Apparatus for Recognizing Three-Dimensional Motion Using Linear Discriminant Analysis

• US 20080285807A1
• Filed: 12/05/2006
• Published: 11/20/2008
• Est. Priority Date: 12/08/2005
• Status: Abandoned Application

• Current Assignee: Electronics and Telecommunications Research Institute
• Original Assignee: Electronics and Telecommunications Research Institute
• Inventors: Kim, Hyun-Bin, Lee, In-Ho, Park, Chang-Joon, Lee, Jae-Ho
• Application Number:
  US12/091,625
• Publication Number:
  US 20080285807A1
• Time in Patent Office: Days
• US Class Current: 382/107
• CPC Class Codes: G06T 7/20 Analysis of motion motion estimation for coding, decoding, compressing or decompressing digital video signals H04N19/43, H04N19/51

Abstract

Provided is an apparatus and method for recognizing a three-dimensional (3D) motion using Linear Discriminant Analysis (LDA). The apparatus includes: a 3D motion capturing means for creating motion data for every motion; a motion recognition learning means for analyzing the created motion data, creating a linear discrimination feature component for discriminating corresponding motion data, extracting/storing a reference motion feature, and recognizing each of the extracted/stored reference motion features as a corresponding motion; and a motion recognition operating means for extracting a motion feature from motion data, searching a reference motion feature corresponding to the extracted input motion feature, and recognizing a motion corresponding to the searched reference motion feature as a 3D motion.

Claims

1. An apparatus for recognizing a three-dimensional (3D) motion using Linear Discriminant Analysis (LDA), comprising:

a 3D motion capturing means for creating motion data for every motion by using a marker-free motion capturing process for a motion of an actor;

a motion recognition learning means for analyzing the created motion data on multiple types of motions using the LDA, creating a linear discrimination feature component for discriminating corresponding motion data, extracting/storing a reference motion feature on each type of motions based on the created linear discrimination feature component, and recognizing each of the extracted/stored reference motion features as a corresponding motion; and

a motion recognition operating means for extracting a motion feature based on the created linear discrimination feature component from motion data on an input motion to be the created 3D recognition object, searching a reference motion feature corresponding to the extracted input motion feature among the stored reference motion features, and recognizing a motion corresponding to the searched reference motion feature as a 3D motion on the input motion.

2. The apparatus as recited in claim 1, further comprising:

a motion command transmitting means for transmitting the recognized 3D motion to a motion command of a character;

a key input creating means for creating a key input value corresponding to the transmitted motion command transmitted from the motion command transmitting means; and

a 3D virtual motion controlling means for controlling a 3D virtual motion of the character according to the created key input value.

3. The apparatus as recited in claim 1, wherein the motion recognition learning means includes:

a motion data analyzing means for analyzing the created motion data on multiple types of motions using the LDA;

a feature component creating means for creating a linear discrimination feature component for discriminating the analyzed motion data obtained in the motion data analyzing means; and

a motion feature learning means for extracting/storing a reference motion feature on each type of motions based on the created linear discrimination feature component and recognizing each of the extracted/stored reference motion features as a corresponding motion.

4. The apparatus as recited in claim 3, wherein the feature component creating means creates a linear discrimination feature component W_opt according to the LDA method using Equations 1 and 2 below;

$\begin{array}{cc}{W}_{\mathrm{opt}}=\underset{w}{\text{arg}<br><br>\max }<br><br>\frac{<br><br>W^T<br><br>S_B<br><br>W<br><br>}{<br><br>W^T<br><br>S_W<br><br>W<br><br>}=\left[\begin{array}{cccc}{w}_1& w_2& \mathrm{\dots <br><br>}& w_m\end{array}\right]& \mathrm{Eq}.<br><br>1\end{array}$ $\begin{array}{cc}{S}_B=\underset{i=1}{\overset{c}{\sum <br><br>}}<br><br>N_i<br><br>\left({\mathrm{\mu <br><br>}}_i-\stackrel{\_}{\mathrm{\mu <br><br>}}\right)<br><br>{\left({\mathrm{\mu <br><br>}}_i-\stackrel{\_}{\mathrm{\mu <br><br>}}\right)}^T<br><br><br><br>S_W=\underset{i=1}{\overset{c}{\sum <br><br>}}<br><br>\underset{x_k\in <br><br>X_i}{\sum <br><br>}<br><br>\left(x_k-{\mathrm{\mu <br><br>}}_i\right)<br><br>{\left(x_k-{\mathrm{\mu <br><br>}}_i\right)}^T& \mathrm{Eq}.<br><br>2\end{array}$

where, S_B is a between-class scatter matrix;

S_W is a within-class scatter matrix;

X_i is a class of each motion;

μ

_i is mean motion data of the motion class X_i;

c is the total number of classes; and

N_i is the number of motion data included in each class.

5. The apparatus as recited in claim 3, wherein the motion feature learning means recognizes the motion on the extracted/stored reference feature as a single motion, which is a still motion, or a combination motion, which is a motion combining determination results of the continued motions.

6. The apparatus as recited in claim 1, wherein the motion recognition operating means includes:

a motion feature extracting means for extracting a motion feature based on the linear discrimination feature component created in the motion feature extracting means from the motion data on an input motion, which is an object of the 3D recognition object which is generated in the 3D motion capturing means; and

a motion recognizing means for searching a reference motion feature at the minimum statistical distance from the extracted input motion feature extracted in the motion feature extracting means among the stored reference motion features and recognizing a motion corresponding to the searched reference motion feature as the 3D motion on the input motion.

7. The apparatus as recited in claim 6, wherein the statistical distance between the input motion feature and the reference motion feature is measured in the motion recognizing means, is according to a Mahalanobis distance f(g_s) measuring method using Equation 3 below;

f(g_s)=(g_s−

g)^TS_g<sup>−

1</sup>(g_s−

g) 

 

Eq. 3

where g_s is an inputted sample;

g is a mean of each group; and

S_g is a covariance of each group.

8. A method for recognizing a three-dimensional (3D) motion using Linear Discriminant Analysis (LDA), comprising the steps of:

a) creating motion data for every motion by performing a marker-free motion capturing process on a motion of an actor;

b) extracting a motion feature based on a pre-stored linear discrimination feature component from motion data on an input motion, which is an object of 3D recognition created in the step a);

c) searching a reference motion feature, which has the minimum statistical distance from the extracted input motion feature, among the pre-stored reference motion features; and

d) recognizing a motion corresponding to the searched reference motion feature as a 3D motion corresponding to the input motion.

9. The method as recited in claim 8, further comprising the steps of:

e) creating and storing the linear discrimination feature component for discriminating the motion data by analyzing the created motion data on multiple motions using the LDA;

f) extracting and storing a reference motion feature on each type of motions based on the created linear discrimination feature component generated in the step e); and

g) recognizing each of extracted/stored reference motion features as a corresponding motion.

10. The method as recited in claim 8, further comprising the steps of:

h) transmitting the 3D motion recognized in the step

d) to a motion command of a character;

i) creating a key input value corresponding to the transmitted motion command; and

j) controlling a 3D virtual motion of the character according to the created key input value.

11. The method as recited in claim 10, wherein in the step g), the 3D motion feature is recognized as a single motion, which is a still motion, or a combination motion, which is a motion combining determination results of the continued motions.

12. The method as recited in claim 10, wherein in the statistical distance measuring procedure of the step c), the statistical distance between the input motion feature and the reference motion feature is measured according to a Mahalanobis distance f(g_s) measuring method using Equation 4 below;

f(g_s)=(g_s−

g)^TS_g<sup>−

1</sup>(g_s−

g) 

 

Eq. 4

Where g_s is an inputted sample;

g is a mean of each group; and

S_g is a covariance of each group.