Learning assessment method and device using a virtual tutor

• US 8,021,160 B2
• Filed: 10/06/2006
• Issued: 09/20/2011
• Est. Priority Date: 07/22/2006
• Status: Expired due to Fees

• Current Assignee: Industrial Technology Research Institute
• Original Assignee: Industrial Technology Research Institute
• Inventors: Chang, Yao-Jen
• Primary Examiner(s): Thai; Xuan
• Assistant Examiner(s): Yip; Jack
• Application Number:
  US11/539,178
• Publication Number:
  US 20080020363A1
• Time in Patent Office: 1,810 Days
• Field of Search: 434/308, 434/262, 434/350, 434/247, 434/252, 434/258, 345/429, 345/473, 345/619, 345/629, 345/630, 345/646, 382/107, 382/108, 473/222, 473/266
• US Class Current: 434/262
• CPC Class Codes: G09B 5/00 Electrically-operated educational appliances working with questions and answers G09B7/00; simulators G09B9/00; advertising or displaying in general G09F

Abstract

Disclosed is a learning assessment method and device using a virtual tutor. The device comprises at least one action acquisition module, a virtual tutor synthesis module, and a learning assessment module. The method captures and analyzes a first and a second action-feature for a first and a second targets respectively, and constructs an intrinsic model of the second target based on a reference data of the second target. A virtual tutor is synthesized by applying the first action-feature to the intrinsic model such that the virtual tutor exhibits the intrinsic characteristics of the second target but performs a synthesized action-feature similar to the first action-feature. The method then assesses the difference between the synthesized action-feature and the second action-feature.

Claims

1. A method using a synthesized virtual tutor in a learning assessment device for providing performance assessment of a subject performing an action imitation task, comprising the steps of:

using at least a video action acquisition and analysis module to acquire a first action video of a first target for analyzing a first action-feature of said first target performing a first action;

using said at least a video action acquisition and analysis module to acquire a second action video of a second target for analyzing a second action-feature of said second target performing a second action imitating said first action;

establishing an intrinsic model of said second target by using reference data of said second target, said intrinsic model being constructed with a multidimensional morphable model by using image textures and motion flows of said second target;

generating a synthetic video with a virtual tutor having said image textures and motion flows of said second target but exhibiting a synthesized action-feature with behavior similar to said first action-feature based on a behavior model of said first target, said behavior model being constructed by using a set of reference data of said first target with a model transfer process and a model adaptation process, said model transfer process being composed of image texture matching and motion flow matching procedures for finding a set of prototype images for said first target with a matching-by-synthesis approach; and

assessing image texture and motion flow differences between said second action-feature and said synthesized action-feature through a learning assessment module;

wherein said image textures and motion flows of said second target are trained from a set of prototype images selected from said reference data of said second target with said multidimensional morphable model, and said synthetic video with said virtual tutor is generated by compositing said image textures and motion flows of said second target using said intrinsic model of said second target to form each frame of said synthetic video by warping and combining said prototype images with parameters generated according to said behavior model of said first target; and

wherein said model transfer process adopts said matching-by-synthesis approach further comprising the steps of;

establishing a set of key correspondences between a reference image of said first target and a reference image of said second target to derive dense point correspondences as an image warping function between said first and second targets;

generating a set of synthetic prototype motion flows for said first target by warping motion flows of said prototype images of said second target with said image warping function, and searching for an initial set of prototype images from said reference data of said first target whose motion flows are best matched to said synthetic prototype motion flows;

generating a set of synthetic prototype image textures by warping and combining said initial set of prototype images using linear dependency between said prototype images of said second target; and

iteratively searching for an updated set of prototype images from said reference data of said first target whose image textures and motion flows are best matched to said synthetic prototype image textures and motion flows, and taking said updated set prototype images as said set of prototype images for said first target.

2. The method as claimed in claim 1, wherein said first target is a learner and said second target is a target being imitated.

3. The method as claimed in claim 1, wherein said first target is a target being imitated and said second target is a learner.

4. The method as claimed in claim 1, wherein said performance assessment includes the comparison of said second action-feature and said synthesized action-feature of said virtual tutor based on differences in image textures and motion flows, and the provision of a corresponding assessment result or correction guidelines.

5. The method as claimed as in claim 1, wherein said synthetic video with said virtual tutor is generated with the steps of:

providing said intrinsic model and a behavior model of said second target trained with a sufficient amount of said reference data of said second target;

establishing an intrinsic model of said first target and said behavior model of said first target according to said set of reference data of said first target by using said model transfer process and said model adaptation process with said intrinsic model of said second target and said behavior model of said second target;

using said intrinsic model of said second target to form the intrinsic model of said virtual tutor;

using said behavior model of said first target to form the behavior model of said virtual tutor; and

generating image texture and motion flow parameters based on said behavior model of said virtual tutor and synthesizing images with said intrinsic model of said virtual tutor to form a synthetic video.

6. The method as claimed as in claim 5, wherein said model adaptation process adopts a gradient descent linear regression method to build an adapted behavior model for said first target by adapting an original behavior model of said second target and calculating linear transformation relation between said adapted behavior model of said first target and said original behavior model of said second target to minimize differences in image textures and motion flows generated by said adapted behavior model with respect to said reference data of said first target, said linear transformation being adapted iteratively by using a gradient descent approach that updates said linear transformation with partial derivatives of said differences with respect to parameters of said linear transformation.

7. The method as claimed in claim 6, wherein said gradient descent linear regression method includes the following steps to calculate linear transformation relation between said adapted behavior model of said first target and said original behavior model of said second target:

dividing a plurality of elements in said original behavior model into a plurality of regression groups, each said group g sharing a common linear transformation matrix R_g to transform a mean vector μ

_p of any element p of the group g to μ

_p^adapt=R_gξ

_p, where ε

_p=[1 μ

_p]^T is an extended mean vector; and

adapting said linear transformation matrix R_g to minimize synthesis error in terms of differences in image textures and motion flows generated by said adapted behavior model with respect to said reference data of said first target for elements in said group g.

8. The method as claimed in claim 7, wherein said dividing said plurality of elements into said plurality of groups is based on the characteristics of said elements in a multidimensional morphable model parameter space.

9. A device using a synthesized virtual tutor for providing performance assessment of a subject performing an action imitation task, comprising:

at least a video action acquisition and analysis module for acquiring a first action video of a first target and analyzing a first action-feature of said first target performing a first action, and acquiring a second action video of a second target and analyzing a second action-feature of said second target performing a second action imitating said first action;

a virtual tutor synthesis module for generating a synthetic video with a virtual tutor having image textures and motion flows of said second target but exhibiting a synthesized action-feature with behavior similar to said first action-feature based on a behavior model of said first target, said behavior model being constructed by using a set of reference data of said first target with a model transfer process and a model adaptation process, said model transfer process being composed of image texture matching and motion flow matching procedures for finding a set of prototype images for said first target with a matching-by-synthesis approach; and

a learning assessment module for assessing image texture and motion flow differences between said second action-feature and said synthesized action-feature;

wherein said image textures and motion flows of said second target are trained from a set of prototype images selected from said reference data of said second target with said multidimensional morphable model, and said synthetic video with said virtual tutor is generated by compositing said image textures and motion flows of said second target using said intrinsic model of said second target to form each frame of said synthetic video by warping and linearly combining said prototype images with parameters generated according to said behavior model of said first target; and

wherein said model transfer module adopts said matching-by-synthesis approach through comprising the steps of;

establishing a set of key correspondences between a reference image of said first target and a reference image of said second target to derive dense point correspondences as an image warping function between said first and second targets;

generating a set of synthetic prototype motion flows for said first target by warping motion flows of said prototype images of said second target with said image warping function, and searching for an initial set of prototype images from said reference data of said first target whose motion flows are best matched to said synthetic prototype motion flows;

generating a set of synthetic prototype image textures by warping and combining said initial set of prototype images using linear dependency between said prototype images of said second target; and

iteratively searching for an updated set of prototype images from said reference data of said first target whose image textures and motion flows are best matched to said synthetic prototype image textures and motion flows, and taking said updated set prototype images as said set of prototype images for said first target.

10. The device as claimed in claim 9, wherein said learning assessment module further includes a comparison module and a correction guidance module, said comparison module generates related information on differences in image textures and motion flows between said second action video and said synthetic video with said virtual tutor, and said correction guidance module provides guidelines in accordance with said related information.

11. The device as claimed in claim 9, wherein said model adaptation module adopts a gradient descent linear regression method to build an adapted behavior model for said first target by adapting an original behavior model of said second target and calculate linear transformation relation between said adapted behavior model and said original behavior model of said second target to minimize differences in image textures and motion flows generated by said adapted behavior model with respect to said reference data of said first target, said linear transformation being adapted iteratively by using a gradient descent approach that updates said linear transformation with partial derivatives of said differences with respect to parameters of said linear transformation.