文章快速检索 高级检索

1. 北京交通大学 信息科学研究所, 北京 100044;
2. 北京交通大学 现代信息科学和网络技术北京市重点实验室, 北京 100044

Slow feature extraction algorithm of human actions in video
CHEN tingting1,2 , RUAN Qiuqi1, AN Gaoyun1
1. Institute of Information Science, Beijing Jiaotong University, Beijing 100044, China;
2. Beijing Key Laboratory of Advanced Information Science and Network technology, Beijing Jiaotong University, Beijing 100044, China
Abstract:Extracting important and distinguishable features from complex human actions is the key for human actions analysis. In recent years, classical feature analysis methods are mostly linear feature analysis technologies, which result in error results for non-linear processing. this paper proposes a method of extracting slow features. First, the image sequence of frame difference was obtained by the difference between the consecutive frames and some feature points of selected beginning frame were detected. Next, the feature points were tracked by optical flow method and the training cuboids were collected. Finally, the slow feature functions were learned with the collected training cuboids, then the slow features could be extracted and represented. In the experiment, slow features of each action were extracted and compared with each other. the results show that the extracted slow features vary slowly with time and action interclass has good discrimination, which suggests that this method can extract slow features from human actions effectively.
Key words: human action     training cuboids     slow feature function     slow feature     frame difference

1 慢特征分析(SFA)原理

1.1 SFA的数学描述

1.2 SFA算法具体实现步骤

1)如果变换是线性的，即gj(x)=wjTx，其中x是输入，wj是权值。不失一般性，假设x均值为0，方差为1，即〈xt=0，〈x2t=1。因为yj(t)=gj(x(t))=wjTx(t)，所以方程(2)中，

B=〈xxTt,则〈yi(t)yj(t)〉=wiTBwj，只要选择合适的权值，使得wiTBwj=δij，则式(4)中的限制条件被满足。

,则，把式(3)中的限制条件整合到目标函数(1)中，则有

2)非线性变换可以视为非线性空间的线性变换[5]。函数h(x)的非线性扩展定义为

a)使用非线性函数h(x)对原始信号进行扩展，并把h(x)归一化，使其均值为0，即z:=h(x)－h0,其中h0=〈h(x)〉t，归一化使条件(2)有效。本文使用二次扩展，即h(x)=(x1,x2,...,xI,x1x1,x1x2,...,xIxI)。

b)解决一般化的特征值问题AW=BWΛ，其中，B:=〈zzTt。假设矩阵A和矩阵B的维数都是M，前K个特征向量w1,w2,...,wK(KM)和最小的特征值相联系，即λ1λ2≤…≤λK，对应的非线性慢特征函数g1(x),g2(x),...,gK(x)：

1.3 改进的慢特征分析算法(D-SFA)

2.1 收集训练立方体

 图 1 训练立方体的获取过程 Fig. 1 the process of obtaining training cuboids
2.2 D-SFA算法提取慢特征

2.3 ASD特征表示

3 实验结果及分析 3.1 数据库介绍

 图 2 Weizmann人体行为数据库样图 Fig. 2 Sample images of each type of action in the Weizmann database
3.2 实验结果及分析

 图 3 SFA算法提取的慢特征随时间变化 Fig. 3 the diagram of slow features extracted by SFA vary with time

 图 4 D-SFA算法提取的慢特征随时间变化 Fig. 4 the diagram of slow features extracted by D-SFA vary with time

 Action Walk Run Jump Pjump Bend Wave1 Wave2 Skip Jack Side Walk 0.070 900 151.790 0 133.570 0 132.840 0 127.840 0 128.470 0 124.620 0 135.160 0 160.230 0 145.480 0 Run 229.140 0 0.062 400 218.820 0 229.860 0 224.930 0 237.400 0 230.600 0 232.590 0 252.640 0 232.000 0 Jump 699.784 5 673.711 9 0.012 000 596.557 3 574.379 3 594.044 5 583.238 9 637.634 4 657.620 8 605.668 7 Pjump 599.942 5 657.441 9 491.771 5 0.001 600 541.067 3 553.955 2 532.864 2 633.454 2 747.772 2 637.911 7 Bend 671.993 2 810.703 0 716.960 0 680.960 2 0.000 700 936.513 2 640.215 0 646.039 7 836.871 7 669.577 6 Wave1 111.170 0 103.880 0 109.310 0 109.600 0 993.861 6 0.002 800 106.190 0 929.660 7 100.890 0 908.275 0 Wave2 671.422 7 658.995 6 637.138 4 596.024 6 512.697 4 710.915 5 0.000 784 677.386 6 605.973 7 707.146 2 Skip 154.830 0 152.960 0 135.650 0 154.300 0 144.120 0 144.650 0 136.380 0 0.062 100 159.480 0 169.340 0 Jack 199.040 0 221.760 0 186.610 0 179.800 0 180.100 0 195.600 0 194.540 0 218.020 0 0.011 900 213.670 0 Side 186.260 0 200.920 0 173.890 0 171.580 0 169.160 0 165.030 0 172.510 0 181.580 0 198.180 0 0.035 200
4 结束语

 [1] VENKAtASUBRAMANIAN V, RENGASWAMY R, KAVURI S N, et al. A review of process fault detection and diagnosis: Part III: process history based methods[J]. Computers & Chemical Engineering, 2003, 27(3): 327-346. [2] CHERRY G A, QIN S J. Multiblock principal component analysis based on a combined index for semiconductor fault detection and diagnosis[J]. IEEE transactions on Semiconductor Manufacturing, 2006, 19(2): 159-172. [3] DUNIA R, QIN S J. Joint diagnosis of process and sensor faults using principal component analysis[J]. Control Engineering Practice, 1998, 6(4): 457-469. [4] SCHLKPOF B, SMOLA A, MVLLER K R. Nonlinear component analysis as a kernel eigenvalue problem[J]. Neural Computation, 1998, 10(5): 1299-1319. [5] WISKOtt L, SEINOWSKI t L. Slow feature analysis: unsupervised learning of invariances [J]. Neural Computation, 2002, 14(4): 715-770. [6] BERKES P, WISKOtt L. Slow feature analysis yields a rich repertoire of complex cell properties[J]. Journal of Vision, 2005, 5(6): 579-602. [7] XIA Qi, GAO Jianbin, XU Chunxiang. A new watermarking algorithm based on slowly feature analysis[C]//International Conference on Apperceiving Computing and Intelligence Analysis. Chengdu, China, 2008: 70-72. [8] GAO Jianbin, LI Jianping, XIA Qi. Slowly feature analysis of Gabor feature for face recognition[C]//2008 International Conference on Apperceiving Computing and Intelligence Analysis. Chengdu, China, 2008: 177-180. [9] HUANG Yaping, ZHAO Jiali, tIAN Mei, et al. Slow feature discriminant analysis and its application on handwritten digit recognition [C]//International Joint Conference on Neural Networks. Atlanta, USA, 2009: 1294-1297. [10] MA Kuijun, tAO Qing, WANG Jue. Nonlinear blind source separation using slow feature analysis with random features[C]//2010 20th International Conference on Pattern Recognition. Istanbul, turkey, 2010: 830-833. [11] KVHNL t, KUMMERt F, FRItSCH J. Monocular road segmentation using slow feature analysis[C]//2011 IEEE Intelligent Vehicles Symposium (IV). Baden-Baden, Germany, 2011: 800-806. [12] DENG Xiaogang, tIAN Xuemin, HU Xiangyang. Nonlinear process fault diagnosis based on slow feature analysis[C]//2012 10th World Congress on Intelligent Control and Automation. Beijing, China, 2012: 3152-3156. [13] ZHANG Zhang, tAO Dacheng. Slow feature analysis for human action recognition[J]. IEEE transactions on Pattern Analysis and Machine Intelligence, 2012, 34(3): 436-450. [14] 王丽辉, 袁保宗. 三维散乱点云模型的特征点检测[J]. 信号处理, 2011, 27(6): 932-938.WANG Lihui, YUAN Baozong. Feature point detection for 3D scattered point cloud model[J]. Signal Processing, 2011, 27(6): 932-938. [15] 马龙, 王鲁平, 陈小天, 等. 噪声环境下光流场估计方法[J]. 信号处理, 2012, 28(1): 87-91.MA Long, WANG Luping, CHEN Xiaotian, et al. Determining optical flow field in the presence of noise[J]. Signal Processing, 2012, 28(1): 87-91 [16] 江志军, 易华蓉. 一种基于图像金字塔光流的特征跟踪方法[J]. 武汉大学学报:信息科学版, 2007, 32(8): 680-683.JIANG Zhijun, YI Huarong. An image pyramid-based feature detection and tracking algorithm[J]. Geomatics and Information Science of Wuhan University, 2007, 32(8): 680-683.
DOI:10.3969/j.issn.1673-4785.201407002

0

#### 文章信息

CHEN tingting, RUAN Qiuqi, AN Gaoyun

Slow feature extraction algorithm of human actions in video

CAAI transactions on Intelligent Systems, 2015, 10(03): 381-386.
DOI:10.3969/j.issn.1673-4785.201407002