CN112381056B - A cross-domain pedestrian re-identification method and system integrating multiple source domains - Google Patents

Abstract

The invention discloses a cross-domain pedestrian re-identification method and system integrating multiple source domains. The method includes: when training a cross-domain person re-identification model, acquiring multiple sets of source domain data sets, and using a set of source domain data sets to train a representation learning network and a metric learning network to perform the cross-domain person re-identification model on the cross-domain person re-identification model. After training, the trained cross-domain pedestrian re-identification model is obtained. When performing pedestrian re-identification, the pedestrian samples to be identified are input into the trained cross-domain pedestrian re-identification model, and the pedestrian re-identification result is obtained. By adopting the method and system of the present invention, the data of multiple source domains are integrated in the training process, and the feature representation of pedestrians can be better learned. Compared with a single training model, the accuracy of pedestrian re-identification can be improved, which effectively solves the problem of The problem of model performance degradation caused by differences between domains.

Description

A cross-domain pedestrian re-identification method and system integrating multiple source domains

technical field

The invention relates to the technical field of pedestrian re-identification, in particular to a cross-domain pedestrian re-identification method and system integrating multiple source domains.

Background technique

Pedestrian re-identification, also known as pedestrian re-identification, is a technology that uses computer vision technology to determine whether there is a specific pedestrian in an image or video sequence. In recent years, due to the rapid development of deep learning, the performance of person re-identification algorithms has also been unprecedentedly improved. When implementing a deep learning-based person re-identification system, when the model trained by a single training set is applied to the actual scene, the performance drops very significantly. This is because when the trained model is directly used in the actual scene, it is often The performance of the model is degraded because of the inter-domain differences between the pedestrians in the actual scene and the training dataset.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a cross-domain pedestrian re-identification method and system integrating multiple source domains, which can effectively solve the problem of model performance degradation caused by differences between domains and improve the accuracy of pedestrian re-identification.

For achieving the above object, the present invention provides the following scheme:

A pedestrian re-identification method, comprising:

obtaining a pedestrian sample pair to be identified; the pedestrian sample pair to be identified includes two pedestrian samples to be identified;

The pedestrian samples to be identified are input into a trained cross-domain pedestrian re-identification model, and a pedestrian re-identification result is obtained; the cross-domain pedestrian re-identification model includes multiple representation learning networks and a metric learning network, and the representation The learning network adopts the ResNet-50 network structure, and the metric learning network adopts a three-layer fully connected network structure;

The training method of the cross-domain pedestrian re-identification model specifically includes:

Obtain multiple sets of source domain data sets; each group of said source domain data sets includes a plurality of pedestrian samples to be trained, and each said pedestrian sample to be trained corresponds to a pedestrian label; the number of said source domain data sets is the same as all The above means that the number of learning networks is the same;

A set of source domain data sets are used to train a representation learning network and a metric learning network to train the cross-domain person re-identification model, and a trained cross-domain person re-identification model is obtained.

Optionally, the pedestrian sample to be identified is input into the trained cross-domain pedestrian re-identification model to obtain a pedestrian re-identification result, which specifically includes:

Inputting the pedestrian samples to be identified into the trained cross-domain pedestrian re-identification model, a plurality of pedestrian feature groups to be identified are obtained; each pedestrian feature group includes two pedestrian features;

Calculate the distance between the two pedestrian features of each pedestrian feature group to be identified, and obtain a plurality of pedestrian feature distances to be identified;

Determine whether the number of pedestrian feature distances to be identified that is less than the preset distance exceeds half of the total number of pedestrian feature distances to be identified; if so, determine that the pedestrian samples in the pair of pedestrian samples to be identified are the same; if not, then It is determined that the pedestrian samples in the pair of pedestrian samples to be identified are different.

Optionally, the method of using a set of source domain data sets to train a representation learning network and a metric learning network is used to train the cross-domain pedestrian re-identification model to obtain a trained cross-domain pedestrian re-identification model, which specifically includes:

Select a set of source domain data sets and a representation learning network; the source domain data set includes a positive sample pair and a negative sample pair, the positive sample pair includes two pedestrian samples with the same pedestrian, and the negative sample pair includes different pedestrians two pedestrian samples of , and one pedestrian sample in the negative sample pair is the same as one pedestrian sample in the positive sample pair;

Input the selected source domain dataset into the selected representation learning network to obtain multiple first pedestrian features;

Inputting a plurality of the first pedestrian features into the metric learning network to obtain a plurality of second pedestrian features;

Calculate the pedestrian feature distance corresponding to the positive sample pair and the pedestrian feature distance corresponding to the negative sample pair according to the second pedestrian feature;

The loss value is calculated according to the pedestrian feature distance corresponding to the positive sample pair and the pedestrian feature distance corresponding to the negative sample pair, and taking the minimum loss value as the optimization goal, the gradient descent method is used to optimize the cross-domain pedestrian re-identification model. parameter;

Determine whether all source domain datasets have been selected. If so, get the trained cross-domain pedestrian re-identification model. If not, select a source domain dataset from the unselected source domain dataset, and select a source domain dataset from the unselected source domain dataset. Select a representation learning network from the selected representation learning network, and then return to the step "input the selected source domain dataset into the selected representation learning network to obtain multiple first pedestrian features".

optional,

The calculation formula of the pedestrian feature distance is as follows:

为

与

的行人特征距离，

为第t个源域数据集中第i个行人样本，

为第t个源域数据集中第j个行人样本，

为

对应的第二行人特征，

为

对应的第二行人特征，C为度量学习网络；In the formula,

for

and

The pedestrian characteristic distance of ,

is the i-th pedestrian sample in the t-th source domain dataset,

is the jth pedestrian sample in the tth source domain dataset,

for

The corresponding second pedestrian feature,

for

The corresponding second pedestrian feature, C is the metric learning network;

The formula for calculating the loss value is as follows:

L _triple =[d _p -d _n +α] ₊

In the formula, L _triple is the loss value, d _p is the pedestrian feature distance corresponding to the negative sample pair, d _n is the pedestrian feature distance corresponding to the positive sample pair, α is the preset interval between d _p and d _n , [d _p -d _n +α] ₊ means that when d _n <d _p +α, the loss value remains unchanged, and when d _{n ≥} d _p +α, the loss value is 0.

The present invention also provides a pedestrian re-identification system, comprising:

A pedestrian re-identification module is used to obtain a pair of pedestrian samples to be identified; the pair of pedestrian samples to be identified includes two pedestrian samples to be identified; it is also used to input the pair of pedestrian samples to be identified into a trained cross-domain In the pedestrian re-identification model, a pedestrian re-identification result is obtained; the cross-domain pedestrian re-identification model includes a plurality of representation learning networks and a metric learning network, the representation learning network adopts the ResNet-50 network structure, and the metric learning network adopts Three-layer fully connected network structure;

The training module is used to train the cross-domain person re-identification model to obtain a trained cross-domain person re-identification model; the training method specifically includes:

Obtain multiple sets of source domain data sets; each group of said source domain data sets includes a plurality of pedestrian samples to be trained, and each said pedestrian sample to be trained corresponds to a pedestrian label; the number of said source domain data sets is the same as all The above means that the number of learning networks is the same;

A set of source domain data sets are used to train a representation learning network and a metric learning network to train the cross-domain person re-identification model, and a trained cross-domain person re-identification model is obtained.

Optionally, the pedestrian re-identification module specifically includes:

a pedestrian feature group generation unit, configured to input the pedestrian samples to be identified into the trained cross-domain pedestrian re-identification model to obtain a plurality of pedestrian feature groups to be identified; each pedestrian feature group includes two pedestrian features;

A pedestrian feature distance calculation unit to be identified, used to calculate the distance between two pedestrian features of each pedestrian feature group to be identified, to obtain a plurality of pedestrian feature distances to be identified;

The first judgment unit is used for judging whether the number of pedestrian feature distances to be identified that is less than the preset distance exceeds half of the total number of pedestrian feature distances to be identified; if so, then determine the pedestrian in the pedestrian sample pair to be identified. The samples are the same; if not, it is determined that the pedestrian samples in the pair of pedestrian samples to be identified are different.

Optionally, the training module specifically includes:

The selection unit is used to select a set of source domain data sets and a representation learning network; the source domain data set includes a positive sample pair and a negative sample pair, the positive sample pair includes two pedestrian samples with the same pedestrian, the negative sample pair The sample pair includes two pedestrian samples with different pedestrians, and one pedestrian sample in the negative sample pair is the same as one pedestrian sample in the positive sample pair;

The first pedestrian feature generation unit is used to input the selected source domain data set into the selected representation learning network to obtain a plurality of first pedestrian features;

A second pedestrian feature generation unit, configured to input a plurality of the first pedestrian features into the metric learning network to obtain a plurality of second pedestrian features;

a pedestrian feature distance calculation unit for the sample pair, configured to calculate the pedestrian feature distance corresponding to the positive sample pair and the pedestrian feature distance corresponding to the negative sample pair according to the second pedestrian feature;

The model optimization unit is used to calculate the loss value according to the pedestrian feature distance corresponding to the positive sample pair and the pedestrian feature distance corresponding to the negative sample pair, and take the minimum loss value as the optimization goal, and use the gradient descent method to optimize the cross-domain Parameters in the pedestrian re-identification model;

The second judgment unit is used to judge whether all the source domain data sets have been selected. If so, obtain a trained cross-domain pedestrian re-identification model. If not, select a source domain data set from the unselected source domain data sets. set, and select a representation learning network from the unselected representation learning networks, and then execute the first pedestrian feature generation unit.

optional,

The calculation formula of the pedestrian feature distance is as follows:

为

与

的行人特征距离，

为第t个源域数据集中第i个行人样本，

为第t个源域数据集中第j个行人样本，

为

对应的第二行人特征，

为

对应的第二行人特征，C为度量学习网络；In the formula,

for

and

The pedestrian characteristic distance of ,

is the i-th pedestrian sample in the t-th source domain dataset,

is the jth pedestrian sample in the tth source domain dataset,

for

The corresponding second pedestrian feature,

for

The corresponding second pedestrian feature, C is the metric learning network;

The formula for calculating the loss value is as follows:

L _triple =[d _p -d _n +α] ₊

In the formula, L _triple is the loss value, d _p is the pedestrian feature distance corresponding to the negative sample pair, d _n is the pedestrian feature distance corresponding to the positive sample pair, α is the preset interval between d _p and d _n , [d _p -d _n +α] ₊ means that when d _n <d _p +α, the loss value remains unchanged, and when d _{n ≥} d _p +α, the loss value is 0.

Compared with the prior art, the beneficial effects of the present invention are:

The invention proposes a cross-domain pedestrian re-identification method and system integrating multiple source domains. When training the cross-domain pedestrian re-identification model, multiple sets of source domain data sets are obtained, and a set of source domain data sets are used to train one Representation learning network and metric learning network are used to train cross-domain pedestrian re-identification model, and a trained cross-domain pedestrian re-identification model is obtained. In the pedestrian re-identification model, the pedestrian re-identification result is obtained. The invention integrates the data of multiple source domains in the training process, and can better learn the feature representation of pedestrians. Compared with a single training model, the accuracy of pedestrian re-identification can be improved, and the model caused by differences between domains can be effectively solved. performance degradation issues.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the present invention. In the embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative labor.

FIG. 1 is a flowchart of a cross-domain pedestrian re-identification method integrating multiple source domains in an embodiment of the present invention;

2 is a schematic diagram of a training phase in an embodiment of the present invention;

FIG. 3 is a schematic diagram of a testing stage in an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

The purpose of the present invention is to provide a cross-domain pedestrian re-identification method and system integrating multiple source domains, which can effectively solve the problem of model performance degradation caused by differences between domains and improve the accuracy of pedestrian re-identification.

In order to make the above objects, features and advantages of the present invention more clearly understood, the present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

Example

FIG. 1 is a flowchart of a cross-domain pedestrian re-identification method integrating multiple source domains according to an embodiment of the present invention. As shown in FIG. 1 , a cross-domain pedestrian re-identification method integrating multiple source domains includes:

Step 101: Obtain a pedestrian sample pair to be identified; the pedestrian sample pair to be identified includes two pedestrian samples to be identified. Among them, pedestrian samples are pedestrian pictures.

Step 102: Input the pedestrian samples to be identified into the trained cross-domain pedestrian re-identification model to obtain a pedestrian re-identification result; the cross-domain pedestrian re-identification model includes multiple representation learning networks and a metric learning network, and the representation learning network adopts ResNet-50 network structure, the metric learning network adopts a three-layer fully connected network structure.

Step 102 specifically includes:

Input the pedestrian samples to be identified into the trained cross-domain pedestrian re-identification model to obtain a plurality of pedestrian feature groups to be identified; each pedestrian feature group includes two pedestrian features;

Calculate the distance between the two pedestrian features of each pedestrian feature group to be identified, and obtain a plurality of pedestrian feature distances to be identified;

Determine whether the number of pedestrian feature distances to be identified that is less than the preset distance exceeds half of the total number of pedestrian feature distances to be identified; if so, determine that the pedestrian samples in the pair of pedestrian samples to be identified are the same; The pedestrian samples in the pedestrian sample pair are different.

in,

The training method of cross-domain person re-identification model, including:

Obtain multiple sets of source domain datasets; each source domain dataset includes multiple pedestrian samples to be trained, and each pedestrian sample to be trained corresponds to a pedestrian label; the number of source domain datasets is the same as the number of representation learning networks ;

A set of source domain datasets are used to train a representation learning network and a metric learning network to train a cross-domain person re-identification model, and a trained cross-domain person re-identification model is obtained, including:

Select a set of source domain datasets and a representation learning network; the source domain dataset includes positive sample pairs and negative sample pairs, the positive sample pair includes two pedestrian samples with the same pedestrian, and the negative sample pair includes two pedestrian samples with different pedestrians. And one pedestrian sample in the negative sample pair is the same as one pedestrian sample in the positive sample pair; (that is, two different pictures of the same pedestrian are selected as positive samples, and the image of another pedestrian and one of the positive samples are used as negative samples. sample).

Input the selected source domain dataset into the selected representation learning network to obtain multiple first pedestrian features.

A plurality of first pedestrian features are input into the metric learning network to obtain a plurality of second pedestrian features.

According to the second pedestrian feature, the pedestrian feature distance corresponding to the positive sample pair and the pedestrian feature distance corresponding to the negative sample pair are calculated.

The loss value is calculated according to the corresponding pedestrian feature distance of the positive sample pair and the corresponding pedestrian feature distance of the negative sample pair, and taking the minimum loss value as the optimization goal, the gradient descent method is used to optimize the parameters in the cross-domain pedestrian re-identification model.

Determine whether all source domain datasets have been selected. If so, get the trained cross-domain pedestrian re-identification model. If not, select a source domain dataset from the unselected source domain dataset, and select a source domain dataset from the unselected source domain dataset. Select a representation learning network from the selected representation learning network, and then return to the step "input the selected source domain dataset into the selected representation learning network to obtain multiple first pedestrian features".

in,

The calculation formula of pedestrian feature distance is as follows:

为

与

的行人特征距离，

为第t个源域数据集中第i个行人样本，

为第t个源域数据集中第j个行人样本，

为

对应的第二行人特征，

为

对应的第二行人特征，C为度量学习网络；In the formula,

for

and

The pedestrian characteristic distance of ,

is the i-th pedestrian sample in the t-th source domain dataset,

is the jth pedestrian sample in the tth source domain dataset,

for

The corresponding second pedestrian feature,

for

The corresponding second pedestrian feature, C is the metric learning network;

The formula for calculating the loss value is as follows:

L _triple =[d _p -d _n +α] ₊

In the formula, L _triple is the loss value, d _p is the pedestrian feature distance corresponding to the negative sample pair, d _n is the pedestrian feature distance corresponding to the positive sample pair, α is the preset interval between d _p and d _n , [d _p -d _n +α] ₊ means that when d _n <d _p +α, the loss value remains unchanged, and when d _{n ≥} d _p +α, the loss value is 0. The purpose of setting α is to have a small separation between d _p and d _n so that d _n should always be larger than d _p .

The present invention also provides a cross-domain pedestrian re-identification system integrating multiple source domains, including a pedestrian re-identification module and a training module.

The pedestrian re-identification module is used to obtain the pair of pedestrian samples to be identified; the pair of pedestrian samples to be identified includes two pedestrian samples to be identified; it is also used to input the pair of pedestrian samples to be identified into the trained cross-domain pedestrian re-identification model , the person re-identification results are obtained; the cross-domain person re-identification model includes multiple representation learning networks and a metric learning network, the representation learning network adopts the ResNet-50 network structure, and the metric learning network adopts a three-layer fully connected network structure.

Pedestrian re-identification module, including:

The pedestrian feature group generation unit is used to input the pedestrian samples to be identified into the trained cross-domain pedestrian re-identification model to obtain a plurality of pedestrian feature groups to be identified; each pedestrian feature group includes two pedestrian features;

A pedestrian feature distance calculation unit to be identified, used to calculate the distance between two pedestrian features of each pedestrian feature group to be identified, to obtain a plurality of pedestrian feature distances to be identified;

The first judgment unit is used for judging whether the number of pedestrian feature distances to be identified that is less than the preset distance exceeds half of the total number of pedestrian feature distances to be identified; if so, determine that the pedestrian samples in the pedestrian sample pair to be identified are the same; if If not, it is determined that the pedestrian samples in the pair of pedestrian samples to be identified are different.

The training module is used to train the cross-domain person re-identification model to obtain a trained cross-domain person re-identification model; the training method specifically includes:

Obtain multiple sets of source domain datasets; each source domain dataset includes multiple pedestrian samples to be trained, and each pedestrian sample to be trained corresponds to a pedestrian label; the number of source domain datasets is the same as the number of representation learning networks ;

A set of source domain datasets are used to train a representation learning network and a metric learning network to train the cross-domain person re-identification model, and the trained cross-domain person re-identification model is obtained.

Training modules, including:

The selection unit is used to select a set of source domain data sets and a representation learning network; the source domain data set includes a positive sample pair and a negative sample pair, the positive sample pair includes two pedestrian samples with the same pedestrian, and the negative sample pair includes different pedestrians. Two pedestrian samples, and one pedestrian sample in the negative sample pair is the same as one pedestrian sample in the positive sample pair;

The first pedestrian feature generation unit is used to input the selected source domain data set into the selected representation learning network to obtain a plurality of first pedestrian features;

The second pedestrian feature generation unit is used to input a plurality of first pedestrian features into the metric learning network to obtain a plurality of second pedestrian features;

The pedestrian feature distance calculation unit of the sample pair is used to calculate the pedestrian feature distance corresponding to the positive sample pair and the pedestrian feature distance corresponding to the negative sample pair according to the second pedestrian feature;

The model optimization unit is used to calculate the loss value according to the pedestrian feature distance corresponding to the positive sample pair and the pedestrian feature distance corresponding to the negative sample pair, and take the minimum loss value as the optimization goal, and use the gradient descent method to optimize the cross-domain pedestrian re-identification model. parameter;

The second judgment unit is used to judge whether all the source domain data sets have been selected. If so, obtain a trained cross-domain pedestrian re-identification model. If not, select a source domain data set from the unselected source domain data sets. set, and select a representation learning network among the unselected representation learning networks, and then execute the first pedestrian feature generation unit.

in,

The calculation formula of pedestrian feature distance is as follows:

为

与

的行人特征距离，

为第t个源域数据集中第i个行人样本，

为第t个源域数据集中第j个行人样本，

为

对应的第二行人特征，

为

对应的第二行人特征，C为度量学习网络；In the formula,

for

and

The pedestrian characteristic distance of ,

is the i-th pedestrian sample in the t-th source domain dataset,

is the jth pedestrian sample in the tth source domain dataset,

for

The corresponding second pedestrian feature,

for

The corresponding second pedestrian feature, C is the metric learning network;

The formula for calculating the loss value is as follows:

L _triple [d _p -d _n +α] ₊

In the formula, L _triple is the loss value, d _p is the pedestrian feature distance corresponding to the negative sample pair, d _n is the pedestrian feature distance corresponding to the positive sample pair, α is the preset interval between d _p and d _n , [d _p -d _n +α] ₊ means that when d _n <d _p +α, the loss value remains unchanged, and when d _{n ≥} d _p +α, the loss value is 0.

The present invention further illustrates the cross-domain pedestrian re-identification method by fusing multiple source domains through the following description.

Get m labeled source domain datasets:

表示在第t个源域数据集中行人样本

和对应的行人标签

Among them, S ^t represents the source domain dataset;

represents a sample of pedestrians in the t-th source domain dataset

and the corresponding pedestrian labels

Get test data (target domain):

T={x _i |i=1,2,...,n}

Among them, _xi is the pedestrian sample in the target domain dataset.

The goal of the cross-domain pedestrian re-identification model that fuses multiple source domains is to input any pedestrian samples in the target domain T into the model for x _i and x _j to obtain the corresponding pedestrian features, which are determined by judging whether the distance between the pedestrian features is less than a threshold. whether it is the same pedestrian.

The cross-domain person re-identification model that fuses multiple source domains includes two core modules: a representation learning network and a fusion feature network (i.e., a metric learning network).

1.m representation learning networks, implemented using ResNet-50

For the source domain dataset S ^t , let the t-th representation learning network be denoted as H ^t (S ^t ,ω ^t ), and H ^t is determined by the t-th source domain dataset S ^t in (cross-domain fusion of multiple source domains Person re-identification model), ω ^t is the learned parameter of H ^t after training on the source domain dataset S ^t .

输入到表示学习网络H^t(S^t,ω^t)后得到的行人特征为：

Pedestrian samples in the t-th source domain

The pedestrian features obtained after inputting to the representation learning network H ^t (S ^t ,ω ^t ) are:

2. A fusion feature network, implemented using 3 fully connected layers

其中

为第t个源域中的行人样本

输入表示学习网络H^t(S^t,ω^t)后得到的行人特征，θ为模型完成训练后度量学习网络的参数。Let the fusion feature network be expressed as

in

is the pedestrian sample in the t-th source domain

The input represents the pedestrian feature obtained after learning the network H ^t (S ^t ,ω ^t ), and θ is the parameter of the metric learning network after the model is trained.

输入到融合多个源域的跨域行人重识别模型后得到的行人特征为：

Pedestrian samples in the t-th source domain

The pedestrian features obtained after inputting to the cross-domain person re-identification model that fuses multiple source domains are:

After inputting any pedestrian samples in the target domain T to x _i and x _j , merging the cross-domain pedestrian re-identification model of multiple source domains to obtain the distance of pedestrian features:

The cross-domain person re-identification model integrating multiple source domains is designed as follows:

The cross-domain person re-identification model integrating multiple source domains is mainly composed of m representation learning networks and one metric learning network. The representation learning network is mainly responsible for obtaining the pedestrian features input to the model, and the metric learning network is used to measure the distance of pedestrian features.

Training phase:

其中

为正样本对，

为负样本对，输入到对应的表示学习网络H^t(S^t,ω^t)中会得到对应的3个行人特征

和

再将这些行人特征输入到度量学习网络，通过度量学习网络后会得到新的3个行人特征

和

损失函数采用三元组损失函数(

其中d_n是负样本对之间的距离，d_p是正样本对之间的距离)，采用梯度下降策略进行融合多个源域的跨域行人重识别模型的优化训练。最后通过三元组损失函数对融合多个源域的跨域行人重识别模型进行参数学习。当m个源域数据集都输入模型后训练完成。As shown in Figure 2, the m representation learning networks all use the ResNet-50 network structure, and the metric learning network uses a 3-layer fully connected network structure. When training the model with the t-th source domain dataset, the pedestrian sample triples are

in

is a positive sample pair,

is a negative sample pair, and inputting it into the corresponding representation learning network H ^t (S ^t ,ω ^t ) will get the corresponding three pedestrian features

and

These pedestrian features are then input into the metric learning network, and three new pedestrian features will be obtained after passing through the metric learning network.

and

The loss function adopts the triple loss function (

where d _n is the distance between pairs of negative samples and d _p is the distance between pairs of positive samples), and the gradient descent strategy is used to optimize the training of cross-domain person re-identification models that fuse multiple source domains. Finally, the parameters of the cross-domain person re-identification model fused with multiple source domains are learned through the triple loss function. When the m source domain datasets are all input into the model, the training is completed.

For example, the training step (the 1st source domain dataset training (m total))

1) Input the first source domain triple pedestrian sample into the model;

2) After the representation learning network, a set of pedestrian features are obtained;

3) After the fusion feature network, a new set of pedestrian features is obtained;

4) The triple loss is obtained by calculating the distance between the positive and negative sample pairs;

5) Update the network weights through the gradient descent strategy to optimize the training model;

Test phase:

As shown in Figure 3, two pedestrian samples (x _i , x _j ) in the target domain are input, and m groups of pedestrian features are output after m representation learning networks, and new m groups of pedestrian features can be obtained through the metric learning network for each group of pedestrian features. , the distance between two pedestrian features can be obtained by calculation, and a threshold is set artificially. If the calculated distance is less than the set threshold, it is considered to be the same pedestrian, otherwise it is not the same pedestrian. Finally, through the voting mechanism, that is, the minority obeys the majority, it is determined whether the input picture is the same person.

The above method effectively improves the robustness of the training model, and the trained model is directly used in practical work, which effectively improves the performance of the model. The main reason for the performance improvement of the model is that the data from multiple source domains is integrated in the training process, and a metric network is introduced, which can better learn the feature representation of pedestrians, and the model can be further improved through the voting mechanism during the testing process. performance.

The principles and implementations of the present invention are described herein using specific examples. The descriptions of the above embodiments are only used to help understand the method and the core idea of the present invention; meanwhile, for those skilled in the art, according to the present invention There will be changes in the specific implementation and application scope. In conclusion, the contents of this specification should not be construed as limiting the present invention.

Claims (6)

1. A pedestrian re-identification method is characterized by comprising the following steps:

acquiring a pedestrian sample pair to be identified; the pedestrian sample pair to be identified comprises two pedestrian samples to be identified;

inputting the to-be-recognized pedestrian sample pair into a trained cross-domain pedestrian re-recognition model to obtain a pedestrian re-recognition result; the cross-domain pedestrian re-identification model comprises a plurality of representation learning networks and a measurement learning network, wherein the representation learning networks adopt a ResNet-50 network structure, and the measurement learning networks adopt a three-layer full-connection network structure;

the training method of the cross-domain pedestrian re-recognition model specifically comprises the following steps:

acquiring a plurality of groups of source domain data sets; each group of source domain data sets comprises a plurality of pedestrian samples to be trained, and each pedestrian sample to be trained corresponds to a pedestrian label; the number of the source domain data sets is the same as that of the representation learning networks;

training the cross-domain pedestrian re-recognition model by adopting a method of training a representation learning network and a metric learning network by adopting a group of source domain data sets to obtain a trained cross-domain pedestrian re-recognition model;

the method for training the cross-domain pedestrian re-recognition model by adopting a group of source domain data sets to train an expression learning network and a measurement learning network is used for training the cross-domain pedestrian re-recognition model to obtain the trained cross-domain pedestrian re-recognition model, and specifically comprises the following steps:

selecting a group of source domain data sets and a representation learning network; the source domain dataset comprises a positive sample pair and a negative sample pair, the positive sample pair comprising two pedestrian samples that are the same as a pedestrian, the negative sample pair comprising two pedestrian samples that are different from a pedestrian, and one pedestrian sample of the negative sample pair being the same as one pedestrian sample of the positive sample pair;

inputting the selected source domain data set into the selected representation learning network to obtain a plurality of first pedestrian characteristics;

inputting a plurality of first pedestrian features into the metric learning network to obtain a plurality of second pedestrian features;

calculating the pedestrian characteristic distance corresponding to the positive sample pair and the pedestrian characteristic distance corresponding to the negative sample pair according to the second pedestrian characteristic;

calculating a loss value according to the pedestrian characteristic distance corresponding to the positive sample pair and the pedestrian characteristic distance corresponding to the negative sample pair, and optimizing parameters in the cross-domain pedestrian re-identification model by adopting a gradient descent method with the minimum loss value as an optimization target;

and judging whether all the source domain data sets are completely selected, if so, obtaining a trained cross-domain pedestrian re-recognition model, otherwise, selecting one source domain data set from the unselected source domain data sets, selecting one representation learning network from the unselected representation learning networks, and returning to the step of inputting the selected source domain data set into the selected representation learning network to obtain a plurality of first pedestrian characteristics.

2. The method according to claim 1, wherein the step of inputting the to-be-recognized pedestrian sample pair into the trained cross-domain pedestrian re-recognition model to obtain a pedestrian re-recognition result specifically comprises:

inputting the to-be-recognized pedestrian sample pair into a trained cross-domain pedestrian re-recognition model to obtain a plurality of to-be-recognized pedestrian feature groups; each pedestrian feature group comprises two pedestrian features;

calculating the distance between two pedestrian features of each to-be-identified pedestrian feature group to obtain a plurality of to-be-identified pedestrian feature distances;

judging whether the number of the characteristic distances of the pedestrians to be identified, which is smaller than the preset distance, exceeds half of the total number of the characteristic distances of the pedestrians to be identified; if so, determining that the pedestrian samples in the to-be-identified pedestrian sample pair are the same; and if not, determining that the pedestrian samples in the to-be-identified pedestrian sample pair are different.

3. The pedestrian re-identification method according to claim 1,

the calculation formula of the pedestrian characteristic distance is as follows:

in the formula,

is composed of

And

the characteristic distance of the pedestrian is calculated,

for the ith pedestrian sample in the tth source domain data set,

for the jth pedestrian sample in the tth source domain data set,

is composed of

A corresponding second pedestrian characteristic is provided,

is composed of

Corresponding second pedestrian characteristics, C is a metric learning network;

the loss value is calculated as follows:

L_triple＝[d_p-d_n+α]₊

in the formula, L_tripleTo a loss value, d_pFor the negative sample to the corresponding pedestrian characteristic distance, d_nIs the positive sample to the corresponding pedestrian characteristic distance, alpha is d_pAnd d_nA predetermined interval of [ d ]_p-d_n+α]₊Is shown when d_n＜d_pAt + alpha, the loss value remains unchanged, when d_n≥d_pAt + α, the loss value is 0.

4. A pedestrian re-identification system, comprising:

the pedestrian re-identification module is used for acquiring a pedestrian sample pair to be identified; the pedestrian sample pair to be identified comprises two pedestrian samples to be identified; the pedestrian re-recognition method is also used for inputting the to-be-recognized pedestrian sample pair into a trained cross-domain pedestrian re-recognition model to obtain a pedestrian re-recognition result; the cross-domain pedestrian re-identification model comprises a plurality of representation learning networks and a measurement learning network, wherein the representation learning networks adopt a ResNet-50 network structure, and the measurement learning networks adopt a three-layer full-connection network structure;

the training module is used for training the cross-domain pedestrian re-recognition model to obtain a trained cross-domain pedestrian re-recognition model; the training method specifically comprises the following steps:

acquiring a plurality of groups of source domain data sets; each group of source domain data sets comprises a plurality of pedestrian samples to be trained, and each pedestrian sample to be trained corresponds to a pedestrian label; the number of the source domain data sets is the same as that of the representation learning networks;

training the cross-domain pedestrian re-recognition model by adopting a method of training a representation learning network and a metric learning network by adopting a group of source domain data sets to obtain a trained cross-domain pedestrian re-recognition model;

the training module specifically comprises:

a selecting unit for selecting a group of source domain data sets and a representation learning network; the source domain dataset comprises a positive sample pair and a negative sample pair, the positive sample pair comprising two pedestrian samples that are the same as a pedestrian, the negative sample pair comprising two pedestrian samples that are different from a pedestrian, and one pedestrian sample of the negative sample pair being the same as one pedestrian sample of the positive sample pair;

the first pedestrian characteristic generation unit is used for inputting the selected source domain data set into the selected representation learning network to obtain a plurality of first pedestrian characteristics;

the second pedestrian characteristic generating unit is used for inputting a plurality of first pedestrian characteristics into the metric learning network to obtain a plurality of second pedestrian characteristics;

the pedestrian characteristic distance calculation unit of the sample pair is used for calculating the pedestrian characteristic distance corresponding to the positive sample pair and the pedestrian characteristic distance corresponding to the negative sample pair according to the second pedestrian characteristic;

the model optimization unit is used for calculating a loss value according to the pedestrian characteristic distance corresponding to the positive sample pair and the pedestrian characteristic distance corresponding to the negative sample pair, and optimizing parameters in the cross-domain pedestrian re-identification model by adopting a gradient descent method with the minimum loss value as an optimization target;

and the second judgment unit is used for judging whether all the source domain data sets are completely selected, if so, obtaining a trained cross-domain pedestrian re-recognition model, otherwise, selecting one source domain data set from the unselected source domain data sets, selecting one representation learning network from the unselected representation learning networks, and then executing the first pedestrian feature generation unit.

5. The pedestrian re-identification system according to claim 4, wherein the pedestrian re-identification module specifically comprises:

the pedestrian feature group generation unit is used for inputting the to-be-identified pedestrian sample pair into a trained cross-domain pedestrian re-identification model to obtain a plurality of to-be-identified pedestrian feature groups; each pedestrian feature group comprises two pedestrian features;

the pedestrian feature distance calculation unit is used for calculating the distance between two pedestrian features of each pedestrian feature group to be recognized to obtain a plurality of pedestrian feature distances to be recognized;

the first judging unit is used for judging whether the number of the characteristic distances of the pedestrians to be identified, which is smaller than the preset distance, exceeds half of the total number of the characteristic distances of the pedestrians to be identified; if so, determining that the pedestrian samples in the to-be-identified pedestrian sample pair are the same; and if not, determining that the pedestrian samples in the to-be-identified pedestrian sample pair are different.

6. The pedestrian re-identification system according to claim 4,

the calculation formula of the pedestrian characteristic distance is as follows:

in the formula,

is composed of

And

the characteristic distance of the pedestrian is calculated,

for the ith pedestrian sample in the tth source domain data set,

for the jth pedestrian sample in the tth source domain data set,

is composed of

A corresponding second pedestrian characteristic is provided,

is composed of

Corresponding second pedestrian characteristics, C is a metric learning network;

the loss value is calculated as follows:

L_triple＝[d_p-d_n+α]₊

in the formula, L_tripleTo a loss value, d_pFor the negative sample to the corresponding pedestrian characteristic distance, d_nIs the positive sample to the corresponding pedestrian characteristic distance, alpha is d_pAnd d_nA predetermined interval of [ d ]_p-d_n+α]₊Is shown when d_n＜d_pAt + alpha, the loss value remains unchanged, when d_n≥d_pAt + α, the loss value is 0.

Images