US 12,738,088 B2Granted

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

AssigneeUbtech Robotics Corp Ltd

InventorsShuping Hu, Kan Wang, Huan Tan, Jianxin Pang

Life of this patent

Estimating the term…

Loading patent…

Abstract

A person re-identification method, a storage medium, and a terminal device are provided. In the method, a preset ratio-based triplet loss function is used as a loss function during training. The ratio-based triplet loss function limits a ratio of a positive sample feature distance to a negative sample feature distance to be less than a preset ratio threshold. The positive sample feature distance is a distance between a reference image feature and a positive sample image feature, and the negative sample feature distance is a distance between the reference image feature and a negative sample image feature. Compared with the existing absolute distance-based triplet loss function, in the case of small inter-class differences and large intra-class differences, the ratio-based triplet loss function can effectively improve the stability of model training, the features extracted by the trained model are more discriminative and robust, thereby improving the accuracy of person re-identification results.

Claims3

Dependent claims

A computer-implemented person re-identification method for a terminal device having a camera, comprising:

obtaining, from the terminal device, a person re-identification task including at least a triplet image set generated based on at least an image captured by the camera of the terminal device;
and
obtaining a person re-identification result by processing the obtained person re-identification task using a preset person re-identification model trained by taking a preset ratio-based triplet loss function that limits a ratio of a positive sample feature distance to a negative sample feature distance to less than a preset ratio threshold as a loss function;
wherein the positive sample feature distance is a distance between a reference image feature and a positive sample image feature, and the negative sample feature distance is a distance between the reference image feature and a negative sample image feature;
and wherein the ratio threshold is a variable, and the ratio threshold is negatively correlated with the negative sample feature distance;
wherein the ratio-based triplet loss function is:
L
triplet
⁢
_
⁢
ratio
=
1
N
tp
⁢
∑
i
=
1
P
∑
a
=
1
A
{
max
p
=
1
⁢
…
⁢
A
D
⁡
(
f
i
a
,
f
i
p
)
-
β
×
min
n
=
1
⁢
…
⁢
A
j
=
1
⁢
…
⁢
P
j
≠
i
D
⁡
(
f
i
a
,
f
i
P
)
}
+
;
where, Ntp is a number of triples, i and j are serial numbers, fα, fp and fn are the reference image feature, the positive sample image feature and the negative sample image feature, respectively, D is a function for calculating distance, β is the ratio threshold, {*}+=max {*, 0}, max is a function for calculating maximum value, min is a function for calculating minimum value, P and A are a number of pedestrians in a training batch of the model and a number of images of each person to be re-identified, respectively, and Ltriplet_ratio is the ratio-based triplet loss function;
and
wherein the ratio threshold is a variable satisfying an equation of:
β
=
F
⁡
(
D
⁡
(
f
i
a
,
f
i
n
)
)
=
{
β
1
D
⁡
(
f
i
a
,
f
j
n
)
<
D
1
an
k
×
D
⁡
(
f
i
a
,
f
j
n
)
+
b
D
1
an
≤
(
D
⁡
(
f
i
a
,
f
j
n
)
≤
D
2
an
β
2
D
⁡
(
f
i
a
,
f
j
n
)
>
D
2
an
;
where,
k
=
β
1
-
β
2
D
1
an
-
D
2
an
,
b
=
β
2
-
β
1
-
β
2
D
1
an
-
D
2
an
×
D
2
an
,
β1 is a preset upper limit of ratio threshold, β2 is a preset lower limit of ratio threshold,
D
1
a
⁢
n
is a preset lower limit of distance,
D
2
an
is a preset upper limit of distance, and F is a preset functional relationship.

220 more elements

A non-transitory computer-readable storage medium for storing one or more computer programs, wherein the one or more computer programs comprise:

instructions for obtaining, from a terminal device, a person re-identification task including at least a triplet image set generated based on at least an image captured by a camera of the terminal device;
and
instructions for obtaining a person re-identification result by processing the obtained person re-identification task using a preset person re-identification model trained by taking a preset ratio-based triplet loss function that limits a ratio of a positive sample feature distance to a negative sample feature distance to less than a preset ratio threshold as a loss function;
wherein the positive sample feature distance is a distance between a reference image feature and a positive sample image feature, and the negative sample feature distance is a distance between the reference image feature and a negative sample image feature;
and wherein the ratio threshold is a variable, and the ratio threshold is negatively correlated with the negative sample feature distance;
wherein the ratio-based triplet loss function is:
L
triplet
⁢
_
⁢
ratio
=
1
N
tp
⁢
∑
i
=
1
P
∑
a
=
1
A
{
max
p
=
1
⁢
…
⁢
A
D
⁡
(
f
i
a
,
f
i
p
)
-
β
×
min
n
=
1
⁢
…
⁢
A
j
=
1
⁢
…
⁢
P
j
≠
i
D
⁡
(
f
i
a
,
f
i
P
)
}
+
;
where, Ntp is a number of triples, i and j are serial numbers, fα, fp and fn are the reference image feature, the positive sample image feature and the negative sample image feature, respectively, D is a function for calculating distance, β is the ratio threshold, {*}+=max {*, 0}, max is a function for calculating maximum value, min is a function for calculating minimum value, P and A are a number of pedestrians in a training batch of the model and a number of images of each person to be re-identified, respectively, and Ltriplet_ratio is the ratio-based triplet loss function;
and
wherein the ratio threshold is a variable satisfying an equation of:
β
=
F
⁡
(
D
⁡
(
f
i
a
,
f
j
n
)
)
=
{
β
1
D
⁡
(
f
i
a
,
f
j
n
)
<
D
1
an
k
×
D
⁡
(
f
i
a
,
f
j
n
)
+
b
D
1
an
≤
D
⁡
(
f
i
a
,
f
j
n
)
≤
D
2
an
β
2
D
⁡
(
f
i
a
,
f
j
n
)
>
D
2
an
;
where,
k
=
β
1
-
β
2
D
1
an
-
D
2
an
,
b
=
β
2
-
β
1
-
β
2
D
1
an
-
D
2
an
×
D
2
an
,
β1 is a preset upper limit of ratio threshold, β2 is a preset lower limit of ratio threshold,
D
1
a
⁢
n
is a preset lower limit of distance,
D
2
a
⁢
n
is a preset upper limit of distance, and F is a preset functional relationship.

221 more elements

A terminal device, comprising:

a camera,
a processor;
a memory coupled to the processor;
and
one or more computer programs stored in the memory and executable on the processor;
wherein, the one or more computer programs comprise:
instructions for obtaining, from the terminal device, a person re-identification task including at least a triplet image set generated based on at least an image captured by the camera;
and
instructions for obtaining a person re-identification result by processing the obtained person re-identification task using a preset person re-identification model trained by taking a preset ratio-based triplet loss function that limits a ratio of a positive sample feature distance to a negative sample feature distance to less than a preset ratio threshold as a loss function;
wherein the positive sample feature distance is a distance between a reference image feature and a positive sample image feature, and the negative sample feature distance is a distance between the reference image feature and a negative sample image feature;
and wherein the ratio threshold is a variable, and the ratio threshold is negatively correlated with the negative sample feature distance;
wherein the ratio-based triplet loss function is:
L
triplet
⁢
_
⁢
ratio
=
1
N
tp
⁢
∑
i
=
1
P
∑
a
=
1
A
{
max
p
=
1
⁢
…
⁢
A
D
⁡
(
f
i
a
,
f
i
p
)
-
β
×
min
n
=
1
⁢
…
⁢
A
j
=
1
⁢
…
⁢
P
j
≠
i
D
⁡
(
f
i
a
,
f
i
P
)
}
+
;
where, Ntp is a number of triples, i and j are serial numbers, fα, fp and fn are the reference image feature, the positive sample image feature and the negative sample image feature, respectively, D is a function for calculating distance, β is the ratio threshold, {*}+=max {*, 0}, max is a function for calculating maximum value, min is a function for calculating minimum value, P and A are a number of pedestrians in a training batch of the model and a number of images of each person to be re-identified, respectively, and Ltriplet_ratio is the ratio-based triplet loss function;
and
wherein the ratio threshold is a variable satisfying an equation of:
β
=
F
⁡
(
D
⁡
(
f
i
a
,
f
j
n
)
)
=
{
β
1
D
⁡
(
f
i
a
,
f
j
n
)
<
D
1
an
k
×
D
⁡
(
f
i
a
,
f
j
n
)
+
b
D
1
an
≤
D
⁡
(
f
i
a
,
f
j
n
)
≤
D
2
an
β
2
D
⁡
(
f
i
a
,
f
j
n
)
>
D
2
an
;
where,
k
=
β
1
-
β
2
D
1
an
-
D
2
an
,
b
=
β
2
-
β
1
-
β
2
D
1
an
-
D
2
an
×
D
2
an
,
β1 is a preset upper limit of ratio threshold, β2 is a preset lower limit of ratio threshold,
D
1
a
⁢
n
is a preset lower limit of distance,
D
2
a
⁢
n
is a preset upper limit of distance, and F is a preset functional relationship.

227 more elements

Dependencies are inferred from the claim text.

Facts

Application: 18/078,027Filed 8 Dec 2022
Priority: CN 202210609192.031 May 2022
Published: US 2023/0386244 A130 Nov 2023
Granted: 15 Sep 2026
Term ends: Not estimatedThe source records don’t support an estimate.

Classification18 codes

G06V 40/103
Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands: static body considered as a whole, e.g. static pedestrian…
G06V 10/761
Image or video pattern matching; Proximity measures in feature spaces: proximity, similarity or dissimilarity measures
G06V 10/776
Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or…:…
G06F 18/211
Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation: selection of the most…