Recognition of hand poses by classification using discrete values

Patent metadata

Field	Value
Publication number	US-9734435-B2
Application number	US-201514985741-A
Country	US
Kind code	B2
Filing date	Dec 31, 2015
Priority date	Dec 31, 2015
Publication date	Aug 15, 2017
Grant date	Aug 15, 2017

Abstract

Computer implemented method for computing a feature dataset classifying a pose of a human hand, comprising:
(a) Selecting a global orientation category (GOC) defining a spatial orientation of a human hand in a 3D space by applying GOC classifying functions on a received image segment depicting the hand. (b) Identifying in-plane rotation by applying in-plane rotation classifying functions on the image segment, the in-plane rotation classifying functions are selected according to said GOC. (c) Aligning the image segment in a 2D plane according to the in-plane rotation. (d) Applying hand pose features classifying functions on the aligned image segment.

Each one of the feature classifying functions outputs a current discrete pose value of an associated hand feature.
(e) Outputting a features dataset defining a current discrete pose value for each of the hand pose features for classifying current hand pose of the hand.

First claim

What is claimed is:

1. A computer implemented method of computing a features dataset classifying a pose of a hand of a human, comprising:
selecting, using a computerized processor, a global orientation category (GOC) defining a spatial orientation of a hand of a human in a three dimensional (3D) space by applying a plurality of GOC classifying functions on a received image segment depicting said hand;
identifying an in-plane rotation by applying a plurality of in-plane rotation classifying functions on said image segment, said plurality of in-plane rotation classifying functions being selected according to said GOC;
aligning said image segment in a 2 dimensional (2D) plane according to said identified in-plane rotation;
applying a plurality of hand feature classifying functions on said aligned image segment, each one of said plurality of feature classifying functions outputting a current discrete pose value of an associated one of a plurality of hand pose features; and
outputting a features dataset defining a respective said current discrete pose value for each of said plurality of hand pose features for classifying a current hand pose of said hand;

wherein said GOC is identified after estimating a center of hand by applying a plurality of center of hand classifying functions on said image segment.

2. The computer implemented method of claim 1 , further comprising said center of hand is derived from a center of mass of said hand, said center of mass is identified by analyzing a depth image data available from an image depicting said hand, said depth image data maps a depth of said hand in said 3D space.

3. The computer implemented method of claim 2 , further comprising, said image is manipulated to remove at least one non-relevant image portions to produce an image segment.

4. The computer implemented method of claim 1 , wherein said plurality of center of hand classifying functions are trained statistical classifiers used for regression analysis.

5. The computer implemented method of claim 1 , wherein said plurality of GOC classifying functions are trained statistical classifiers.

6. A computer implemented method of computing a features dataset classifying a pose of a hand of a human, comprising:
selecting, using a computerized processor, a global orientation category (GOC) defining a spatial orientation of a hand of a human in a three dimensional (3D) space by applying a plurality of GOC classifying functions on a received image segment depicting said hand;
identifying an in-plane rotation by applying a plurality of in-plane rotation classifying functions on said image segment, said plurality of in-plane rotation classifying functions being selected according to said GOC;
aligning said image segment in a 2 dimensional (2D) plane according to said identified in-plane rotation;
applying a plurality of hand feature classifying functions on said aligned image segment, each one of said plurality of feature classifying functions outputting a current discrete pose value of an associated one of a plurality of hand pose features; and
outputting a features dataset defining a respective said current discrete pose value for each of said plurality of hand pose features for classifying a current hand pose of said hand;

wherein said plurality of in-plane rotation classifying functions are trained statistical classifiers.

7. The computer implemented method of claim 6 , wherein said plurality of hand feature classifying functions are trained statistical classifiers.

8. The computer implemented method of claim 6 , further comprising said features dataset includes a score assigned to said current hand pose, said score indicates a probability rate of said current hand pose matching one of a plurality of pre-defined hand poses features datasets.

9. The computer implemented method of claim 6 , further comprising scaling said image segment to comply with a scale of a training dataset used for training at least one of: said plurality of GOC classifying functions, said plurality of in-plane rotation classifying functions and said plurality of hand feature classifying functions.

10. A system for computing a features dataset classifying a pose of a hand of a human, comprising:
a storage storing a plurality of pre-defined hand poses features datasets;
a memory storing a code;
at least one processor coupled to said storage and said memory for executing said stored code, said code comprising:
code instructions to select a global orientation category (GOC) defining a spatial orientation of a hand of a human in a three dimensional (3D) space by applying a plurality of GOC classifying functions on a received image segment depicting said hand;
code instructions to identify an in-plane rotation by applying a plurality of in-plane rotation classifying functions on said image segment, said plurality of in-plane rotation classifying functions being selected according to said GOC;
code instructions to align said image segment in a 2 dimensional (2D) plane according to said identified in-plane rotation;
code instructions to apply a plurality of hand feature classifying functions on said aligned image segment, each one of said plurality of feature classifying functions outputting a current discrete pose value of an associated one of a plurality of hand pose features; and
code instructions to output a features dataset defining a respective said current discrete pose value for each of said plurality of hand pose features for classifying a current hand pose of said hand;
wherein said GOC is identified after estimating a center of hand by applying a plurality of center of hand classifying functions on said image segment.

11. The system of claim 10 , further comprising said center of hand is derived from a center of mass of said hand, said center of mass is identified by analyzing a depth image data available from an image depicting said hand, said depth image data maps a depth of said hand in said 3D space.

12. The system of claim 11 , further comprising said image is manipulated to remove at least one non-relevant image portion to produce an image segment.

13. The system of claim 10 , further comprising said features dataset includes a score assigned to said current hand pose, said score indicates a probability rate of said features dataset matching one of said plurality of pre-defined hand poses features datasets.

14. A software program product for computing a feature dataset classifying a pose of a hand of a human, comprising:
a non-transitory computer readable storage medium;
first program instructions to access a memory storing a plurality of hand pose features datasets;
second program instructions to select, using a computerized processor, a global orientation category (GOC) defining a spatial orientation of a hand of a human in a three dimensional (3D) space by applying a plurality of GOC classifying functions on a received image segment depicting said hand;
third program instructions to identify an in-plane rotation by applying a plurality of in-plane rotation classifying functions on said image segment, said plurality of in-plane rotation classifying functions being selected according to said GOC;
fourth program instructions to align said image segment in a 2 dimensional (2D) plane according to said identified in-plane rotation;
fifth program instructions to apply a plurality of hand feature classifying functions on said aligned image segment, each one of said plurality of feature classifying functions outputting a current discrete pose value of an associated one of a plurality of hand pose features; and
sixth program instructions to output a f

Assignees

• Microsoft Technology Licensing Llc

Inventors

• Krupka Eyal
• Vinnikov Alon
• Karmon Kfir

Classifications

• G06V10/764

  using classification, e.g. of video objects · CPC title

• G06F18/24

  Classification techniques · CPC title

• G06V40/11 Primary

  Hand-related biometrics; Hand pose recognition · CPC title

• G06F18/243

  relating to the number of classes · CPC title

• G06V10/48

  by mapping characteristic values of the pattern into a parameter space, e.g. Hough transformation · CPC title