(MIRU2008) HOG Histograms of Oriented Gradients (HOG)

Size: px

Start display at page:

Download "(MIRU2008) HOG Histograms of Oriented Gradients (HOG)"

くうしょうゆのもと
5 years ago
Views:

1 (MIRU2008) HOG katsu0920@me.cs.scitec.kobe-u.ac.jp, {takigu,ariki}@kobe-u.ac.jp Histograms of Oriented Gradients (HOG) HOG Shape Contexts HOG 5.5 Histograms of Oriented Gradients D Human Posture Estimation Using the HOG Features from Monocular Image Katsunori ONISHI, Tetsuya TAKIGUCHI, and Yasuo ARIKI Graduate School of Science and Technology, Kobe University, -, Rokkodai, Nada, Kobe, Hyogo Organization of Advanced Science and Technology, Kobe University, -, Rokkodai, Nada, Kobe, Hyogo katsu0920@me.cs.scitec.kobe-u.ac.jp, {takigu,ariki}@kobe-u.ac.jp Abstract the markers. This paper shows a method to estimate the D human posture from monocular image without using A D human body is expressed by a multi-joint model, and each of that joint angle describes a posture. This method estimates the posture with the Histograms of Oriented Gradients(HOG) features that can express the shape of the object in the input image obtained from one camera. In addition, the feature dimension of the background area is reduced by principal component analysis performed at every block of HOG. Joint angles in Human multi-joint model are estimated by linear regression analysis applied to its feature vector extracted from an input image. As a result of comparison experiment with the Shape Contexts features, the estimation error was reduced by about 5.5 degrees. Key words Posture estimation, D, Histograms of Oriented Gradients, Principal component analysis, Linear regression analysis, Monocular image. [] [2]

2 入力画像背景差分正規化輝度勾配算出セル毎に勾配方向と強度をヒストグラム化. ブロック毎に正規化 HOG [7] [0] [6] [8] [9] [] [2] HOG HOG HOG 2. HOG HOG 2. Histograms of Oriented Gradients gradient Histograms of Oriented Gradients (HOG) [] [4] Scale- Invariant Feature Transform (SIFT) [5] HOG SIFT SIFT SIFT (keypoint) HOG HOG 2.. HOG 2(a) (x, y) I(x, y) f x (x, y) = I(x +, y) I(x, y) f y (x, y) = I(x, y + ) I(x, y ) x, y x, y f x f y x y m(x, y) θ(x, y) () m(x, y) = f x (x, y) 2 + f y (x, y) 2 (2) θ(x, y) = tan (f y (x, y)/f x (x, y)) () θ(x, y) [ 80, 80 ] [0,80 ] { θ(x, y) + π, if θ(x, y) < 0 θ(x, y) = θ(x, y), otherwise (a) 2 セル : c w c h 画素 HOG (b) (4) 2(b) c w c h m(x, y) θ(x, y)

3 θ(x, y) c b m(x, y) c b 2.. b w b h c b d b = b w b h c b v (i, j), { < = i < = b w, < = j < = b h } h ij 4 2(a) HOG 特徴次元高 h ij = h ij v ϵ (ϵ = ) (5) h ij 4 低オーバーラップ & 正規化ブロック : c b方向 b w b h セル 2. ( ) y HOG HOG 5. [6] ( 6)

HOG 特徴 x n 枚画像 A 6 推定人体モデル特徴 HOG x R d y R m y = Ax +

:= arg min A y n Ax i y i 2 (7) i= m n Y (y y 2 y n )

y A AX = Y X T A T = Y T 4. Shape Contexts [8] 4.

i ) p i n p i h i (k) = {q p i : (q p i ) bin(k)} (9)

log-polar 7 60 shape context shape context 00 k-means

4 HOG 特徴 x n 枚画像 A 6 推定人体モデル特徴 HOG x R d y R m y = Ax + ε (6) A m d ε x y n {(y i, x i ) i = n}( HOG ) (A ) A := arg min A y n Ax i y i 2 (7) i= m n Y (y y 2 y n ) d n X (x x 2 x n ) A := arg min AX Y 2 (8) A HOG x A y A AX = Y X T A T = Y T 4. Shape Contexts [8] 4. [6] 7(a) 7(b) n P = {p,, p n }, p i R 2 p i p i q(q p i ) p i n p i h i (k) = {q p i : (q p i ) bin(k)} (9) p i shape context 7(c) log-polar p i (a) (b) (c) log-polar 7 60 shape context shape context 00 k-means shape context shape context (a)

(b) http://www.nada.kth.se/ hedvig/data.html (c) 2.

c w = 0, c h = 0, c b = 9, b w =, b h = 0 0 9 d b = b w b h c b

25 2 2 9 7 5 Shape Contexts HOG Block PCA 9 HOG HOG RMS error

5 (b) hedvig/data.html (c) 2. (a) (b) (c) HOG c w = 0, c h = 0, c b = 9, b w =, b h = d b = b w b h c b = HOG HOG 90% 8 HOG y ( ) 24 RMS error (degree) Shape Contexts HOG Block PCA 9 HOG HOG RMS error (degree) Shape Contexts HOG Block PCA 腕部脚部全身 HOG 0 5

( ) 5 0 Shape Contexts HOG Block PCA RMS error (degree) 25 20 5 0 5 0 直立手を上げる手を広げる歩く走る 2 2 9 5.

886-89, 2005. [4],, H.Bon-Woo,,, (MIRU2007), 492-497, 2007. [5] D.

Pattern Recognition, VOL.2, 882-888, 2004 [7] M.Lee, I.

Analysis and Machine Intelligence, VOL.28, No.6, 2006. [8] G.Mori and J.

6 ( ) 5 0 Shape Contexts HOG Block PCA RMS error (degree) 直立手を上げる手を広げる歩く走る HOG HOG [] -Computer Vision and Image Media, Vol.2006, No.5, pp.75-92, [2] Boosting (MIRU2007), 77-82, [] N.Dalal and B.Triggs Histograms of Oriented Gradients for Human Detection, IEEE Conputer Vision and Pattern Recognition, , [4],, H.Bon-Woo,,, (MIRU2007), , [5] D.Lowe, Distinctive Image Features from Scale-Invariant Keypoints, International Journal of Computer Vision, 60(2), 9-0, [6] A.Agarwal and B.Triggs, D Human Pose from Silhouettes by Relevance Vector Regression, IEEE Conference on Computer Vision and Pattern Recognition, VOL.2, , 2004 [7] M.Lee, I.Cohen, A Model-Based Approach for Estimating Human D Poses in Static Images, IEEE Transactions on Pattern Analysis and Machine Intelligence, VOL.28, No.6, [8] G.Mori and J.Malik, Estimating Human Body Configurations using Shape Context Matching, European Conference on Computer Vision, 50-80, [9] G.Mori and J.Malik, Recovering D Human Body Configurations using Shape Contexts, IEEE Transactions on Pattern Analysis and Machine Intelligence, VOL.28, No.7, , [0],,,, CG, Technical report of IEICE. PRMU, VOL.04, No.57, 79-84, [], B.Stenger, Tree-Based Filtering, (MIRU2006), 6-69, [2],,,, (MIRU2007), , 2007.

Duplicate Near Duplicate Intact Partial Copy Original Image Near Partial Copy Near Partial Copy with a background (a) (b) 2 1 [6] SIFT SIFT SIF

Duplicate Near Duplicate Intact Partial Copy Original Image Near Partial Copy Near Partial Copy with a background (a) (b) 2 1 [6] SIFT SIFT SIF Partial Copy Detection of Line Drawings from a Large-Scale Database Weihan Sun, Koichi Kise Graduate School of Engineering, Osaka Prefecture University E-mail: sunweihan@m.cs.osakafu-u.ac.jp, kise@cs.osakafu-u.ac.jp