KR20190068000A - Person Re-identification System in Multiple Camera Environments - Google Patents

Abstract

According to an embodiment of the present invention, a multiple image same person re-identifying system is the system for re-identifying the same person in multiple image environments. A method for re-identifying the same person comprises the steps of: extracting an object estimated as human from multiple images; discriminating a person who is unique to a person extracted from a single image; extracting 14 feature points corresponding to a person′s joint; extracting a feature vector for a person image of the extracted person; re-identifying a person presumed to be the same person as the person selected by a user; re-identifying the person again in which the user selects among the persons listed as re-identification candidates; selecting a person, who is determined by the user to be the same person, among a re-identified person candidate list; and displaying location information of the image, in which the person selected as the same person is present, on a map. Thus, the user can search for a person in a plurality of images different from one another.

Description

[Background Art] [0002] In a multi-image environment,

The present invention relates to a system and a technique for re-identifying the same person in a multi-

It is about.

The number of CCTV installations in public institutions and private facilities is increasing annually. Collection from CCTV

Image data is utilized in various ways. Especially in the criminal investigation process,

CCTV image data is widely used for collecting evidence such as inspections and copper lines. Various

The process of detecting the same person in CCTV video and grasping the line of the suspect

(A) the identification of the victim (s), the harm (s) of the crime, and the collection of evidence

Is an essential process. However, in a real criminal investigation,

And it consumes an excessive amount of time and manpower. Korea Registration

Patent No. 10-0995569 (registered on November 15, 2010)

Describe a human perverse recognition system that appears in images through the stars. The above system

Recognizes the same person among multiple images in proximity and recognizes the character's wandering. Then

Because it is difficult to identify the same person in a place that is somewhat distant from the image

There is a difficulty in recognition of roaming. Therefore, regardless of the location information of the image data

A technique for identifying the same person based on the feature points of the person extracted from the image

Is required.

Extraction of human objects from multiple cameras and the same discrimination as one pipeline

So that the user can search for the person to be searched in different images.

The multi-image same re-identification system according to an embodiment of the present invention includes a multi-

The same identity re-identification system in the home environment is disclosed. Room for Identification

The method includes extracting an object estimated as a person from multiple images; In a single image

Determining a unique person for the extracted person; 14 corresponding to the human joints

Extracting feature points of the feature points; Characteristics of the person image on the extracted person

Extracting a vector; A person who is presumed to be the same person as the person selected by the user.

Re-identifying; Among the people listed as candidates for re-identification,

Re-identifying again; Of the re-identified person candidates list,

Selecting a person to be judged; Image of the person selected as the same person

And displaying the location information on the map.

In the multi-image environment, when the user identifies the same person,

There is a difficulty in searching for water. The characters in each video

And then reassigns them to the user through the re-identification model.

Identical list is given the same unique number. When a real person identifies the same person

It is judged to be the same person based on the impressions between the characters. The re-identification of this system

In addition to the process of re-identification of people,

Extract and determine the points. By the body part of the figure extracted from multiple images

The extraction point is extracted and a human re-identification learning model is created.

It can track the path of the same person by displaying the location of the displayed images

Figure 1 is a block diagram of an exemplary embodiment of the present invention,
It is a process to identify a person.
FIG. 2 is a block diagram illustrating an exemplary embodiment of the present invention. Referring to FIG. 2,
After extracting images for each human body area, learn about each feature point and develop a model
FIG.

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings.

Explain in detail. However, it should be understood that the present invention is limited or limited by the embodiments

. Like reference symbols in the drawings denote like elements.

1 is a block diagram of an embodiment of the present invention,

And FIG.

Referring to FIG. 1, in the human object recognition (S110), a person

Extract the object.

At this time, the human object recognition (S110) is based on the deep learning object recognition model,

. The step of extracting the human object is specifically a frame of 1 second

, A step of estimating a range estimated as a human object for each frame

. The Deep Learning Object Recognition Model is based on

Therefore, it is necessary to divide the image into frames. The input multiple images

Divide the frame by 1 second for each. For each frame,

We perform human object recognition with Ning object recognition model. Deep Learning Object Recognition Model

Predicts the x-coordinate, y-coordinate, width, and width. Based on the four predicted values

To handle a square bounding box for a human object. Each of these processes

Apply to the image.

Referring to FIG. 1, the extracted human objects are stored in a database,

The user selects a person to be searched in the set (S120).

At this time, a unique number is assigned to the extracted human objects in step S110,

It is stored in the database. The assigned unique number is S140, and in S150,

It is used in the star process. When a user selects a person to be searched,

You can search by each household. If you select the person you are looking for,

And the re-identification is performed with the person of the unique number (S130).

Referring to FIG. 1, it is assumed that, based on the person selected by the user,

The re-identification is performed (S130).

At this time, re-identification is performed using the re-identification deep learning model. Re-identification deep learning model

The learning process will be described in detail with reference to FIG.

2 is a flowchart showing a learning process of the person re-identification deep learning model.

Referring to FIG. 2, the object recognition deep learning model (S210)

And extracts the human object.

At this time, the object recognition deep learning model uses the same principle as the object recognition model of (S110)

And extracts the object.

Referring to FIG. 2, a unique number is assigned to the human object images extracted in step S210

; In step S220, a person image for metric learning is sampled

.

A unique number is assigned to the human object images extracted in step S210,

A person with a specific number (standard figure), the same person as the reference person

The object image, and the person object image different from the reference person. All of the characters

Since the matrix learning is performed for the unique number, the unique number is selected for each sequence.

Sample a portrait image with the same unique number based on the selected person's unique number

. Means another image of the same person. And the figure number is different.

Sampling the portrait image of a person with a different unique number.

Referring to FIG. 2, an image of the reference human body part extracted at (S220)

(S230).

At this time, the image of human body part is divided into head, body, leg,

Lee, arms are divided into left and right. Head, torso, left arm, right arm, left foot, right foot 7

It is divided into dogs. Extracting an image of a human body part;

Referring to FIG. 2, minutiae points of each part extracted in S230 are extracted (S240).

The reason for extracting minutiae points for each part is that the unique features of each person image

It is for learning the settlement. In general, a person is a person

The principle of discriminating the same person is based on the impression of the person. versus

Because CTV images of the part are low-resolution image quality,

It is difficult to discriminate. For this reason, most of the human re-identification algorithms are box-

Feature points are extracted from the portrait image to identify the same person.

In carrying out the person re-identification of the present invention,

Is to separate and extract the characteristic points of the original. Body image in person image

And the human re-identification deep-run model is used to extract the feature points of each body part

.

Referring to FIG. 2, a matrix learning (S250) is performed on the minutiae extracted for each human body region

do.

At this time, triplet loss is used for metric learning. The definition of triplet loss is

As follows.

는 기준이 되는 Class로 정의된다. a = anchor

Is defined as a reference class. a = anchor

는

(기준이 되는 인물)와 같은 Class(소속)로 정의된다. p =

The

(The reference person) and the same class (belonging). p =

positive

는

(기준이 되는 인물)과 다른 Class로 정의된다. n = negative

The

Is defined as a different class from the standard (the reference person). n = negative

과

의 거리는 최대하는 The triplet loss is the minimum distance between fia and fip,

and

The distance of

Direction.

In the re-identification of people, metric learning measures the distance between feature points of each person

It is for the purpose of classification. Extraction from re-identification model without metric learning

If the same person is judged from the distance between the feature points, high accuracy can not be obtained

All. Therefore, through the metric learning, feature points in the same class are close to each other

In the running direction, feature points in different classes are

do.

Referring to FIG. 2, (S260)

Create a human re-identification model after performing the hum.

Claims (6)

We extract information of human object from multiple images using a deep learning model,
Storing the water in a database; Among the people stored in the database
Selecting a person to be searched by a user; Based on selected characters
Re-identifying the same person; Used for re-identified persons presumed to be the same person
Re-identifying based on the self-selected character; Presumed to be the same person
A step of displaying the positional information of the image data on which the person is displayed on the map
Multi-imaging human driver tracking system
The method according to claim 1,
The step of storing the person extracted from the image in the database includes:
A human object recognition deep learning model is used to extract a frame from an image,
How to detect people by frame
The method according to claim 1,
Among the people stored in the database, the user selects the person
Shows a list of the detected persons in time of the video data,
How to choose the person you are looking for
The method according to claim 1,
The step of re-identifying the same person based on the selected person includes:
Based on the human re-identification deep learning model, it extracts feature points for each person
Based on the extracted feature points,
Includes a way to show a list of rules.
5. The method of claim 4,
The step of extracting feature points for each person using the deep learning person re-
The image of the person is extracted by the body parts of the head, torso, arms, and legs,
Based on the feature points of the image, the distance between the feature points decreases
Learning model
The method according to claim 1,
In the re-identified person list, the user selects a person presumed to be the same person
The step is to re-identify the person as a person whose distance from the selected person is close to the feature point
How to Reconfigure

Images