IPSJ SIG Technical Report Vol.2011-CVIM-175 No /1/ , Evaluation of Age Estimation by Gaussian Process Regression

Size: px

Start display at page:

Download "IPSJ SIG Technical Report Vol.2011-CVIM-175 No /1/ , Evaluation of Age Estimation by Gaussian Process Regression"

あいりこやぎ
5 years ago
Views:

1 , Evaluation of Age Estimation by Gaussian Process Regression using a Very Large-scale Gait Database MAYU OKUMURA, 1 YASUSHI MAKIHARA 1 and YASUSHI YAGI 1 While many gait recognition methods have been recently proposed, the studies about gait-based age estimation is unexplored and the number of subjects and the distribution of subjects ages of the existing gait databases are insufficient for this purpose. In this paper we evaluate the age estimation by Gaussian process regression using a very large-scale gait database which includes 1,730 subjects with ages ranging from 2 to 94 years. First, gait feature as an input vector is related with an age as an output scalar in the Gaussian process regression framework, and the parameters of Gaussian process regression are estimated by the training pairs of input vectors and output scalars. Then, given a new gait feature, a Gaussian distribution of an age for the gait defined by expectation and variance, is estimated with the trained parameters and the training pairs. In the experiment, 879 and 851 subjects in the large-scale database are used for training and testing, respectively. As a result, the performance of the age estimation for 30s females and 40s males was the best. 1. ( ) 1) 2)3)4)5)6) 7)8) 9)10)11) 12) USF 13) Soton (Southampton) 14) CASIA 15) 16)17)18)19)20)21) ,730 22) 1 The Institute of Scientific and Industrial Research, Osaka University 1 c 2011 Information Processing Society of Japan

2 23) Lanitis 24) 4 25) 26) 27) 23) 36) 35) Fu 28) Guo 29) Xiao 30) Geng 33) PCA 32) 31) Zhang 35) 2.2 Daves 9) (3-5 ) (30-52 ) Begg 10) ( ) ( ) 11) 6 80 (6-12 ) (13-64 ) (65-80 ) 3 37) 4 25 Lu 12) m 3 m 4 m 2 (1) (2) pixel 30 fps Dive Into the Movie (DIM) 38) 39) ,730 ( ) 3 2 c 2011 Information Processing Society of Japan

情報処理学会研究報告 (a) Camera 1 (a) Top view (b) Side view 図 3 被験者の性別年齢分布 Fig. 3 Statistics of subjects gender and age.

スは以下のような長所が挙げられる (1) 被験者数は既存の大規模歩容データベースの約 10 倍以上であるこの被験者数に図 4 GSV の例 Fig. 4 An example of GSV.

予め定めた高さになるようにアスペクト比を保ったままスケーリングし水必要な特性である平中心の位置合わせを行う本研究ではスケーリング後の画像サイズを高さ 32 画素幅被験者の年齢分布は幅広く 2 歳から 94

の歩容シルエットボリューム (GSV : Gait Silhouette Volume) を作成する図 4 に GSV の作め歩容に基づく年齢推定や各年代間での歩容認証性能の比較などにおいてより統成例を示す 4.

1 歩容シルエットボリュームの作成 T (N) x,y n=0 g(x, y, n)g(x, y, n + N) C(N) = T (N) T (N) x,y n=0 g(x, y, n)2 x,y n=0 g(x, y, n +

3 情報処理学会研究報告 (a) Camera 1 (a) Top view (b) Side view 図 3 被験者の性別年齢分布 Fig. 3 Statistics of subjects gender and age. 図 1 歩容撮影システム Fig. 1 Gait measurement system. (b) Camera 2 図 2 撮影画像例 Fig. 2 Examples of the captured images. スは以下のような長所が挙げられる (1) 被験者数は既存の大規模歩容データベースの約 10 倍以上であるこの被験者数に図 4 GSV の例 Fig. 4 An example of GSV. より年齢推定の評価の信頼性が著しく改善される (2) (3) 男性と女性の割合が同程度になるように構築しているこれは性別に制限を設けることのできない実世界での年齢推定を想定する際より信頼性の高い評価をするために中心を算出し予め定めた高さになるようにアスペクト比を保ったままスケーリングし水必要な特性である平中心の位置合わせを行う本研究ではスケーリング後の画像サイズを高さ 32 画素幅被験者の年齢分布は幅広く 2 歳から 94 歳までとなっている特に子供の人数は大 22 画素としたそして最後にそれらの画像を時間軸方向に重ね合わせることで時空間人の人数に匹敵し 50 歳までの年代毎に 100 人以上の被験者を有しているそのたの歩容シルエットボリューム (GSV : Gait Silhouette Volume) を作成する図 4 に GSV の作め歩容に基づく年齢推定や各年代間での歩容認証性能の比較などにおいてより統成例を示す 4.2 歩行周期の検出計的に信頼性の高い結果を得ることができる歩行は周期運動であるため歩行周期 Ngait を GSV の時間軸方向への正規化自己相関を 4. 歩容特徴抽出最大にする時間シフト量として以下のように算出する 4.1 歩容シルエットボリュームの作成 T (N) x,y n=0 g(x, y, n)g(x, y, n + N) C(N) = T (N) T (N) x,y n=0 g(x, y, n)2 x,y n=0 g(x, y, n + N)2 本研究ではアピアランスベースの歩容特徴を用いるためまず前処理として歩容シルエット抽出を背景差分とグラフカットに基づく領域分割により抽出する (1) 次に人物領域のスケーリングと位置合わせを行う各フレームから人物領域の高さと水平 Information Processing Society of Japan

4 5 3 (GEI) 1 2 Fig. 5 Examples of frequency-domain features among three subjects. In each subject, left image indicates direct current element (GEI), middle and right images indicate 1- and 2-times frequency elements. 6 Fig. 6 Diagram for Gaussian process. T (N) = N total N 1 (2) g(x,y,n) n (x,y) GSV C(N) N GSV N total 30 fps N [20, 40] frame ([0.66, 1.33] ) ) 41)42) GSV (3) (i+1)n gait 1 G i (x,y,k) = n=in gait g(x,y,n)e jω 0kn (3) ω 0 = 2π/N gait N gait G i (x,y,k) i GSV k (DFT) DFT (4) A i (x,y,k) = 1 N gait G i (x,y,k) (4) A i (x,y,k) N gait G i (x,y,k) 5 (k = 0) ( ) (k = 1,2) N A 2,112 (= ) (k = 0) 41) ( Gait Energy Image (GEI) 42) ) FREQ GEI ) D x y N x D N X = (x 1,x 2,,x N ) N y y = (y 1,y 2,,y N ) N x y D = (X,y) D x y P(y x,d) x φ f (x) = φ(x) T w w x m x n k k(x m,x n ) = φ(x m ) T φ(x n ) [ ] = vexp 1 x m x n 2 2 r 2 (5) Θ = [v,r] x f P( f x,d) f 4 c 2011 Information Processing Society of Japan

5 σ 2 f 23) f = k T (K + S) 1 y (6) σ 2 f = k(x,x ) k T (K + S) 1 k (7) K i j k(x i,x j ) k i k(x,x i ) S σ 2 f (x) y 6 σ [ ] P(y f ) = 1 (y f )2 exp 2πσ 2σ 2 y P(y x,d) y σ 2 y = f (8) σ 2 = σ 2 f + σ 2 (9) x y 5.2 Θ σ X y P(y X) = p(y f) p(f X)d f = 1 (2π) N 2 K + S 1 2 exp [ 1 2 yt (K + S) 1 y (10) l = 1 2 yt (K + S) 1 y 1 2 log K + S N 2 logπ ] (10) l Θ σ , ( 1 ) (GEI) ( 2 ) (FREQ) ( 3 ) (GP) ( 4 ) + (GEI+GP) ( 5 ) + (FREQ+GP) (MAE:Mean Absolute Error) MAE t k y k ȳ k (11) MAE = 1 t y k ȳ k (11) k t t e l t e<l (12) CumScore(l) = t e<l t (%) (12) (GEI, FREQ, GP, GEI+GP, FREQ+GP) MAE 1 7 FREQ FREQ GEI 1 2 GEI GEI 5 c 2011 Information Processing Society of Japan

6 1 MAE Table 1 MAE of the age estimation for each gait feature. Gait Feature Male Female Total GEI FREQ GP GEI+GP FREQ+GP FREQ GEI GP GEI+GP FREQ+GP GEI FREQ 8 MAE GP MAE MAE 11 36) 7. (a) GEI (b) FREQ , (c) GP (d) GEI+GP 6 c 2011 Information Processing Society of Japan

7 10 Fig. 10 The impact of the gender and age group. 11 Fig. 11 Bias in age estimation. (e) FREQ+GP 7 Fig. 7 Distribution of the estimated ages for true ages. 9 8 Fig. 9 Correlation between gait period and age. Fig. 8 Cumulative Score. JST CREST 1) Zhang, D., Wang, Y. and Bhanu, B.: Ethnicity Classification Based on Gait Using Multiview Fusion, IEEE Computer Society and IEEE Biometrics Council Workshop on Biometrics 2010, San Francisco, CA, USA, pp.1 6 (2010). 2) Bobick, A. and Johnson, A.: Gait Recognition using Static Activity-specific Parameters, Proc. of IEEE Conf. on Computer Vision and Pattern Recognition, Vol.1, pp (2001). 3) Gavrila, D.M.: The Visual Analysis of Human Movement: A Survey, Computer Vision and Image Understandings, Vol.73, No.1, pp (1999). 4) Liu, Y., Collins, R. and Tsin, Y.: Gait Sequence Analysis using Frieze Patterns, Vol.2351, Springer Berlin / Heidelberg (2002). 5) Han, J. and Bhanu, B.: Individual recognition using gait energy image, Vol.28, pp (2006). 6) Vol.48, No.SIG1(CVIM17), pp (2007). 7) Huang, G. and Wang, Y.: Gender classification based on fusion of multi-view gait sequences, Proc. of the 8th Asian Conf. on Computer Vision, Vol.1, pp (2007). 8) Li, X., Maybank, S., Yan, S., Tao, D. and Xu, D.: Gait Components and Their Application to Gender Recognition, Trans. on Systems, Man, and Cybernetics, Part C, Vol.38, No.2, pp (2008). 9) Davis, J.: Visual Categorization of Children and Adult Walking Styles, Proc. of the Int. Conf. on Audio- and Video-based Biometric Person Authentication, pp (2001). 10) Begg, R.: Support Vector Machines for Automated Gait Classification, IEEE Trans. on 7 c 2011 Information Processing Society of Japan

8 Biomedical Engineering, Vol.52, No.5, pp (2005). 11) D-2 Vol.84, No.7, pp (2001). 12) Lu, J. and Tan, Y.-P.: Ordinary Preserving Manifold Analysis for Human Age Estimation, IEEE Computer Society and IEEE Biometrics Council Workshop on Biometrics 2010, San Francisco, CA, USA, pp.1 6 (2010). 13) Sarkar, S., Phillips, J., Liu, Z., Vega, I., Grother, P. and Bowyer, K.: The HumanID Gait Challenge Problem: Data Sets, Performance, and Analysis, Trans. of Pattern Analysis and Machine Intelligence, Vol.27, No.2, pp (2005). 14) Nixon, M., Carter, J., Shutler, J. and Grant, M.: Experimental plan for automatic gait recognition, Technical report, Southampton (2001). 15) Yu, S., Tan, D. and Tan, T.: A Framework for Evaluating the Effect of View Angle, Clothing and Carrying Condition on Gait Recognition, Proc. of the 18th Int. Conf. on Pattern Recognition, Vol.4, Hong Kong, China, pp (2006). 16) Little, J. and Boyd, J.: Recognizing People by Their Gait: The Shape of Motion, Videre: Journal of Computer Vision Research, Vol.1, No.2, pp.1 13 (1998). 17) Tanawongsuwan, R. and Bobick, A.: A Study of Human Gaits across Different Speeds, Technical report, Georgia Tech (2003). 18) Gross, R. and Shi, J.: The CMU Motion of Body (MoBo) Database, Technical report, CMT (2001). 19) Chalidabhongse, T., Kruger, V. and Chellappa, R.: The UMD database for human identification at a distance, Technical report, University of Meryland (2001). 20) : MIT AI Database, 21) : OU-ISIR Gait Database, 22) 13 (2010). 23) Carl EdwardRasmussen, C. K. I.W.: Gaussian Processes for Machine Learning, The MIT Press (2006). 24) Lanitis, A., Draganova, C. and Christodoulou, C.: Comparing different classifiers for automatic age estimation, TSMC (2004). 25) Lanitis, A., Taylor, C.J. and Cootes, T.F.: Toward automatic simulation of aging effects on face images, TPAMI (2002). 26) Yan, S., Zhou, X., Liu, M., M.Hasegawa-Johnson and Huang, T.S.: Regression from patchkernel, CVPR (2008). 27) Guo, G., Mu, G., Fu, Y. and Huang, T.S.: Human age estimation using bio-inspired features, CVPR (2009). 28) Fu, Y. and Huang, T.S.: Human age estimation with regression on discriminative aging manifold, TMM (2008). 29) Guo, G., Fu, Y., Dyer, C.R. and Huang, T.S.: Image-based human age estimation by manifold learning and locally adjusted robust regression, TIP (2008). 30) Xiao, B., Yang, X., Xu, Y. and Zha, H.: Learning distance metric for regression by semidefinite programming with application to human age estimation, ACMMM (2009). 31) Geng, X. and Smith-Miles, K.: Facial age estimation by multilinear subspace analysis, ICASSP (2009). 32) Geng, X., Smith-Miles, K. and Zhou, Z.H.: facial age estimation by nonlinear aging pattern subspace, ACMMM (2008). 33) Geng, X., Zhou, Z.H. and Smith-Miles, K.: Automatic age estimation based on facial aging patterns, TPAMI (2007). 34) Geng, X., Zhou, Z.H., Zhang, Y., Li, G. and Dai, H.: Learning from facial aging patterns for automatic age estimation, ACMMM (2006). 35) Zhang, Y. and Yeung, D.-Y.: Multi-Task Warped Gaussian Processes for Personalized Age Estimation, CVPR2010 (2010). 36) Snelson, E., Rasmussen, C. E. and Ghahramani, Z.: Warped Gaussian Processes, NIPS16 (2004). 37) D Vol.J92-D, No.8, pp (2009). 38) : Dive Into the Movie project, DIM/home e.html. 39) Okumura, M., Makihara, Y., Nakamura, S., Morishima, S. and Yagi, Y.: The Online Gait Measurement for the Audience-Participant Digital Entertainment, Invited Workshop on Vision Based Human Modeling and Synthesis in Motion and Expression, Xi an, China (2009). 40) Makihara, Y., Sagawa, R., Mukaigawa, Y., Echigo, T. and Yagi, Y.: Gait Recognition Using a View Transformation Model in the Frequency Domain, Proc. of the 9th European Conf. on Computer Vision, Vol.3, Graz, Austria, pp (2006). 41) Liu, Z. and Sarkar, S.: Simplest Representation Yet for Gait Recognition: Averaged Silhouette, Proc. of the 17th Int. Conf. on Pattern Recognition, Vol.1, pp (2004). 42) Han, J. and Bhanu, B.: Individual Recognition Using Gait Energy Image, Trans. on Pattern Analysis and Machine Intelligence, Vol.28, No.2, pp (2006). 8 c 2011 Information Processing Society of Japan

(a) (b) 2 2 (Bosch, IR Illuminator 850 nm, UFLED30-8BD) ( 7[m] 6[m]) 3 (PointGrey Research Inc.Grasshopper2 M/C) Hz (a) (b

(a) (b) 2 2 (Bosch, IR Illuminator 850 nm, UFLED30-8BD) ( 7[m] 6[m]) 3 (PointGrey Research Inc.Grasshopper2 M/C) Hz (a) (b (MIRU202) 202 8 AdrianStoica 89 0395 744 89 0395 744 Jet Propulsion Laboratory 4800 Oak Grove Drive, Pasadena, CA 909, USA E-mail: uchino@irvs.ait.kyushu-u.ac.jp, {yumi,kurazume}@ait.kyushu-u.ac.jp 2 nearest