PROXY TRAINING DATA FOR HUMAN BODY TRACKING

• US 20110228976A1
• Filed: 03/19/2010
• Published: 09/22/2011
• Est. Priority Date: 03/19/2010
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Fitzgibbon, Andrew, Finnochio, Mark, Moore, Richard, Shotton, Jamie, Cook, Mat
• Granted Patent: US 8,213,680 B2
• Time in Patent Office: Days
• US Class Current: 382/103
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 10/754 involving a deformation of the sample pattern or of the reference pattern; Elastic matching
• CPC Class Codes: G06V 10/774 Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06V 40/20 Movements or behaviour, e.g. gesture recognition recognition of facial expressions G06V40/16

Abstract

Synthesized body images are generated for a machine learning algorithm of a body joint tracking system. Frames from motion capture sequences are retargeted to several different body types, to leverage the motion capture sequences. To avoid providing redundant or similar frames to the machine learning algorithm, and to provide a compact yet highly variegated set of images, dissimilar frames can be identified using a similarity metric. The similarity metric is used to locate frames which are sufficiently distinct, according to a threshold distance. For realism, noise is added to the depth images based on noise sources which a real world depth camera would often experience. Other random variations can be introduced as well. For example, a degree of randomness can be added to retargeting. For each frame, the depth image and a corresponding classification image, with labeled body parts, are provided. 3-D scene elements can also be provided.

Claims

1. A processor-implemented method for generating proxy training data for human body tracking, comprising the processor-implemented steps of:

accessing at least one motion capture sequence which identifies poses of an actor'"'"'s body during a time period in which the actor performs a movement;

performing retargeting to a plurality of body types and dissimilar pose selection, based on the at least one motion capture sequence, to provide a plurality of dissimilar poses; and

rendering each of the dissimilar poses according to a 3-D body model for a respective body type of the plurality of body types, to provide a respective depth image of the dissimilar pose, and to provide a respective classification image which identifies body parts of the dissimilar pose, where a plurality of 3-D body models are used, one for each body type, and the respective depth image and the respective classification image comprise pixel data which is usable by a machine learning algorithm for human body tracking.

2. The processor-implemented method of claim 1, wherein:

the retargeting retargets the at least one motion capture sequence to the plurality of body types, to provide a plurality of retargeted motion capture sequences, each retargeted motion capture sequence identifies poses for a respective body type of the plurality of body types; and

the dissimilar pose selection selects the dissimilar poses from the plurality of retargeted motion capture sequences, such that the dissimilar poses are a subset of poses in the plurality of retargeted motion capture sequences.

3. The processor-implemented method of claim 1, wherein:

the dissimilar pose selection selects the dissimilar poses from the at least one motion capture sequence, such that the dissimilar poses are a subset of poses in the at least one motion capture sequence; and

the retargeting retargets the dissimilar poses to the plurality of body types, to provide a plurality of retargeted motion capture sequences, each retargeted motion capture sequence identifies poses for a respective body type of the plurality of body types.

4. The processor-implemented method of claim 1, further comprising the process-implemented step of:

introducing a variation comprising at least one of a varied bone length and a varied joint angle into at least one of the dissimilar poses.

5. The processor-implemented method of claim 1, wherein:

the 3-D body model is rendered based on at least one of joint positions and joint angles of a corresponding skeleton.

6. The processor-implemented method of claim 1, wherein:

the rendering adds 3-D scene elements to at least one of the dissimilar poses; and

at least one of the 3-D scene elements is tailored to a corresponding body pose of the dissimilar poses.

7. The processor-implemented method of claim 6, further comprising the process-implemented step of:

extending the poses of the actor'"'"'s body to at least one additional pose by combining joint angles from two or more motion capture frames of the at least one motion capture sequence.

8. The processor-implemented method of claim 1, wherein:

the plurality of body types include variations in gender, age, body type, hair and clothing.

9. The processor-implemented method of claim 1, wherein:

the rendering varies at least one of the dissimilar poses by varying at least one of a virtual camera height, a virtual camera angle and a virtual camera field of view.

10. The processor-implemented method of claim 1, wherein:

the rendering varies at least one of the dissimilar poses by varying a body position for the at least one of the dissimilar poses.

11. The processor-implemented method of claim 1, wherein:

the rendering adds 3-D scene elements to at least one of the dissimilar poses; and

the rendering combines two or more of the 3-D body models in different poses but with a common virtual camera position and common 3-D scene elements.

12. Tangible computer readable storage having computer readable software embodied thereon for programming at least one processor to perform a method for generating proxy training data for human body tracking, the method comprising:

accessing at least one motion capture sequence which identifies poses of an actor'"'"'s body during a time period in which the actor performs a movement;

performing retargeting to a plurality of body types and dissimilar pose selection, based on the at least one motion capture sequence, to provide a plurality of dissimilar poses; and

rendering each of the dissimilar poses according to a 3-D body model for a respective body type of the plurality of body types, to provide a respective depth image of the dissimilar pose, and to provide a respective classification image which identifies body parts of the dissimilar pose, where a plurality of 3-D body models are used, one for each body type, the respective depth image and the respective classification image comprise pixel data which is usable by a machine learning algorithm for human body tracking, and the rendering adds 3-D scene elements to at least one of the dissimilar poses.

13. The tangible computer readable storage of claim 12, wherein:

the 3-D scene elements include a ground place, background wall, ceiling and furniture.

14. The tangible computer readable storage of claim 12, wherein:

at least one of the 3-D scene elements is tailored to a corresponding body pose of the plurality of dissimilar poses.

15. A processor-implemented method for generating proxy training data for human body tracking, comprising the processor-implemented steps of:

accessing at least one motion capture sequence which identifies poses of an actor'"'"'s body during a time period in which the actor performs a movement;

performing dissimilar pose selection and retargeting to a plurality of body types, based on the at least one motion capture sequence, to provide a plurality of dissimilar poses; and

rendering each of the dissimilar poses according to a 3-D body model for a respective body type of the plurality of body types, to provide a respective depth image of the dissimilar pose, and to provide a respective classification image which identifies body parts of the dissimilar pose, where a plurality of 3-D body models are used, one for each body type, the respective depth image and the respective classification image comprise pixel data which is usable by a machine learning algorithm for human body tracking, and the rendering adds noise to at least one of the dissimilar poses.

16. The processor-implemented method of claim 15, wherein:

the noise is added by adding head hair to the at least one of the dissimilar poses.

17. The processor-implemented method of claim 15, wherein:

the noise is added by quantizing a depth resolution of the at least one of the dissimilar poses, with a precision dependent on depth.

18. The processor-implemented method of claim 15, wherein:

the noise is added by identifying at least one edge in the at least one of the dissimilar poses, and replacing a depth of at least one pixel at the at least one edge with a depth of a neighboring pixel further from the at least one edge.

19. The processor-implemented method of claim 15, wherein:

the noise is added by identifying at least one thin structure which is thinner than a specified number of pixels in the at least one of the dissimilar poses, and replacing pixels of the at least one thin structure with background pixels.

20. The processor-implemented method of claim 15, wherein:

the noise is added by identifying first and second edges in the at least one of the dissimilar poses having a depth discontinuity greater than a threshold, and replacing pixels between the first and second edges with background pixels.