reducing characterized gait amount of image data, the present invention employs a gait energy map, a plurality of image gait upcoming week period weighted average method of a synthesized image, this image comprising an image of each gait cycle gait gait contour information, frequency, phase and the like. 这种方法不需要考虑每帧的步态间隔的大小,并且还避免一些偶然因素的影响。 This method does not consider the size of each frame interval gait, and also avoids some of the effects of accidental factors. 对于给定的二值步态周期图像序列Bt (X, y), GEI的定义如下:1 N1 For a given binary image sequences gait cycle Bt (X, y), GEI defined as follows: 1 N1

Figure CN102663374BD00061

[0059] 其中,G(x, y)是这个周期序列的步态能量图,N1是完整步态周期序列的长度,Bt (X,y)是一个周期中的第t个步态图像,X, y代表二维图像平面坐标。 [0059] wherein, G (x, y) is the periodic sequence gait energy image-on, N1 is the length of the complete sequence of the gait cycle, Bt (X, y) is the image of a t-th gait cycle, X y represents coordinates of a two-dimensional image plane. [0060] (3)融合WPD+ (2D) 2PCA 特征选择 [0060] (3) fusion WPD + (2D) 2PCA feature selection

[0061] 二维离散小波变换将每帧图像分为低频分量A和高频分量,其中包括水平分量H、垂直分量V和对角分量D,该方法仅保留了图像信号的低频分量A,损失了高频分量。 [0061] The two-dimensional discrete wavelet transform of each frame image into low frequency components and high frequency components A, which includes a horizontal component H, and the vertical component V D diagonal components, which retains only the low frequency components of the image signal A, the loss of the high-frequency component. 小波包分解(WPD)不仅保留了图像信号的低频分量,还保留了高频分量。 Wavelet packet decomposition (the WPD) not only retains the low frequency component image signal, but also retains the high-frequency component.

[0062] 小波包分解后的图像特征维数较高,采用经典的PCA算法对进行特征提取时,即采用奇异值分解的方法来求相关矩阵的特征值和特征向量时,计算量很大。 [0062] After the image feature dimension higher wavelet packet decomposition, classical PCA algorithm when the feature extraction method which uses singular value decomposition to find the eigenvalues ​​and eigenvectors of the correlation matrix, the large amount of calculation. 而二维主成分分析(2DPCA)直接对矩阵进行计算,计算量相对少很多。 And the two-dimensional principal component analysis (2DPCA) direct calculation of the matrix, relatively less calculation amount. 但2DPCA在图像特征提取方面也有其弊端。 However, the image feature extraction 2DPCA also has its drawbacks. 假设图像大小是64*64,经2DPCA后需要64*k(其中k为变换后选取特征列向量个数,且k < 64)个数据来表示图像,为了满足精度要求,通常k要大于5,这样至少需要320个数据来表示图像,特征向量的维数仍较高。 Assumed that the image size is 64 * 64, 64 * required after 2DPCA k (where the number of row select feature vector of the k transform, and k <64) to represent the image data, in order to meet the accuracy requirements, usually greater than 5 k, this requires at least 320 data represents the image dimension feature vector is still high. 完全主成分分析((2D)2PCA)可以进一步降低特征向量的维数,从而降低识别耗费,且其识别性能上与2DPCA相当,甚至优于2DPCA。 Principal Component Analysis completely ((2D) 2PCA) can be further reduced dimension feature vector, thereby reducing the cost to identify, and on which identification 2DPCA performance equivalent or superior to 2DPCA. 为更有效地利用图像的信息来进行步态识别,基于上述考虑,本文提出了一种融合小波包分解(WPD)和(2D) 2PCA的方法。 A more effective use of the image information to gait recognition, based on the above considerations, the proposed method of wavelet packet decomposition fusion (the WPD) and (2D) 2PCA of.

[0063] 具体实现过程如下: [0063] The specific implementation process is as follows:

[0064] 首先对每幅步态能量图G(x,y)进行二级WPD分解,得到20个分解后的图像。 [0064] First, each web gait energy image G (x, y) for two WPD decomposed image 20 obtained after decomposition. 即:利用小波包性质对原始步态能量图进行第一级wro分解,得低频Al、水平高频H1、垂直高频V1、对角高频Dl的图像,再分别对这四个图像进行第二级Wro分解,依次得到低频A2,水平高频H2,垂直高频V2,对角高频D2,低频HA2,水平高频HH2,垂直高频HV2,对角高频HD2,低频VA2,水平高频VH2,垂直高频W2,对角高频VD2,低频DA2,水平高频DH2,垂直高频DV2,对角高频DD2。 That is: the nature of the original wavelet packet gait energy image a first decomposition stage wro give low Al, horizontal high frequency H1, the vertical high V1, Dl-frequency image angle, and then each of these four images Wro two decomposed successively to obtain low A2, horizontal high frequency H2, the vertical high-frequency V2, the high frequency diagonal D2, low HA2, the HH2 horizontal frequency, vertical frequency HV2, the HD2 diagonal high-frequency, VA2 of low frequency, high levels of frequency VH2, vertical high W2, diagonal frequency VD2, low DA2, horizontal high frequency DH2, vertical high DV2, DD2 diagonal frequency.

[0065] 其次,分别对WPD分解后的图像(Al,-,DD2)进行(2D)2PCA变换,(2D)2PCA变换的具体过程如下: [0065] Next, each decomposed image after WPD (Al, -, DD2) for (2D) 2PCA transform specific process (2D) 2PCA transformed as follows:

[0066] i)计算总体样本均值矩阵 [0066] i) calculating the total sample mean matrix

Figure CN102663374BD00071

[0068] 其中,Ak为第k个样本图像,M为样本总数 [0068] wherein, Ak is k-th sample image, M being the total number of samples

[0069] ii)计算样本协方差矩阵 [0069] ii) calculate the sample covariance matrix

[0070] [0070]

Figure CN102663374BD00072

[0072] 分别计算协方差矩阵G1、G2的特征值和标准正交特征向量,前dl、d2个较大的非零特征值的特征向量组成的矩阵为u,v。 [0072] are calculated eigenvectors of the covariance matrix G1, eigenvalues ​​and orthonormal eigenvectors G2, the front dl, d2, larger non-zero eigenvalues ​​is composed of u, v. 其中,[ among them,[

Figure CN102663374BD00073

[0073] iii)生成(2D)2PCA变换后的特征矩阵Y' [0073] iii) generating (2D) characteristics after transformation matrix 2PCA Y '

[0074] Y' = UAV [0074] Y '= UAV

[0075] 其中,A为任意的样本矩阵。 [0075] wherein, A is an arbitrary matrix of samples.

[0076] 最后,分别对变换后的20个图像序列进行最近邻分类,分别对每个图像序列利用INN计算识别率。 [0076] Finally, each of the 20 sequences in the transformed image nearest neighbor classification, recognition rates were calculated using the INN for each image sequence. [0077] 步骤三、利用MCAB算法进行分类 [0077] Step three, the use of classification algorithms MCAB

[0078] (I)确定重抽样的原始属性集合 [0078] (I) determined resampling the original set of attributes

[0079] [0079]

Figure CN102663374BD00081

[0080] 以O。 [0080] In O. 为例,选择上步骤中最后的识别率识别率大于等于50%的特征构成待重抽样的原始特征集合Attribute Set = {Al,H1,VI,A2,H2,V2,HA2,HH2}。 For example, the step of selecting the final recognition rate in the recognition rate of 50% or more features configured to be re-sampling the original feature set Attribute Set = {Al, H1, VI, A2, H2, V2, HA2, HH2}. 设有20 个人的步态图像,即类别个数为20。 Gait with 20 individual images, i.e., the number of classes is 20.

[0081 ] (2)利用MCAB算法进行分类识别 [0081] (2) using a classification algorithm MCAB

[0082] Stepl:确定重抽样属性集合中的个数为8,确定迭代次数10 ; [0082] Stepl: determining the number of re-sampling the set of attributes to 8, 10 to determine the number of iterations;

[0083] Step2:For t = 1: 10,迭代10 次; [0083] Step2: For t = 1: 10, 10 iterations;

[0084] 第一次迭代:从Attribute Set中有放回地抽样8次,每次抽取一个,得到属性集Al = {Al,HI, HH2,V2, H2,VI,Hl, Al},由于每次都是随机地有放回的抽取,故有些属性可能会出现多次,例如:A1,H1,有些属性可能不会出现,例如:A2,HA2。 [0084] The first iteration: the sampling with replacement from the Attribute Set 8 times, each time an extract, to give attribute set Al = {Al, HI, HH2, V2, H2, VI, Hl, Al}, since each times are random with replacement of extraction, so some property may appear multiple times, for example: A1, H1, some properties may not appear, for example: A2, HA2. 每个训练样本实例Xk的属性只取A1 得到向量Slk = [Al(k) Hl(k) HH2(k) V2(k) H2(k) Vl(k) Hl(k)Al(k)],Xk 用Slk 来表示,用INN算法训练出基分类器Ill (X) — Y,计算权值: Each training sample Xk instance attribute vector obtained only take A1 Slk = [Al (k) Hl (k) HH2 (k) V2 (k) H2 (k) Vl (k) Hl (k) Al (k)], Slk represented by Xk, an algorithm trained by INN-yl classifier Ill (X) - Y, weights are calculated:

Figure CN102663374BD00082

[0086] 其中,!T1 = Ii1(X)正确分类的个数减去Ii1 (X)错误分类的个数。 [0086] wherein,! T1 = Ii1 (X) minus the number of correct classification number Ii1 (X) misclassification.

[0087] 第十次迭代:从Attribute Set中有放回地重抽样8次得到属性集Altl = {V2,HH1,A2,H2,HH2,HA2,H1,A1},每个实例样本Xk 的属性只取Altl 得到向量Sltlk= [V2(k) HHl(k) A2(k)H2(k) HH2(k) HA2(k)Hl(k) Al(k)] Ji 用Sltlk 来表示,用INN 算法训练出基分类器h1(l (x) —Y,计算权值 [0087] Tenth iterations: Attribute Set from there back to the resampling 8 times to obtain attribute set Altl = {V2, HH1, A2, H2, HH2, HA2, H1, A1}, each instance of the properties of the sample Xk just take Altl resulting vector Sltlk = [V2 (k) HHl (k) A2 (k) H2 (k) HH2 (k) HA2 (k) Hl (k) Al (k)] Ji by Sltlk represented by INN algorithm training a classifier based h1 (l (x) -Y, weights are calculated

Figure CN102663374BD00083

[0089] 其中,r10 = h10(x)正确分类的个数减去h1Q(x)错误分类的个数。 [0089] where, r10 = h10 (x) minus the number of correct classification number h1Q (x) misclassification.

[0090] Step3:对于任意测试样本X, [0090] Step3: For any test sample X,

Figure CN102663374BD00091

[0092] 其中,Y为类别集合,ye Y;t为迭代次数,取值为1...T。 [0092] wherein, Y is a category set, ye Y; t is the iteration number, a value of 1 ... T.

[0093] Step3中公式进一步说明如下:对于任意测试样本X,各个基分类器的分类情况如下: [0093] Step3 further described in the following formula: test sample for any X, classification of each group classification is as follows:

[0094] Ii1 (X) - > I, h2 (X) - > 2, h3 (X) - > 2, h4 (X) - > 3, h5 (X) - > I, h6 (x) - > 2, h7 (x)- >5, h8(x)_ > 19, h9(x) - > I, h10 (x) - > 3。 [0094] Ii1 (X) -> 2, h3 (X) - -> 2, h4 (X) -> 3, h5 (X) -> I, h6 (x) -> 2> I, h2 (X) , h7 (x) -> 5, h8 (x) _> 19, h9 (x) -> I, h10 (x) -> 3.

[0095] 类别I的权值之和为:BJaJa9 ; Weight values ​​[0095] for the class I and: BJaJa9;

[0096] 类别2的权值之和为:a2+a3+a6 ; And the sum of weights [0096] Category 2 is: a2 + a3 + a6;

[0097] 类别3的权值之和为:a4+a1Q ; [0097] Weight of 3 and the sum of categories of: a4 + a1Q;

[0098] 类别5的权值之和为:a7 ; [0098] The weights Category 5 and is: a7;

[0099] 类别19的权值之和为:a8 ; [0099] Class 19 and the sum of weights is: a8;

[0100] 其余类别标号权值为:0。 [0100] Reference numeral a weight of the remaining categories: 0.

[0101] 最终测试样本的类别标号为:权值之和最大对应的类别标号, [0101] Reference numeral final category of test sample: the sum of the maximum weight category corresponding reference numerals,

[0102] 即^axiaJaifHa9, a2+a3+a6, a4+a10, a7, a8}所对应的类别标号。 [0102] That ^ axiaJaifHa9, a2 + a3 + a6, a4 + a10, a7, a8} category corresponding reference numerals. 当有两个或多个最大的权值之和,则给测试样本赋予这几种类别中的任意一个类别标号。 When there are two or more values ​​of the maximum power and then to test this sample imparting any one of several categories category label.

[0103] 下面详述说明本发明的实验结果: [0103] The following detailed description of the experimental results of the present invention:

[0104] 本发明的实验采用的数据库是中科院自动化所提供的CASIA Dataset A数据库,该数据库共包含20个人,每个人分别有3个不同的视角(0°、45°和90° ),每个视角分别拍摄4个序列,总计包括240个序列。 [0104] Database experiment of the present invention is to provide CASIA CASIA Dataset A database, which contains a total of 20 individuals, respectively, each of three different viewing angle (0 °, 45 ° and 90 °), each Perspective sequences were captured 4, comprising a total of 240 sequences. 这些彩色图像序列以25帧每秒的速率拍摄,原始尺寸为352*240像素点,平均长度约为100帧。 These color image sequences at a rate of 25 frames per second shooting, the original size is 352 * 240 pixels, an average length of about 100.

[0105] 为了验证MCAB算法,我们在NLPR库中,在每个人相同视角的4个序列中任意选择三个序列作为训练数据,剩余的一个序列做测试。 [0105] In order to verify MCAB algorithm, we NLPR database, arbitrarily selected three sequences in perspective of the same four sequences as training data in each, the remaining one sequence for testing.

[0106] 表1是20个人在三种不同视角(0°,45° ,90° )条件下,采用Haar小波作为小波包分解的基函数,对小波包分解的每个子图像分别进行(2D)2PCA变换,最后采用前述的基于L2范式的最近邻分类器进行识别的结果,最后我们统计了迭代次数不同时的识别率。 [0106] Table 1 is 20 people at three different viewing angles (0 °, 45 °, 90 °) conditions using Haar wavelet basis functions as wavelet packet decomposition, each sub-image for wavelet packet decomposition of the (2D), respectively 2PCA transformation, the final result of the use of identification based on L2 paradigm nearest neighbor classifier foregoing, and finally we counted the number of iterations recognition rate is not simultaneous.

[0107]表1WPD+ (2D) 2PCA 的识别率 [0107] Table 1WPD + (2D) of the recognition rate 2PCA

[0108] [0108]

Figure CN102663374BD00101

[0109] 表2是以表1的识别结果为基础,然后调用MCAB算法进行识别的结果,最后我们统计了迭代次数不同时的识别率。 [0109] Table 2 is based on Table 1 is based on the recognition result, and then call MCAB result of the recognition algorithm, and finally we counted the number of iterations is not the same recognition rate.

[0110] 表2MCAB算法的识别率 [0110] Algorithm recognition rate table 2MCAB

[0111] [0111]

Figure CN102663374BD00102

[0112] 从表2可以看出:相比于单纯使用小波分解的各子图像系数进行识别,当迭代次数T较大时,MCAB算法的识别性能更高,且对不同视角(0°,45°,90° )都能获得较高的识别率。 [0112] As can be seen from Table 2: Compared to simply using wavelet decomposition coefficients of each sub-image to identify, when a large number of iterations T, MCAB higher recognition performance algorithms, and for different viewing angles (0 °, 45 °, 90 °) can obtain a higher recognition rate.

Claims (2)

1.一种基于多步态特征属性的多类别Bagging步态识别方法,包括人体步态序列的预处理、特征提取,最后根据最近邻分类原理与基于多步态特征属性的多类别Bagging即MCAB算法相结合,把测试样品归到相应的类中,其特征在于:具体步骤如下: 步骤一、预处理对已经背景分离的人体运动目标图像依次进行形态学处理、目标提取、以及图像归一化处理; 步骤二、特征提取经过步态周期的检测,建立步态能量图,将步态能量图作为不同步态序列的代表,利用WPD+(2D) 2PCA方法计算20个特征的识别率;WPD即二级小波包分解; 步骤三、分类识别选择步骤二中识别率大于等于50%的特征作为待重抽样的原始特征集合AttributeSet,利用MCAB算法对测试样本进行最终识别;具体如下(1)在MCAB算法中,首先需要对原始训练实例的属性集合进行η次有放回的重抽样,η为经过步骤二后识 A multi-class Bagging gait recognition method based on multi-attribute Gait, includes a pre-sequence of human gait, feature extraction, and finally in accordance with the principles of multi-nearest neighbor classifier based on multiple categories Bagging Gait properties i.e. MCAB algorithm are combined, the test sample is normalized to the respective class, characterized in that: the following steps: a step, pretreatment of human motion has been separated from the background object image sequentially morphological processing, object extraction and image normalization process; step two, the feature extraction tested gait cycle, the establishment of energy FIG gait, the gait energy image as representing different sequences of gait, the recognition rate is calculated using the 20 features WPD (2D) 2PCA method +; WPD i.e. two wavelet packet decomposition; step three, classification selection step two recognition rate of 50% or greater as an original feature set of features to be re-sampling AttributeSet, the test sample using the final recognition algorithm MCAB; as follows (1) at MCAB algorithm, first needs to properties of the original collection of training examples of [eta] re-sampling with replacement times, [eta] is a two step after identifying 率大于等于50%的属性个数,这η个属性可能会出现多次,有的可能一次也不会出现,由这η个属性构成新的训练实例,再由这些新的训练实例构造弱分类器; (2)利用MCAB算法进行分类识别1:确定重抽样属性的个数n,确定迭代次数T ;其中η为识别率大于等于50%的属性的个数,T取值为10~100 ; ii:For t=l:T执行以下3步; a):从Attribute Set中重抽样η个属性得到属性集At,对训练集S中的每个样本,只取At中的属性,得到新的属性St ; b):在属性St上用I NN算法训练出基分类器ht(x) — Y; c):计算权值 Rate is greater than or equal to 50% of the number of attributes, which η attribute may appear multiple times, once some may not occur by this new configuration attributes η training examples, then the configuration of these new training examples of weak classifiers ; a (2) using a classification algorithm MCAB: determining the number of re-sampling attribute n, T determines the number of iterations; wherein η is a recognition rate more than 50% equal to the number of attributes, T value of from 10 to 100; ii: for t = l: T 3 perform the following steps; a): the re-sampling from the attribute set attributes obtained η At attribute set, the training set S of each sample, just take At attributes, get new properties St; b): St on the property with a group I NN algorithm trained classifier ht (x) - Y; c): calculated weight
Figure CN102663374BC00021
其中,rt=ht (χ)正确分类的个数减去ht(x)错误分类的个数; ii1:对于任意测试样本X,集成分类器H(X)为: Wherein, rt = number ht (χ) obtained by subtracting the number of correct classification misclassified ht (x); ii1: test sample for any X, the integrated classifier H (X) is:
Figure CN102663374BC00022
其中,Y为类别集合,ye Y ;t为迭代次数,取值为I…T。 Wherein, Y is a category set, ye Y; t is the iteration number, a value of I ... T.
2.根据权利要求1所述的基于多步态特征属性的多类别Bagging步态识别方法,其特征是,所述的步态能量图的提取步骤如下: 在进行步态周期检测之后,通过对一个周期内的步态序列图像处理生成的步态能量图为:/ 式中,G(x, y)是这个周期序列的步态能量图,N1是完整步态周期序列的长度,Bt(x, y)是一个周期中的第t个步态图像,χ, y代表二维图像平面坐标。 The multi-class recognition method based on multi-Bagging gait gait characteristic properties according to claim 1, characterized in that said extraction step gait energy map is as follows: after performing a gait cycle detection, by gait generating a sequence of image processing in the gait cycle energy Pictured: / formula, G (x, y) is the periodic sequence gait energy image-on, N1 is the length of the complete sequence of the gait cycle, Bt (x , y) is the image of a t-th gait cycle, χ, y coordinates of the representative two-dimensional image plane.
CN 201210134185 2012-04-28 2012-04-28 Multi-class Bagging gait recognition method based on multi-characteristic attribute CN102663374B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN 201210134185 CN102663374B (en) 2012-04-28 2012-04-28 Multi-class Bagging gait recognition method based on multi-characteristic attribute

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN 201210134185 CN102663374B (en) 2012-04-28 2012-04-28 Multi-class Bagging gait recognition method based on multi-characteristic attribute

Publications (2)

Publication Number Publication Date
CN102663374A CN102663374A (en) 2012-09-12
CN102663374B true CN102663374B (en) 2014-06-04

Family

ID=46772858

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 201210134185 CN102663374B (en) 2012-04-28 2012-04-28 Multi-class Bagging gait recognition method based on multi-characteristic attribute

Country Status (1)

Country Link
CN (1) CN102663374B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103577805A (en) * 2013-10-28 2014-02-12 汕头大学 Gender identification method based on continuous gait images
CN104794449B (en) * 2015-04-27 2017-11-28 青岛科技大学 FIG obtaining energy hog Gait Characteristics and identification methods based on human
CN104820825B (en) * 2015-04-27 2017-12-22 北京工业大学 Face Recognition Based on Adaboost Algorithm Optimization

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6628821B1 (en) 1996-05-21 2003-09-30 Interval Research Corporation Canonical correlation analysis of image/control-point location coupling for the automatic location of control points
CN101630364A (en) 2009-08-20 2010-01-20 天津大学 Method for gait information processing and identity identification based on fusion feature

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6628821B1 (en) 1996-05-21 2003-09-30 Interval Research Corporation Canonical correlation analysis of image/control-point location coupling for the automatic location of control points
CN101630364A (en) 2009-08-20 2010-01-20 天津大学 Method for gait information processing and identity identification based on fusion feature

Also Published As

Publication number Publication date
CN102663374A (en) 2012-09-12

Similar Documents

Publication Publication Date Title
Wang et al. Action recognition with trajectory-pooled deep-convolutional descriptors
Belongie et al. Matching with shape contexts
Arica et al. BAS: a perceptual shape descriptor based on the beam angle statistics
Bashir et al. Gait recognition without subject cooperation
Leibe et al. Interleaving object categorization and segmentation
Paisitkriangkrai et al. Strengthening the effectiveness of pedestrian detection with spatially pooled features
Urtasun et al. 3D tracking for gait characterization and recognition
Evangelidis et al. Skeletal quads: Human action recognition using joint quadruples
Lu et al. Gait recognition for human identification based on ICA and fuzzy SVM through multiple views fusion
Leibe et al. Combined object categorization and segmentation with an implicit shape model
Liu et al. Learning expressionlets on spatio-temporal manifold for dynamic facial expression recognition
CN101350069B (en) Computer implemented method for constructing classifier from training data and detecting moving objects in test data using classifier
Singh et al. Muhavi: A multicamera human action video dataset for the evaluation of action recognition methods
Li et al. Structured sparse subspace clustering: A unified optimization framework
Minhas et al. Human action recognition using extreme learning machine based on visual vocabularies
Guo et al. Action recognition from video using feature covariance matrices
Nam et al. Local decorrelation for improved pedestrian detection
Sankaranarayanan et al. Triplet probabilistic embedding for face verification and clustering
Lu et al. Shape guided contour grouping with particle filters
Du et al. Skeleton based action recognition with convolutional neural network
Liu et al. Face recognition using dual-tree complex wavelet features
Wen et al. Handwritten Bangla numeral recognition system and its application to postal automation
Casanova et al. Plant leaf identification using Gabor wavelets
WO1997044757A2 (en) Principal component analysis of image/control-point location coupling for the automatic location of control points
Zhao et al. Online human gesture recognition from motion data streams

Legal Events

Date Code Title Description
C06 Publication
C10 Entry into substantive examination
C14 Grant of patent or utility model
CF01 Termination of patent right due to non-payment of annual fee