Granted Patent B2

US 12,307,742 · App. 17/914,799 · Granted May 20, 2025

Inventors: Runze Zhang (Jiangsu, CN); Liang Jin (Jiangsu, CN); Zhenhua Guo (Jiangsu, CN)

Assignee: INSPUR SUZHOU INTELLIGENT TECHNOLOGY CO., LTD.

G06V10/764G06V40/10

Abstract

Claims (67)

1. A person Re-identification (Re-ID) method, comprising:

acquiring a sample set to be trained;

training a pre-constructed person Re-ID model by a data re-sampling method and a cross-validation method based on the sample set to be trained to obtain a trained person Re-ID model; and

performing person Re-ID based on the trained person Re-ID model;

wherein persons in any two groups are of different classes after the sample set to be trained is grouped according to the cross-validation method;

wherein the training the pre-constructed person Re-ID model by the data re-sampling method and the cross-validation method based on the sample set to be trained to obtain the trained person Re-ID model comprises:

determining a hyperparameter type, hyperparameter search space, and hyperparameter adjustment priority of the pre-constructed person Re-ID model, wherein the hyperparameter type comprises an image number of each person, a boundary of image number, an initial learning rate, a training cycle, a triplet loss threshold, and a Rerank parameter set; and

training the pre-constructed person Re-ID model based on the hyperparameter type, the hyperparameter search space, and the hyperparameter adjustment priority;

wherein a hyperparameter search space of the image number of each person comprises {2, 4, 8}; a hyperparameter search space of the boundary of image number comprises {10, 30, 50, 70, 100}; a hyperparameter search space of the initial learning rate comprises {0.00035, 0.001, 0.003, 0.01}; a hyperparameter search space of the training cycle comprises {80, 120, 160, 240}; a hyperparameter search space of the triplet loss threshold comprises {0.3, 1.2, 4.8, 10.0, 20.0}; a hyperparameter search space of k1 in the Rerank parameter set comprises {1, 5, 10, 15, 20}; a hyperparameter search space of k2 in the Rerank parameter set comprises {1, 2, 3, 4, 5, 6}; and a hyperparameter search space of λ in the Rerank parameter set comprises {0.3, 0.6, 0.9}; and

wherein a sequence of the hyperparameter adjustment priority from high to low comprises: the training cycle, the initial learning rate, the image number of each person, the triplet loss threshold, the boundary of image number, k1, k2, and λ.

2. The method according to claim 1 , wherein the training the pre-constructed person Re-ID model by the data re-sampling method and the cross-validation method based on the sample set to be trained to obtain the trained person Re-ID model comprises:

removing an image of a person with an image number greater than a first preset value from the sample set to be trained to obtain a remaining sample set;

determining a first target person with an image number less than a second preset value in the remaining sample set;

duplicating an image of the first target person to obtain a first duplicate image;

setting the first duplicate image and the remaining sample set as a first target sample set; and

training the pre-constructed person Re-ID model by the cross-validation method based on the first target sample set.

3. The method according to claim 2 , wherein the training the pre-constructed person Re-ID model by the cross-validation method based on the first target sample set comprises:

determining a second target person with an image number greater than the second preset value and less than the first preset value in the first target sample set, duplicating an image of the second target person to obtain a second duplicate image, and setting the second duplicate image and the first target sample set as a second target sample set, so as to perform training based on the second target sample set.

4. The method according to claim 2 , wherein the first preset value comprises 100, the second preset value comprises 4, and the cross-validation method comprises a 4-fold cross-validation method.

5. The method according to claim 1 , wherein the training the pre-constructed person Re-ID model by the data re-sampling method and the cross-validation method based on the sample set to be trained comprises:

dividing the sample set to be trained into N grouped sample sets according to person classes, a value of N being equal to a fold count of the cross-validation method;

selecting N−1 grouped sample sets in the N grouped sample sets as a training set, and setting a remaining grouped sample set as a validation set; and

training the pre-constructed person Re-ID model with the training set, validating the trained person Re-ID model with the validation set, and setting an average value of values of mAP and Rank1 computed with the validation set as an accuracy value of the trained person Re-ID model, so as to perform optimum selection on the trained person Re-ID model based on the accuracy value.

6. A person Re-identification (Re-ID) device, comprising:

a memory, configured to store a computer program;

a processor, configured to execute the computer program to implement a person Re-ID method, wherein the person Re-ID method comprises:

acquiring a sample set to be trained;

training a pre-constructed person Re-ID model by a data re-sampling method and a cross-validation method based on the sample set to be trained to obtain a trained person Re-ID model; and

performing person Re-ID based on the trained person Re-ID model;

wherein persons in any two groups are of different classes after the sample set to be trained is grouped according to the cross-validation method;

wherein the training the pre-constructed person Re-ID model by the data re-sampling method and the cross-validation method based on the sample set to be trained to obtain the trained person Re-ID model comprises:

determining a hyperparameter type, hyperparameter search space, and hyperparameter adjustment priority of the pre-constructed person Re-ID model, wherein the hyperparameter type comprises an image number of each person, a boundary of image number, an initial learning rate, a training cycle, a triplet loss threshold, and a Rerank parameter set; and

training the pre-constructed person Re-ID model based on the hyperparameter type, the hyperparameter search space, and the hyperparameter adjustment priority;

wherein a hyperparameter search space of the image number of each person comprises {2, 4, 8}; a hyperparameter search space of the boundary of image number comprises {10, 30, 50, 70, 100}; a hyperparameter search space of the initial learning rate comprises {0.00035, 0.001, 0.003, 0.01}; a hyperparameter search space of the training cycle comprises {80, 120, 160, 240}; a hyperparameter search space of the triplet loss threshold comprises {0.3, 1.2, 4.8, 10.0, 20.0}; a hyperparameter search space of k1 in the Rerank parameter set comprises {1, 5, 10, 15, 20}; a hyperparameter search space of k2 in the Rerank parameter set comprises {1, 2, 3, 4, 5, 6}; and a hyperparameter search space of λ in the Rerank parameter set comprises {0.3, 0.6, 0.9}; and

wherein a sequence of the hyperparameter adjustment priority from high to low comprises: the training cycle, the initial learning rate, the image number of each person, the triplet loss threshold, the boundary of image number, k1, k2, and λ.

7. A non-transitory computer-readable storage medium, having a computer program stored thereon which, when executed by a processor, implements a person Re-identification (Re-ID) method, wherein the person Re-ID method comprises:

acquiring a sample set to be trained;

training a pre-constructed person Re-ID model by a data re-sampling method and a cross-validation method based on the sample set to be trained to obtain a trained person Re-ID model; and

performing person Re-ID based on the trained person Re-ID model;

wherein persons in any two groups are of different classes after the sample set to be trained is grouped according to the cross-validation method;

wherein the training the pre-constructed person Re-ID model by the data re-sampling method and the cross-validation method based on the sample set to be trained to obtain the trained person Re-ID model comprises:

determining a hyperparameter type, hyperparameter search space, and hyperparameter adjustment priority of the pre-constructed person Re-ID model, wherein the hyperparameter type comprises an image number of each person, a boundary of image number, an initial learning rate, a training cycle, a triplet loss threshold, and a Rerank parameter set; and

training the pre-constructed person Re-ID model based on the hyperparameter type, the hyperparameter search space, and the hyperparameter adjustment priority;

wherein a hyperparameter search space of the image number of each person comprises {2, 4, 8}; a hyperparameter search space of the boundary of image number comprises {10, 30, 50, 70, 100}; a hyperparameter search space of the initial learning rate comprises {0.00035, 0.001, 0.003, 0.01}; a hyperparameter search space of the training cycle comprises {80, 120, 160, 240}; a hyperparameter search space of the triplet loss threshold comprises {0.3, 1.2, 4.8, 10.0, 20.0}; a hyperparameter search space of k1 in the Rerank parameter set comprises {1, 5, 10, 15, 20}; a hyperparameter search space of k2 in the Rerank parameter set comprises {1, 2, 3, 4, 5, 6}; and a hyperparameter search space of λ in the Rerank parameter set comprises {0.3, 0.6, 0.9}; and

wherein a sequence of the hyperparameter adjustment priority from high to low comprises: the training cycle, the initial learning rate, the image number of each person, the triplet loss threshold, the boundary of image number, k1, k2, and λ.

8. The person Re-identification (Re-ID) device according to claim 6 , wherein the training the pre-constructed person Re-ID model by the data re-sampling method and the cross-validation method based on the sample set to be trained to obtain the trained person Re-ID model comprises:

removing an image of a person with an image number greater than a first preset value from the sample set to be trained to obtain a remaining sample set;

determining a first target person with an image number less than a second preset value in the remaining sample set;

duplicating an image of the first target person to obtain a first duplicate image;

setting the first duplicate image and the remaining sample set as a first target sample set; and

training the pre-constructed person Re-ID model by the cross-validation method based on the first target sample set.

9. The person Re-identification (Re-ID) device according to claim 8 , wherein the training the pre-constructed person Re-ID model by the cross-validation method based on the first target sample set comprises:

determining a second target person with an image number greater than the second preset value and less than the first preset value in the first target sample set, duplicating an image of the second target person to obtain a second duplicate image, and setting the second duplicate image and the first target sample set as a second target sample set, so as to perform training based on the second target sample set.

10. The person Re-identification (Re-ID) device according to claim 8 , wherein the first preset value comprises 100, the second preset value comprises 4, and the cross-validation method comprises a 4-fold cross-validation method.

11. The person Re-identification (Re-ID) device according to claim 6 , wherein the training the pre-constructed person Re-ID model by the data re-sampling method and the cross-validation method based on the sample set to be trained comprises:

dividing the sample set to be trained into N grouped sample sets according to person classes, a value of N being equal to a fold count of the cross-validation method;

selecting N−1 grouped sample sets in the N grouped sample sets as a training set, and setting a remaining grouped sample set as a validation set; and

training the pre-constructed person Re-ID model with the training set, validating the trained person Re-ID model with the validation set, and setting an average value of values of mAP and Rank1 computed with the validation set as an accuracy value of the trained person Re-ID model, so as to perform optimum selection on the trained person Re-ID model based on the accuracy value.

12. The non-transitory computer-readable storage medium according to claim 7 , wherein the training the pre-constructed person Re-ID model by the data re-sampling method and the cross-validation method based on the sample set to be trained to obtain the trained person Re-ID model comprises:

removing an image of a person with an image number greater than a first preset value from the sample set to be trained to obtain a remaining sample set;

determining a first target person with an image number less than a second preset value in the remaining sample set;

duplicating an image of the first target person to obtain a first duplicate image;

setting the first duplicate image and the remaining sample set as a first target sample set; and

training the pre-constructed person Re-ID model by the cross-validation method based on the first target sample set.

13. The non-transitory computer-readable storage medium according to claim 12 , wherein the training the pre-constructed person Re-ID model by the cross-validation method based on the first target sample set comprises:

determining a second target person with an image number greater than the second preset value and less than the first preset value in the first target sample set, duplicating an image of the second target person to obtain a second duplicate image, and setting the second duplicate image and the first target sample set as a second target sample set, so as to perform training based on the second target sample set.

14. The non-transitory computer-readable storage medium according to claim 12 , wherein the first preset value comprises 100, the second preset value comprises 4, and the cross-validation method comprises a 4-fold cross-validation method.

Assignments (2)

Priority Claims (1)

Continuity (1)

References Cited (18)