-
目前LLE(局部线性嵌入)算法[1-2]广泛地应用于图像分类与目标识别问题中[3],但经典的LLE算法难以学习多个流形[4-8].本文在文献[9]的LLE算法的基础上,设计了基于LLE的多流形学习方法,通过局部非线性、多流形的方法学习并保留数据集的类结构.假设一个向量数据集为x1,x2,…,xN,ε是一个正实数,k是一个正整数,N(xi)表示xi的近邻点,包含了与xi距离最近的k个向量,即:对于每个xj∈N(xi),其欧式距离‖ xi- xj‖小于ε,将数据样本xi线性地表示为N(xi)中的近邻点.通过最小化(1)式计算权重值{wij}xj∈N(xi):
约束条件为∑xj∈N(xi)wij=1,wij>0;如果xj
$\notin $ N(xi),则wij=0.可将(1)式转换为:ε2=[xi-$\sum\limits_{j = 1}^k {} $ wij xj]2=[$\sum\limits_{j = 1}^k {} $ wij(xi-xj)]2=$\sum\limits_{jk} {} $ wijwik Gjk,式中Gjk=(xi- xj)T(xi-xk)是局部Gram矩阵,其中xj,xk∈N(xi).使用拉格朗日乘法计算重构权重${w_{ij}} = \frac{{\sum\nolimits_k {\mathit{\boldsymbol{G}}_{jk}^{-1}} }}{{\sum\nolimits_{lm} {\mathit{\boldsymbol{G}}_{lm}^{-1}} }}$ ,使用重构权重搜索d维的向量y1,y2,…,yn∈$\mathbb{R}$ d,这些向量定义了$\mathbb{R}$ d中原数据集的d维嵌入,其中d < D,通过最小化(2)式可获得d维的向量.其中
将(2)式改写为以下的矩阵形式:
式中:W是n×n的权重矩阵,mN×N=(I- Wn×n)T(I - Wn×n),Yi均正交.使用拉格朗日乘法计算最优解:
用(4)式对Yk(k=1,2,…,d)求导,YkTm-λkYkT=0则说明了mYk=λkYk,{λk}为m的特征值,{Yk}则是对应的特征向量.通过搜索m底部的特征向量可将E2(Y)最小化. Y0T的一个特征值为λ0=0,所以应当获得m底部的d+1个特征值及其特征向量,然后忽略Y0.
A Tracking Algorithm of Surveillance Video Based on Enhanced Multi-Manifold Learning
-
摘要: 提出了一种多流形局部线性嵌入的流形学习算法,为每个类的流形学习过程设计了一种监督的近邻点选择方法,将流形-流形距离作为度量指标,搜索最优的低维空间.在视频追踪算法中对外部数据库进行图像训练预处理,为人脸检测建立级联分类器,利用均值粒子滤波器结合跟踪校正策略对人脸图像实时跟踪,采用多流形训练的结果从视频流的人脸集中检测出追踪的目标人脸.仿真实验结果表明本算法对不同的数据集均获得了较高的检测率与较高的计算效率.Abstract: The traditional manifold learning algorithms cannot preserve the structure of individual manifolds during multi-class-multi-manifold learning problems, and have obvious influence to the performance of multi-classes identification problems, thus a manifold learning algorithm of multi-manifold locally linear embedding has been proposed. A supervised neighborhood selection method has been designed by this multi-manifold learning algorithm for the manifold learning procedure of each class, and the distances of manifold to manifold have been set as the metric to search the optimal low dimensional space. Image training preprocess of external database has been realized during the video tracking algorithm, the cascade classifier has been constructed for face detection, and the mean particle filter combined with tracking correction strategy has been adopted for real-time tracking of face images, the results of multi-manifold learning training are used to identify the target faces from the face set of video stream. Simulation experiments are implemented based on the large scale video datasets, the results show that the proposed algorithm realizes a high detection accuracy and a high computational efficiency to different video datasets.
-
Key words:
- manifold learning /
- locally linear embedding /
- surveillance video /
- target detection /
- target tracking /
- dimensional reduction .
-
表 算法1 基于局部线性嵌入的多流形学习算法
1.读取数据集X={X1,X2,…,XN}; 2.FOREACH数据分类Cq IN X DO 3. FOREACH数据点xi IN Cq DO 4.搜索kq个近邻点Nε,kq(xi); 5.通过最小化(1)式计算权重wij; 6.搜索权重矩阵W,每个元素的权重为wij; 7.ENDFOR 8.ENDFOR 9.搜索N×N稀疏矩阵F=(I-W)T(I-W); 10.设置d=1,MinDist=0; 11./*迭代地运行以下子程序直至不同分类之间的流形距离最小化*/ 12.WHILE(MinDist < εmmd & & d < N) DO 13.从M个特征向量中搜索d+1个最小特征向量构成集合Y,忽略最小的特征向量; 14. /*从Y中搜索流形嵌入坐标Yq*/ 15.设置initialLength=0; 16. FOREACH数据类Xq IN X DO 17. Yq=Y((initialLength+1),(initialLength+length(Xq))); 18. initialLength=initialLength+length(Xq); 19. ENDFOR 20./*计算流形与流形之间的距离*/ 21. FOREACH Yq,Yr IN Y DO 22.根据定义6搜索D(Yq,Yr); 23. ENDFOR 24./*搜索流形之间的最小距离*/ 25.搜索最小距离:MinDist=min(D(Yq,Yr)); 26. d=d+1; 27. ENDWHILE 表 1 视频识别算法对YTC数据集与COX数据集的识别率/%
视频检测算法 YTC COX12 COX13 COX23 COX21 COX31 COX32 MaxMD 52.6 36.4 19.6 8.9 27.6 19.1 9.6 CDL 69.7 78.4 85.3 79.7 75.6 85.8 81.9 LMKML 70.3 66.0 71.0 56.0 74.0 68.0 60.0 SGM 52.0 26.7 14.3 12.4 26.0 19.0 10.3 GMM 61.0 30.1 24.6 13.0 28.9 31.7 18.9 本算法 73.3 95.1 96.3 94.2 92.3 95.4 94.5 表 2 目标检测算法的计算时间/s
视频检测算法 训练时间 测试时间 MaxMD 无训练 0.1 CDL 433.3 2.6 LMKML 245.3 0.5 SGM 11.9 0.1 GMM 42.3 1.9 本算法 27.1 0.1 -
[1] 李燕燕, 闫德勤, 刘胜蓝, 等.一种基于局部线性嵌入的多流形学习算法[J].小型微型计算机系统, 2012, 33(8):1795-1799. doi: 10.3969/j.issn.1000-1220.2012.08.033 [2] 高翠珍, 胡建龙, 李德玉.保持局部邻域关系的增量Hessian LLE算法[J].计算机科学, 2012, 39(4):217-219. doi: 10.3969/j.issn.1002-137X.2012.04.050 [3] 晏勇.基于SKLLE和SVM的人脸表情识别[J].西南师范大学学报(自然科学版), 2014, 01(1):55-60. doi: http://xbgjxt.swu.edu.cn/jsuns/jscnuhhse/ch/reader/view_abstract.aspx?file_no=X2014-01-055&flag=1 [4] 王鑫, 沃波海, 管秋, 等.基于流形学习的人体动作识别[J].中国图象图形学报, 2014, 19(6):914-923. doi: http://d.old.wanfangdata.com.cn/Thesis/D01014489 [5] ZHANG S Q.Enhanced Supervised Locally Linear Embedding[J].Pattern Recognition Letters, 2009, 30(13):1208-1218. doi: 10.1016/j.patrec.2009.05.011 [6] ZHANG W.An Improved Locally Linear Embedding Algorithm by Using Recurrent Neural Networks[J].Journal of Computational and Theoretical Nanoscience, 2016, 13(1):425-437. doi: 10.1166/jctn.2016.4823 [7] ZHANG L, LENG Y, YANG J, et al.Supervised Locally Linear Embedding Algorithm Based on Orthogonal Matching Pursuit[J].IET Image Processing, 2015, 9(8):626-633. doi: 10.1049/iet-ipr.2014.0841 [8] 任剑锋, 梁雪, 李淑红.基于非线性流形学习和支持向量机的文本分类算法[J].计算机科学, 2012, 39(1):261-263. doi: 10.3969/j.issn.1002-137X.2012.01.060 [9] CHOJNACKI W, BROOKS M J.A Note on the Locally Linear Embedding Algorithm[J].International Journal of Pattern Recognition and Artificial Intelligence, 2009, 23(08):1739-1752. doi: 10.1142/S0218001409007752 [10] 胡一帆, 胡友彬, 李骞, 等.基于视频监控的人脸检测跟踪识别系统研究[J].计算机工程与应用, 2016, 52(21):1-7. doi: 10.3778/j.issn.1002-8331.1602-0035 [11] 吴文通, 李元祥, 韦邦合, 等.局部测地距离估计的增量等距特征映射算法[J].上海交通大学学报(自然科学版), 2013, 47(7):1082-1086. doi: http://d.old.wanfangdata.com.cn/Periodical/shjtdxxb201307014 [12] KAYO O.Locally Linear Embedding Algorithm——Extensions and Applications[M].Oulu:Oulu University Press, 2006. [13] KIM M, KUMAR S, PAVLOVIC V, et al.Face Tracking and Recognition with Visual Constraints in Real-World Videos[C]//2008 IEEE Computer Society Conference on Computer Vision and Pattern Recognition.New York: IEEE Computer Society Press, 2008. [14] HUANG Z, SHAN S, ZHANG H, et al.Benchmarking Still-to-Video Face Recognition via Partial and Local Linear Discriminant Analysis on COX-S2V Dataset[C]//Asian Conference on Computer Vision.New York: Springer-Verlag, 2012: 589-600. [15] LU J, WANG G, MOULIN P.Image Set Classification Using Holistic Multiple Order Statistics Features and Localized Multi-kernel Metric Learning[C]//IEEE International Conference on Computer Vision.New York: IEEE Computer Society Press, 2013: 329-336. [16] doi: http://d.old.wanfangdata.com.cn/OAPaper/oai_arXiv.org_0805.2368 GRETTON A.A Kernel Two-Sample Test[J].Journal of Machine Learning Research, 2012, 13(1):723-773. [17] WANG R, GUO H, DAVIS L S, et al.Covariance Discriminative Learning: A Natural and Efficient Approach to Image Set Classification[C]//IEEE Conference on Computer Vision and Pattern Recognition.New York: IEEE Computer Society Press, 2012: 2496-2503. [18] doi: http://d.old.wanfangdata.com.cn/NSTLQK/NSTL_QKJJ0232544303/ MLIKI H, FENDRI E, HAMMAMI M.Face Recognition Through Different Facial Expressions[J].Journal of Signal Processing Systems, 2015, 81(3):1-14. [19] doi: http://www.wanfangdata.com.cn/details/detail.do?_type=perio&id=090d23b8abc1a53db6770694bc9d87a2 CUI Z, CHANG H, SHAN S, et al.Joint Sparse Representation for Video-Based Face Recognition[J].Neurocomputing, 2014, 135(8):306-312.