Method of learning a parameter to estimate posture of an articulated object and method of estimating posture of an articulated object

• US 9,460,368 B2
• Filed: 10/24/2014
• Issued: 10/04/2016
• Est. Priority Date: 01/23/2014
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: Han, Seung Ju, Han, Jae Joon
• Primary Examiner(s): Kassa, Yosef
• Application Number:
  US14/523,101
• Publication Number:
  US 20150206032A1
• Time in Patent Office: 711 Days
• Field of Search: 382/154, 382/156, 382/157, 382/160, 382/224, 382/278, 358/537, 358/452
• US Class Current: 1/1
• CPC Class Codes: G06V 30/1983 Syntactic or structural pattern recognition, e.g. symbolic string recognition
• CPC Class Codes: G06V 40/107 Static hand or arm

Abstract

A method of learning a parameter to estimate a posture of an articulated object, and a method of estimating the posture of the articulated object are provided. A parameter used to estimate a posture of an articulated object may be iteratively learned based on a depth feature corresponding to an iteration count, and the posture of the articulated object may be estimated based on the learned parameter.

Claims

1. A learning method, comprising:

iteratively learning a parameter based on an estimation position and a depth feature corresponding to the estimation position, the parameter being used to estimate a posture of an articulated object in a training image and the estimation position corresponding to an iteration count, wherein

the iteratively learning includes computing the estimation position based on an estimation position corresponding to a previous iteration count, a depth feature corresponding to the previous iteration count, and a parameter corresponding to the previous iteration count.

2. The learning method of claim 1, wherein the training image includes position information of parts of the articulated object.

3. The learning method of claim 2, wherein the iteratively learning comprises:

computing the depth feature based on the estimation position; and

computing the parameter based on the position information, the estimation position, and the depth feature.

4. The learning method of claim 2, wherein

the iteratively learning includes generating a plurality of offset pairs within an area around the estimation position, and

the computing the depth feature computes the depth feature based on the offset pairs.

5. The learning method of claim 4, wherein as the iteration count increases, a size of the area decreases.

6. The learning method of claim 4, further comprising:

first storing the parameter learned in each iteration count.

7. The learning method of claim 6, further comprising:

second storing a plurality of offset pairs generated within an area around the estimation position, and

computing the depth feature based on the offset pairs.

8. The learning method of claim 1, wherein the iteratively learning comprises:

computing the estimation position X based on

x_k+1=x_k+R_k·

φ

_k(x_k)+b_k

wherein k denotes an iteration count, x_k denotes the estimation position corresponding to a k-th iteration, R_k and b_k denote parameters corresponding to the k-th iteration, φ

_k( ) denotes an operator to compute the depth feature corresponding to the k-th iteration, and x_k+1 denotes the estimation position corresponding to a (k+1)-th iteration.

9. The learning method of claim 1, wherein the iteratively learning comprises:

computing the depth feature based on

10. The learning method of claim 1, wherein the iteratively learning comprises:

computing the parameter based on

11. An estimation method, comprising:

iteratively estimating position information of parts of an articulated object included in an input image based on an estimation position corresponding to an iteration count, a depth feature corresponding to the estimation position, and a parameter corresponding to the iteration count, wherein

the iteratively estimating includes computing the estimation position based on an estimation position corresponding to a previous iteration count, a depth feature corresponding to the previous iteration count, and a parameter corresponding to the previous iteration count.

12. The estimation method of claim 11, further comprising:

estimating a posture of the articulated object based on the position information.

13. The estimation method of claim 11, wherein the iteratively estimating comprises:

computing the depth feature based on offset pairs, the offset pairs indicating offset positions that are based on the estimation position.

14. The estimation method of claim 13, wherein the offset positions are distributed closer to the estimation position, as the iteration count increases.

15. The estimation method of claim 11, further comprising:

first receiving parameters associated with each iteration count that are determined prior to the associated iteration count.

16. The estimation method of claim 15, further comprising:

second receiving offset pairs, the offset pairs indicating offset positions that are based on the estimation position.

17. The estimation method of claim 11, wherein the iteratively estimating comprises:

computing the estimation position x based on

x_k+1=x_k+R_k·

φ

_k(x_k)+b_k

wherein k denotes the iteration count, x_k denotes the estimation position corresponding to a k-th iteration, R_k and b_k denote parameters corresponding to the k-th iteration, φ

_k( ) denotes an operator to compute the depth feature corresponding to the k-th iteration, and x_k+1 denotes the estimation position corresponding to a (k+1)-th iteration.

18. The estimation method of claim 11, wherein the iteratively estimating comprises:

computing the depth feature based on