(Robot Vision) Vision ( (computer) Machine VisionComputer Vision ( ) ( ) ( ) ( ) ( ) 1
DTV 2
DTV D 3
( ( ( ( ( DTV D 4
() 5
A B C D E F G H I A B C D E F G H I I = A + D + G - C - F - I J = A + B + C - G - H - I = I + J = arctan ( I J 6
() Roberts (1964) 7
Laplacian mask 16 8
Edge Background subtraction Fitting Bezer Curve 9
10
2 H x, y = (det( A ) k ( tr( A )) 2 R x Rx R y A = w( u, v) 2 u v Rx R y R y W : u,v : k : 0.04 Rx Ry Harrisoperator Harris 1. (x, y) 2. 3., 11
12 Lucas Kanade f(x,y): g(x,y):lk u,v:( ( ) ( ) + + + + = u,v,, SAD v y u x g v y u x f e x b a x' + = f e y x d c b a y x ' (, ): (,) (a,b,c,d): (e,f):
Docomo P902isN902isSH902is http://www.nttdocomo.co.jp/product/902new/security.html 13
DoCoMo FOMA P902is DoCoMo HP: http://www.nttdocomo.co.jp/product/foma/902i/p902is/topics_03.html 14
15
16
17
HarrWavelet 18
-1,0,1 101 HarrWavelet 19
HarrWavelet 20
21
22
Human vision vs Computer vision 以前の迷信 デジタルコンピュータは画面の端から順に見 ていく 人は一目で重要な部分だけに目が行き そ こだけを詳しく見る しかし 大量の学習データを必要とする 学 デ を す 3DCG生成プロセスに利用される顔器官輪郭検出 顔器官輪郭検出 標準ワイヤーフレーム 3Dレンジデータ 3Dレンジデータと合わせて 3Dレンジデ タと合わせて 2D画像 画像 OKAO Vision 顔器官輪郭検出 3Dモデル生成 画像に整合 表情を合成 標準ワイヤーフレームを対応点が 輪郭点に合致するように変形 あらかじめ決められた各点の移動 量パターンに従い3Dモデルを変形 2005 MITSIMITSI-TOSHIBA PAVILION / dentsu / dentsu tec 23
OKAO Vision OKAO Vision 2005 MITSI-TOSHIBA TOSHIBA PAVILION / dentsu / dentsu tec 24
mapping robot working robot keypoint 25
Scale Invariant Feature Transform example : n=2 26
27
Detection Hessian-based interest point localization L xx (x,y,σ) is the Laplacian of Gaussian of the image It is the convolution of the Gaussian second order derivative with the image Lindeberg showed Gaussian function is optimal for scale-space analysis This paper argues that Gaussian is overrated since the property that no new structures can appear while going to lower resolution is not proven in 2D case 28
Detection Approximated second order derivatives with box filters (mean/average filter) 57 Detection Scale analysis with constant image size 9 x 9, 15 x 15, 21 x 21, 27 x 27 39 x 39, 51 x 51 1 st octave 2 nd octave 29
Detection Non-maximum suppression and interpolation Blob-like feature detector 59 Description Orientation Assignment Circular neighborhood of radius 6s around dthe interest tpoint (s = the scale at which the point was detected) x response y response Side length = 4s Cost 6 operation to 60 compute the response 30
Description Dominant orientation The Haar wavelet responses are represented as vectors Sum all responses within a sliding orientation window covering an angle of 60 degree The two summed response yield a new vector The longest vector is the dominant orientation Second longest is ignored 61 Description Splittheinterestregionupinto4x4squaresub-regions the interest region into 4 square sub regions with 5 x 5 regularly spaced sample points inside Calculate Haar wavelet response d x and d y Weight the response with a Gaussian kernel centered at the interest point Sum the response over each sub-region for d x and d y separately tl feature vector of length th32 In order to bring in information about the polarity of the intensity changes, extract the sum of absolute value of the responses feature vector of length 64 Normalize the vector into unit length 62 31
Description 63 32