# US 2016/0253570 A1 · Feature vector extraction device

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2016/0253570 A1
- Document ID: US20160253570A1
- Status: Published application
- Kind: A1
- Application number: 14/847,678
- Filed: 2015-09-08
- Published: 2016-09-01
- Earliest claimed: 2015-02-27
- Assignees: Kabushiki Kaisha Toshiba
- Applicants: Kabushiki Kaisha Toshiba
- Inventors: Nobuhiro NONOGAKI
- Web page: https://patentgrep.com/patent/US20160253570A1

## CPC

- G06V 40/113: {Recognition of static hand signs}
- G06V 10/467: {Encoded features or binary features, e.g. local binary patterns [LBP]}
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

A feature vector extraction device includes a cell learning unit setting a plurality of cells representing a position and range for counting a feature vector of a target on the basis of a plurality of images containing a target for learning use. A normalizer selects two feature points from among three or more feature points which are set in an image and represent the target for learning use, and normalizes a size and direction of each of the feature points. A feature point calculator calculates a mean position and a variation from the relevant mean position for each of other feature points than the selected two feature points of the normalized feature points. A cell decision unit decides a position of each of the cells on the basis of the mean position and decides a size of the each of the cells on the basis of the variation.

## Claims

### Claim 1

A feature vector extraction device comprising:
a cell learning unit that sets a plurality of cells each representing a position and range for counting a feature vector of a target on the basis of a plurality of images containing a target for learning use, the cell learning unit including
a normalizer that selects two feature points from among three or more feature points which are set in the image and represent the target for learning use, and normalizes a size and direction of each of the feature points,
a feature point calculator that calculates a mean position and a variation from the relevant mean position for each of other feature points than the selected two feature points of the normalized feature points, and
a cell decision unit that decides a position of each of the cells on the basis of the mean position and decides a size of the each of the cells on the basis of the variation.

### Claim 2

A feature vector extraction device comprising:
an image data acquisitor that acquires image data of an image containing a target;
a feature vector calculator that calculates a feature vector of the target on the basis of the image data; and
a feature vector statistical analyzer that counts all feature vectors in a cell for each of a plurality of cells which are respectively set to a plurality of areas in the image containing the target, each cell having a size corresponding to each area.

### Claim 3 (depends on 2)

The feature vector extraction device according to claim 2, further comprising a cell learning unit that sets the plurality of cells each representing a position and range for counting the feature vector of the target on the basis of a plurality of images containing a target for learning use.

### Claim 4 (depends on 2)

The feature vector extraction device according to claim 2, wherein the feature vector statistical analyzer generates a histogram obtained by counting the feature vector for each of the cells.

### Claim 5 (depends on 4)

The feature vector extraction device according to claim 4, further comprising a recognizer that uses the histogram generated by the feature vector statistical analyzer to recognize the target.

### Claim 6 (depends on 3)

The feature vector extraction device according to claim 3, wherein
the cell learning unit includes
a normalizer that selects two feature points from among three or more feature points which are set in each of a plurality of images containing a target for learning use, and normalizes a size and direction of each of the feature points,
a feature point calculator that calculates a mean position and a variation from the relevant mean position for each of other feature points than the selected two feature points of the normalized feature points, and
a cell decision unit that decides a position of each of the cells on the basis of the mean position and decides a size of the each of the cells on the basis of the variation.

### Claim 7 (depends on 6)

The feature vector extraction device according to claim 6, wherein the cell decision unit decides the size of the plurality of cells on the basis of a value obtained by multiplying a variance indicating the variation by a constant.

### Claim 8 (depends on 7)

The feature vector extraction device according to claim 7, wherein the cell decision unit sets the value obtained by multiplication by a constant depending on a kind of the target.

### Claim 9 (depends on 5)

The feature vector extraction device according to claim 5, wherein the recognized target is used as the target for learning use.

### Claim 10 (depends on 2)

The feature vector extraction device according to claim 2, wherein
the feature vector is a binary pattern obtained by converting into a binary a pixel value of each of a plurality of peripheral pixels surrounding a pixel of interest on the basis of a magnitude relationship between the pixel value and a pixel value of the pixel of interest, and
the feature vector calculator
sets a first value of the binary to the pixel value of the peripheral pixel which is larger than a value obtained by dividing the pixel value of the pixel of interest by a predetermined value and smaller than a value obtained by multiplying the pixel value of the pixel of interest by the predetermined value, and
sets a second value of the binary to the pixel value of the peripheral pixel which is equal to or less than a value obtained by dividing the pixel value of the pixel of interest by the predetermined value and equal to or more than a value obtained by multiplying the pixel value of the pixel of interest by the predetermined value.

### Claim 11 (depends on 1)

The feature vector extraction device according to claim 1, further comprising an image size adjuster that extends or reduces the image data to conform to the size of the cell.

### Claim 12 (depends on 2)

The feature vector extraction device according to claim 2, further comprising an image size adjuster that extends or reduces the image data to conform to the size of the cell.

### Claim 13 (depends on 1)

The feature vector extraction device according to claim 1, further comprising a cell size adjuster that extends or reduces the cell to conform to the size of the image data.

### Claim 14 (depends on 2)

The feature vector extraction device according to claim 2, further comprising a cell size adjuster that extends or reduces the cell to conform to the size of the image data.

### Claim 15 (depends on 1)

The feature vector extraction device according to claim 1, wherein the target for learning use is an object having a plurality of joints.

### Claim 16 (depends on 2)

The feature vector extraction device according to claim 2, wherein the target for learning use is an object having a plurality of joints.
