Person re-identification method, computer-readable storage medium, and terminal device

Patent metadata

Field	Value
Publication number	US-12525049-B2
Application number	US-202218088800-A
Country	US
Kind code	B2
Filing date	Dec 27, 2022
Priority date	May 31, 2022
Publication date	Jan 13, 2026
Grant date	Jan 13, 2026

Abstract

A person re-identification method, a storage medium, and a terminal device are provided. In the method, a loss function used during model training is a preset distribution-based triplet loss function constraining a difference between a mean of a negative sample feature distance and a mean of a positive sample feature distance to be larger than a preset difference threshold; where the positive sample feature distance is a distance between a feature of a reference image, and a feature of a positive sample image, and the negative sample feature distance is a distance between the feature of the reference image and a feature of a negative sample image. In this manner, it can constrain the mean of the positive sample feature distance and that of the negative sample feature distance, thereby improving the accuracy of person re-identification results.

First claim

What is claimed is:

1 . A computer-implemented person re-identification method for a terminal device having a camera, wherein the terminal device comprises an output device and a processor electrically coupled to the output device and the camera, and the method comprises:
obtaining, through the camera, an image or a video sequence; obtaining, through the processor from the terminal device, a person re-identification task; processing, through the processor, the person re-identification task using a preset person re-identification model to determine whether there is a person in the image or the video sequence, wherein the person is a target of identification specified by the person re-identification task; and obtaining, through the processor, a person re-identification result, and outputting, through the output device, the person re-identification result; wherein, a loss function used by the person re-identification model during training the person re-identification model is a preset distribution-based triplet loss function constraining a difference between a mean of a negative sample feature distance and a mean of a positive sample feature distance to be larger than a preset difference threshold; wherein the positive sample feature distance is a distance between a feature of a reference image captured by the camera of the terminal device and a feature of a positive sample image captured by the camera of the terminal device, and the negative sample feature distance is a distance between the feature of the reference image and a feature of a negative sample image captured by the camera of the terminal device; and wherein the distribution-based triplet loss function further constrains a first feature distance to be reduced to the mean of the positive sample feature distance and constrains a second feature distance to be increased to the mean of the negative sample feature distance; wherein the first feature distance is the positive sample feature distance larger than the mean of the positive sample feature distance, and the second feature distance is the negative sample feature distance smaller than the mean of the negative sample feature distance.

2 . The method of claim 1 , wherein the distribution-based triple loss function includes a macroscopic constraint function, a positive sample microscopic constraint function, and a negative sample microscopic constraint function; wherein the macroscopic constraint function is for constraining the difference between the mean of the negative sample feature distance and the mean of the positive sample feature distance to be larger than the difference threshold; wherein the positive sample micro constraint function is for constraining the first feature distance to be reduced to the mean of the positive sample feature distance, and the negative sample microcosmic constraint function is for constraining the second feature distance to be increased to the mean of the negative sample feature distance.

3 . The method of claim 2 , wherein the distribution-based triplet loss function is as an equation of:
L triplet_distribution =L macro +λ( L micro ap +L micro an );
where, L macro is the macroscopic constraint function, L micro ap is the positive sample microscopic constraint function, L micro an is the negative sample microscopic constraint function, λ is a preset weight coefficient, and L triplet_distribution is the distribution-based triplet loss function.

4 . The method of claim 3 , wherein the macroscopic constraint function is as an equation of:

L
macro

=

{

∑

i
=
1

N
op

D
⁡
(

f
i
a

,

f
i
p

)

-

∑

j
=
1

N
an

D
⁡
(

f
j
a

,

f
j
n

)

+
α

}

+

;

where, N ap is a number of positive sample feature pairs, N an is a number of negative sample feature pairs, i and j are serial numbers, f a , f p , and f n are the feature of the reference image, the feature of the positive sample image, and the feature of the negative sample image, respectively, D is a function for calculating distance, α is the difference threshold, {*} + =max{*, 0}, and max is a function for calculating maximum value.

5 . The method of claim 3 , wherein the positive sample microscopic constraint function is as an equation of:

L
mirco
ap

=

1

N
ap
act

⁢

{

∑

i
=
1

N
ap
act

(

D
⁡
(

Assignees

• Ubtech Robotics Corp Ltd

Inventors

• Wang Kan
• Pang Jianxin

Classifications

• G06V10/761

  Proximity, similarity or dissimilarity measures · CPC title

• Y02T10/40

  Engine management systems · CPC title

• G06V10/774

  Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06V10/764

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

• G06V10/82

  using neural networks · CPC title