Pedestrian re-identification method and device

Patent metadata

Field	Value
Publication number	US-12394239-B2
Application number	US-202118013795-A
Country	US
Kind code	B2
Filing date	May 7, 2021
Priority date	Aug 25, 2020
Publication date	Aug 19, 2025
Grant date	Aug 19, 2025

Abstract

A method and an electronic device for pedestrian re-identification are provided. The method may include: collecting a target image set including at least two target images, where each target image includes at least one person; extracting a global feature and a head-shoulder feature of each person in each target image in the target image set, where the global feature is an overall appearance feature, and the head-shoulder feature is a feature of a head part and a shoulder part determining a representation feature of each person in each target image based on the global feature and the head-shoulder feature of each person in each target image; and determining a given person in different target images based on the representation feature of each person in each target image.

First claim

What is claimed is:

1. A computer-implemented method for pedestrian re-identification, comprising:
collecting a target image set comprising at least two target images, wherein each target image comprises at least one person;
extracting, using respective pre-trained models, a global feature and a head-shoulder feature of each person in the each target image in the target image set, wherein the global feature is an overall appearance feature, and the head-shoulder feature is a feature of a head part and a shoulder part;
determining a representation feature of the each person in the each target image based on the global feature and the head-shoulder feature of the each person in the each target image; and
determining whether a given person in different target images is the same person by assessing similarity between the respective representation feature of the each person in the each target image;
wherein the determining a representation feature of the each person in the each target image based on the global feature and head-shoulder feature of the each person in the each target image comprises:
acquiring, for the each person in the each target image, a weight value of the global feature and a weight value of the head-shoulder feature that correspond to a reference identification feature in the global feature of the person, wherein the weight value of the global feature that corresponds to the reference identification feature is less than the weight value of the head-shoulder feature that corresponds to the reference identification feature, and the weight value of the global feature and the weight value of the head-shoulder feature that correspond to the reference identification feature vary with the global feature; and
connecting the weighted features of both the global feature and the head-shoulder feature of the person to obtain the representation feature of the person.

2. The method according to claim 1 , wherein the connecting, for the each person in the each target image, weighted features of both the global feature and the head-shoulder feature of the person to obtain the representation feature of the person further comprises:
acquiring a preset weight value of the global feature and a preset weight value of the head-shoulder feature if the reference identification feature does not exist in the global feature of the person; and
connecting the weighted features of both the global feature and the head-shoulder feature of the person to obtain the representation feature of the person.

3. The method according to claim 1 , wherein the reference identification feature comprises: brightness of an image of a person in the target image being less than a preset brightness value, and/or a color of clothing of the or each person in the target image being black, and/or each target image is blurred.

4. The method according to claim 1 , wherein the determining a representation feature of the each person in the each target image based on the global feature and the head-shoulder feature of the each person in the each target image comprises:
inputting, for the each person in the each target image, the global feature of the person into a trained weight adaptive model, to obtain a weight value of the global feature of the person, the weight value being outputted by the weight adaptive model;
calculating a weight value of the head-shoulder feature of the person based on the weight value of the global feature of the person; and
connecting the weighted features of both the head-shoulder feature of the person and the global feature of the person to obtain the representation feature of the person.

5. The method according to claim 1 , wherein the determining a given person in different target images based on the representation feature of the each person in the each target image comprises:
calculating a distance between representation features of any two people in the different target images to form a distance matrix; and
determining the given person in the different target images based on the distance matrix.

6. The method according to claim 1 , wherein extracting the global feature of the each person in the each target image in the target image set comprises:
inputting an image of the each person in the each target image into a trained global deep learning model to obtain the global feature of the each person in the each target image, the global feature being outputted by the trained global deep learning model.

7. The method according to claim 1 , wherein extracting the head-shoulder feature of the each person in the each target image in the target image set comprises:
inputting the image of the each person in the each target image into a trained head-shoulder positioning model to obtain a head-shoulder area image in the image of the each person in the each target image, the head-shoulder area image being outputted by the trained head-shoulder positioning model; and
inputting the head-shoulder area image in the image of the each person in the each target image into a trained head-shoulder deep learning model to obtain the head-shoulder feature of the each person in the each target image, the head-shoulder feature being outputted by the trained head-shoulder deep learning model.

8. The method according to claim 1 , wherein the head-shoulder feature comprises one or more of: a gender, a face, a hairstyle, glasses, a shoulder shape, a scarf, a neck thickness.

9. An electronic device, comprising:
one or more processors; and
a storage apparatus, configured to store one or more programs,
wherein the one or more programs, when executed by the one or more processors, cause the one or more processors to perform operations comprising:
collecting a target image set comprising at least two target images, wherein each target image comprises at least one person;
extracting, using respective pre-trained models, a global feature and a head-shoulder feature of each person in the each target image in the target image set, wherein the global feature is an overall appearance feature, and the head-shoulder feature is a feature of a head part and a shoulder part;
determining a representation feature of the each person in the each target image based on the global feature and the head-shoulder feature of the each person in the each target image; and
determining whether a given person in different target images is the same person by assessing similarity between the respective representation feature of the each person in the each target image;
wherein the determining a representation feature of the each person in the each target image based on the global feature and head-shoulder feature of the each person in the each target image comprises:
acquiring, for the each person in the each target image, a weight value of the global feature and a weight value of the head-shoulder feature that correspond to a reference identification feature in the global feature of the person, wherein the weight value of the global feature that corresponds to the reference identification feature is less than the weight value of the head-shoulder feature that corresponds to the reference identification feature, and the weight value of the global feature and the weight value of the head-shoulder feature that correspond to the reference identification feature vary with the global feature; and
connecting the weighted features of both the global feature and the head-shoulder feature of the person to obtain the representation feature of the person.

10. The electronic device according to claim 9 , wherein the connecting, for the each person in the each target image, weighted features of both the global feature and the head-shoulder feature of the person to obtain the

Assignees

• Beijing Jingdong Shangke Information Technology Co Ltd
• Beijing Jingdong Century Trading Co Ltd

Inventors

• He Lingxiao
• Xu Boqiang
• Liao Xingyu
• Liu Wu
• Mei Tao
• Zhou Bowen

Classifications

• G06V10/44

  Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components · CPC title

• G06V10/751

  Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching · CPC title

• G06V10/70

  using pattern recognition or machine learning (optical pattern recognition or electronic computations therefor G06V10/88) · CPC title

• G06V10/42

  Global feature extraction by analysis of the whole pattern, e.g. using frequency domain transformations or autocorrelation · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title