WO2022147977A1

WO2022147977A1 - Vehicle re-identification method and system based on depth feature and sparse metric projection

Info

Publication number: WO2022147977A1
Application number: PCT/CN2021/103200
Authority: WO
Inventors: 刘凯
Original assignee: 山东交通学院
Priority date: 2021-01-05
Filing date: 2021-06-29
Publication date: 2022-07-14
Also published as: CN112699829B; CN112699829A

Abstract

Disclosed are a vehicle re-identification method and system based on a depth feature and a sparse metric projection. The method comprises: acquiring a target vehicle image; acquiring an image set to be re-identified; performing depth feature extraction on the target vehicle image and each image in said image set to obtain a depth feature of each image; calculating an adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image; calculating an adaptive sparse projection matrix corresponding to the depth feature of each image in said image set; calculating the distance between the depth feature of the target vehicle image and the depth feature of each image in said image set on the basis of the depth feature and the adaptive sparse projection matrix corresponding to the depth feature; repeating the step of a distance calculation module until the distances between the depth feature of the target image and depth features of all the images in said image set are calculated; and selecting an image corresponding to the minimum distance as a re-identification image of the target vehicle.

Description

Vehicle re-identification method and system based on deep feature and sparse metric projection

technical field

The present application relates to the technical field of computer vision, and in particular, to a method and system for vehicle re-identification based on depth feature and sparse metric projection.

Background technique

The statements in this section merely mention the background art related to the present application and do not necessarily constitute prior art.

At present, surveillance cameras are widely installed in cities, suburbs and highways, and a large number of vehicle surveillance images are collected and stored in real time. Cross-camera retrieval and continuous tracking of target vehicles appearing in different areas has become a reality. Traditional methods mainly use license plates. The recognition technology realizes the above functions, but in the real traffic environment, the vehicle has license plate occlusion, duplication, forgery, removal, etc. In this case, the license plate information is used for retrieval, and the target vehicle cannot be accurately located.

In recent years, with the continuous development of computer vision and multimedia technology, vehicle re-identification based on vehicle appearance information in surveillance video has attracted the attention of many researchers due to its important practical value, which involves the driving vehicle recognition technology. The task of vehicle re-identification is to find the images of the target vehicle captured by other cameras given the image of the target vehicle in a certain camera, so as to realize relay tracking across cameras.

However, due to the different positions of the cameras, there will be differences in illumination changes, viewing angles and resolutions. In addition, in complex monitoring scenarios, there are different degrees of occlusion between vehicles, which leads to intra-class differences (the same vehicle is generated from different viewing angles) self-difference) and inter-class similarity (different vehicles form inter-class similarity due to the same model), which makes the vehicle re-identification problem more difficult.

Existing supervised vehicle re-identification methods can be divided into feature learning-based methods and metric learning-based methods. The method based on feature learning expresses vehicle images by designing effective features to improve the matching accuracy of vehicle appearance features. This method has strong interpretability, but the recognition rate is low due to differences in vehicle appearance due to changes in illumination, perspective changes, and occlusions in the actual traffic monitoring environment. The method based on metric learning focuses on using the metric loss function to learn the similarity between vehicle images, and reduces the feature differences caused by illumination changes, viewing angle changes and occlusions through feature projection.

At present, the vehicle re-identification method based on metric learning mainly learns a specific feature projection matrix, so that the transformed features can eliminate the problems of intra-class differences and inter-class similarities caused by changes in perspective. Bai et al. in "Improving triplet-wise training of In the paper "convolutional neural network for vehicle re-identification", a group-sensitive triplet embedding method is designed to perform metric learning in an end-to-end manner. The idea proposed by Liu et al. in "Deep Relative Distance Learning: Tell the Difference between Similar Vehicles" has received a lot of attention from later generations. Deep Relative Distance Learning (DRDL), using the features learned from different branch tasks finally passed the The fully connected layer is integrated to obtain the final mapping feature. In view of the unstable characteristics of the training of the ternary loss function, this paper proposes to construct a positive and negative sample set and use the clustered cluster loss function (Coupled Clusters Loss) to replace the triple loss function as a measure. Learning can make vehicles of the same category more aggregated, and vehicles of different categories more discrete. However, in the process of metric learning, the re-identification model is very sensitive to the position of the image in the feature space. Now the vehicle re-identification metric learning method has the difference between the feature space of the training set and the test set and the generalization of the re-id model under other cameras. Ability not to do in-depth research.

SUMMARY OF THE INVENTION

In order to solve the deficiencies of the prior art, the present application provides a vehicle re-identification method and system based on depth feature and sparse metric projection; fully considering the influence of factors such as lighting conditions, camera parameters, viewing angle and occlusion on vehicle appearance characteristics, through the data space Collection of overcomplete dictionaries and metaprojection matrices. A feature sparse projection matrix is constructed adaptively for each vehicle image feature, which overcomes the diversity of vehicle image feature data distribution, improves the accuracy of vehicle re-identification, and enhances the generalization ability of the re-identification method.

In a first aspect, the present application provides a vehicle re-identification method based on depth feature and sparse metric projection;

Vehicle re-identification methods based on deep feature and sparse metric projection, including:

Obtain the target vehicle image; obtain the image set to be re-identified;

Perform depth feature extraction on the target vehicle image and each image in the set of images to be re-identified to obtain the depth feature of each image;

Calculate the adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image; at the same time, calculate the adaptive sparse projection matrix corresponding to the depth feature of each image in the image set to be re-identified;

Based on the depth feature and the adaptive sparse projection matrix corresponding to the depth feature, the distance between the depth feature of the target vehicle image and the depth feature of each image in the set of images to be re-identified is calculated;

Repeat the steps of the previous step until the distance between the depth feature of the target image and the depth features of all images in the set of images to be re-identified is calculated; the image corresponding to the minimum distance is selected as the re-identified image of the target vehicle.

In a second aspect, the present application provides a vehicle re-identification system based on depth feature and sparse metric projection;

A vehicle re-identification system based on deep feature and sparse metric projection, including:

an acquisition module, which is configured to: acquire an image of the target vehicle; acquire a set of images to be re-identified;

a feature extraction module, which is configured to: perform depth feature extraction on each image in the target vehicle image and the image set to be re-identified to obtain the depth feature of each image;

The projection matrix calculation module is configured to: calculate the adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image; at the same time, calculate the adaptive sparse projection matrix corresponding to the depth feature of each image in the set of images to be re-identified ;

a distance calculation module, which is configured to: based on the depth feature and the adaptive sparse projection matrix corresponding to the depth feature, calculate the distance between the depth feature of the target vehicle image and the depth feature of each image in the set of images to be re-identified;

The output module is configured to: repeat the steps of the distance calculation module until the distance between the depth feature of the target image and the depth features of all images in the set of images to be re-identified is calculated; select the image corresponding to the minimum distance as the target vehicle the re-identified image.

In a third aspect, the present application also provides an electronic device, comprising: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and one or more of the above The computer program is stored in the memory, and when the electronic device runs, the processor executes one or more computer programs stored in the memory, so that the electronic device performs the method described in the first aspect above.

In a fourth aspect, the present application further provides a computer-readable storage medium for storing computer instructions, and when the computer instructions are executed by a processor, the method described in the first aspect is completed.

In a fifth aspect, the present application also provides a computer program (product), including a computer program, which when run on one or more processors, is used to implement the method of any one of the foregoing first aspects.

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

In the actual traffic monitoring scene, the vehicle image imaging process is easily affected by the shooting environment (including lighting conditions, camera parameters, shooting angle and external occlusion and many other factors), and the corresponding features of each vehicle image have a unique data distribution. Vehicle re-identification methods based on traditional metric learning cannot cope with the uniqueness of this feature distribution, resulting in low accuracy of feature distance calculation and vehicle re-identification. Based on this, the present invention proposes a vehicle re-identification method based on deep feature and sparse metric projection, which introduces an adaptive strategy into the traditional metric projection matrix learning process, and learns for each image feature by constructing a data space overcomplete dictionary and a meta-projection matrix. The adaptive sparse projection matrix ensures that all image features are in the same data space after projection. On the one hand, the model maintains good nearest neighbor calculation performance under various data distributions; on the other hand, the distance metric can be better adapted to different types. Practical application scenarios to improve the generalization ability of the system. The experimental results on the vehicle re-identification task confirm the effectiveness of the method proposed in the present invention.

Advantages of additional aspects of the present application will be set forth in part in, and in part will become apparent from, the following description, or may be learned by practice of the present application.

Description of drawings

The accompanying drawings that form a part of the present application are used to provide further understanding of the present application, and the schematic embodiments and descriptions of the present application are used to explain the present application and do not constitute improper limitations on the present application.

1 is a flowchart of an embodiment of the application;

2 is a flowchart of a data space adaptive sparse metric projection learning algorithm according to an embodiment of the present application;

Detailed ways

It should be noted that the following detailed description is exemplary and intended to provide further explanation of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

It should be noted that the terminology used herein is for the purpose of describing specific embodiments only, and is not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly dictates otherwise, the singular is intended to include the plural as well, furthermore, it is to be understood that the terms "including" and "having" and any conjugations thereof are intended to cover the non-exclusive A process, method, system, product or device comprising, for example, a series of steps or units is not necessarily limited to those steps or units expressly listed, but may include those steps or units not expressly listed or for such processes, methods, Other steps or units inherent to the product or equipment.

The embodiments in this application and the features in the embodiments may be combined with each other without conflict.

Example 1

This embodiment provides a vehicle re-identification method based on depth feature and sparse metric projection;

S101: Obtain an image of a target vehicle; obtain a set of images to be re-identified;

S102: Perform depth feature extraction on the target vehicle image and each image in the set of images to be re-identified to obtain the depth feature of each image;

S103: Calculate the adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image;

At the same time, calculating the adaptive sparse projection matrix corresponding to the depth feature of each image in the set of images to be re-identified;

S104: Based on the depth feature and the adaptive sparse projection matrix corresponding to the depth feature, calculate the distance between the depth feature of the target vehicle image and the depth feature of each image in the image set to be re-identified;

S105: Repeat S104 until the distance between the depth feature of the target image and the depth features of all images in the set of images to be re-identified is calculated; the image corresponding to the minimum distance is selected as the re-identified image of the target vehicle.

As one or more embodiments, the S102: perform depth feature extraction on the target vehicle image and each image in the image set to be re-identified to obtain the depth feature of each image; specifically including:

For each image in the target vehicle image and the image set to be re-identified, the improved VGG-19 network is used to extract the depth feature, and the depth feature of each image is obtained;

In the improved VGG-19 network, in order to remove the last two fully connected layers of the VGG-19 network, only the first 16 convolutional layers and the first fully connected layer are retained.

The improved VGG-19 network is pre-trained with the ImageNet dataset.

As one or more embodiments, the S103: the calculation step of calculating the adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image corresponds to the calculation step of the depth feature of each image in the image set to be re-identified The calculation steps of the adaptive sparse projection matrix are consistent.

As one or more embodiments, the S103: Calculate the adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image; specifically include:

S1031: Calculate the sparse coefficient corresponding to the depth feature of the target vehicle image according to the overcomplete dictionary;

S1032: Taking the sparse coefficient corresponding to the depth feature of the target vehicle image as a weight, and performing a weighted summation on the meta-projection matrix to obtain an adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image.

Further, the overcomplete dictionary is obtained using training data.

Further, for the overcomplete dictionary, the obtaining step includes:

S10311: Initialize the overcomplete dictionary D as the K cluster center in the training data space; initialize each element of the sparse coefficient matrix; the training data includes: a known target vehicle image and a known target vehicle re-identification image;

S10312: Calculate feature sparse coding loss function;

S10313: According to the feature sparse coding loss function, adopt an iterative training strategy, first fix the overcomplete dictionary D, use the gradient descent method to update the sparse coefficient matrix, then fix the sparse coefficient matrix, and use the gradient descent method to update the overcomplete dictionary D.

Exemplarily, the specific steps for calculating the overcomplete dictionary and sparse coefficients include:

(11) Initialize the overcomplete dictionary D as K cluster centers in the training data space, and initialize each element of the sparse coefficient matrix to 1/K.

(12) Calculate the feature sparse coding loss function, the loss function is as described in formula (1):

Among them: F is the characteristic matrix of the training data set, α is the sparse coefficient matrix, and λ is the balance coefficient.

(13) According to formula (1), an iterative training strategy is adopted. First, the overcomplete dictionary D is fixed, and the gradient descent method is used to update the sparse coefficient matrix α, and then the sparse coefficient matrix α is fixed, and the gradient descent method is used to update the overcomplete dictionary D.

Further, the meta-projection matrix is also obtained using training data.

Further, the step of obtaining the meta-projection matrix includes:

S10321: Using a joint training strategy, splicing the element projection matrix to construct a composite projection matrix;

S10322: Calculate the loss function of the composite projection matrix;

S10323: According to the loss function of the composite projection matrix, a gradient descent strategy is used to calculate the gradient value of the composite projection matrix, and update the gradient value of the composite projection matrix, that is, to obtain each element projection matrix.

Exemplarily, calculating a set of meta-projection matrices, the method includes the following steps:

Using a joint training strategy to define a composite projection matrix

Compound eigenvectors

and

(21) Calculate the composite projection matrix loss function; the loss function is as described in formula (2):

in:

If the sample

and sample

belong to the same vehicle; then _sil = 1, otherwise, _sil = 0; if

Yes

one of the k-nearest neighbors of , while

and

belong to the same vehicle, then η _ij =1; otherwise, η _ij =0.

(22) According to formula (2), the gradient descent strategy is adopted to calculate

The gradient of , we get formula (3)

Where: σ _β (·)=(1+e ^-βx ) ^-1 ,

_ηij and _sil correspond to the parameters of formula (2);

(23) On the basis of formula (3),

To update, the update rule is:

Where: λ is the step size of iterative update.

As one or more embodiments, the S104: based on the depth feature and the adaptive sparse projection matrix corresponding to the depth feature, calculate the difference between the depth feature of the target vehicle image and the depth feature of each image in the image set to be re-identified distance; specific steps include:

Multiply the depth feature of the target vehicle image and the adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image to obtain the first product;

Multiplying the depth feature of a certain image in the set of images to be re-identified with the adaptive sparse projection matrix corresponding to the depth feature of the image in the set of images to be re-identified to obtain a second product;

Calculate the distance between the first product and the second product;

The distance between the first product and the second product is the distance between the depth feature of the target vehicle image and the depth feature of each image in the image set to be re-identified.

Since the over-complete dictionary and meta-projection matrix are both obtained through the training data set, S101-S105 also includes a training phase and a testing phase; the details are as follows: the method collects images of the same vehicle under different cameras, and the vehicle The images are divided into training image sets and test image sets, and feature extraction is performed on the images to form training data sets and test data sets respectively. Above, use the data space adaptive sparse metric projection matrix to transform image features, and perform distance calculation based on the transformed image features to complete vehicle re-identification.

As shown in Figure 1, the training phase and the testing phase include the following steps:

Step 1): Collect images of the same vehicle under different cameras;

Step 2): Divide the vehicle image into a training image set and a test image set, perform feature extraction on the image, and form a training data set and a test data set respectively;

Step 3): on the training data set, learn the calculation method of the data space adaptive sparse metric projection matrix;

Step 4): On the test data set, use the data space adaptive sparse metric projection matrix to perform image feature transformation, and perform distance calculation based on the transformed image features to complete vehicle re-identification.

Described step 1): for M vehicles, collect images of each vehicle under camera A and camera B to form image sets X and Y respectively.

Described step 2): randomly select N vehicles from M vehicles, extract images belonging to the N vehicles in the image sets X and Y to form a training image set, and images belonging to the remaining M-N vehicles form a test image set.

Step 2) Randomly select N vehicles from M vehicles, extract images belonging to the N vehicles in the image set X and Y to form a training image set, a total of 2*N images, and the images of the remaining M-N vehicles form a test image set , a total of 2*(M-N) images. Perform depth feature extraction on all images in the training image set to form a training data set; perform depth feature extraction on all images in the test image set to form a test data set.

Described step 3) is carried out on the training data set, including:

The projection matrix that defines the feature sample x is:

in

is the sample space overcomplete dictionary

the corresponding set of meta-projection matrices,

is the sparse coefficient.

Described step 4) is carried out on the test data set, including:

4.1) For any image feature x _test in the test data set, calculate its adaptive sparse projection matrix

The method includes the following steps:

4.1.1) Fix the complete dictionary D, and calculate the sparse coefficient of x _test according to steps (11), (12) and (13);

4.1.2) Calculate the adaptive projection matrix of x _test according to formula (5)

4.2) Calculate the distance between the target image feature x _test and the first image feature y ₁ to be re-identified, as shown in formula (6):

4.3) Repeat step 4.2) until the distance calculation between x _test and all the image features to be re-identified in the test data set is completed, and it is considered that the image corresponding to the minimum distance and x _test belong to the same vehicle.

Described step 1) comprises:

1.1) Given a vehicle image, adjust the size to 224*224 pixels;

1.2) The image obtained in step 1.1) is sent to the VGG-19 network for feature extraction to obtain a 4096-dimensional feature vector;

The 16 convolutional layers and the first fully connected layer of the VGG-19 network are used as the feature extraction part, and the last two fully connected layers of the VGG-19 are removed.

After the above 1.2), PCA is further used to perform dimension reduction operation on the feature vector, and a feature vector of 127 dimensions is finally obtained under the condition of retaining 80% of the feature values.

In this embodiment, the specific implementation of the above feature extraction method is described as follows:

Construct a VGG-19 network pre-trained based on the ImageNet dataset, remove the last 2 fully connected layers of the VGG-19 network, and retain the 16 convolutional layers and the first fully connected layer of the VGG-19 network as a deep feature extraction network .

Normalize the vehicle image size to 224*224 pixels;

The image is sent to the deep feature extraction network for feature extraction, and a 4096-dimensional feature vector is obtained;

In order to reduce the number of model parameters and improve the generalization ability of the model, this method uses PCA to perform dimensionality reduction operations on the original features, and finally obtains a 127-dimensional feature vector while retaining 80% of the eigenvalues.

Step 3) On the training data set, learn the data space adaptive sparse projection matrix calculation method.

The data space adaptation refers to learning an adaptive projection matrix for the image feature vectors, so that all image feature vectors are projected in the same data space, thereby ensuring the effectiveness of the nearest neighbor comparison. In practice, in order to improve the efficiency of the algorithm, the approximate learning method based on sparse coding constructs an over-complete dictionary and a meta-projection matrix in the data space, uses the over-complete dictionary to sparsely encode the feature data, and compares the coding coefficients with the meta-projection matrix. Combined, a data space adaptive sparse projection matrix is constructed.

As shown in Figure 2, the flow chart of the learning data space adaptive sparse projection matrix calculation method proposed in the embodiment, the specific learning process is as follows:

3.1) Initialize overcomplete dictionary D and sparse coefficient matrix α

3.2) Use formula (1) to calculate the feature sparse coding loss function;

3.3) Use the iterative gradient optimization strategy to iteratively update the overcomplete dictionary D and sparse coefficient matrix α to complete the optimization of the overcomplete dictionary D and sparse coefficient matrix α

3.4) Use the updated overcomplete dictionary D and the sparse coefficient matrix α to calculate the loss function according to formula (1). If ΔΩ>ε ₁ , go to step 3.3, otherwise it is judged to be converged, and the corresponding D and α are output.

3.5) Calculate the composite projection matrix loss function using formula (2);

3.6) Using formula (3) and formula (4), using the gradient optimization strategy, update

3.7) Using the updated

Calculate the loss function according to formula (2). If ΔΨ>ε ₂ , go to step 3.6, otherwise it is determined to be converged, and the corresponding output is

Step 4): On the basis of the data space adaptive sparse projection matrix calculation method learned in step 3), vehicle re-identification is performed on the test data set; the specific implementation method is as follows:

4.1) In the test data set, the image characteristics of the first vehicle under camera A and the characteristics of all vehicles (M-N vehicles in total) under camera B are calculated according to formula (6) to obtain M-N distance calculation results;

4.2) The M-N distance calculation results obtained in step 4.1) are sorted from small to large, and the camera B image corresponding to the distance calculation result in the first place is the image that belongs to the same vehicle as the camera A image provided by this method;

4.3) Repeat steps 4.1) and 4.2) to complete the distance calculation and vehicle consistency determination between all image features under camera A and all image features under camera B.

The present application provides a vehicle re-identification method based on depth feature and sparse metric projection. The method collects images of the same vehicle under different cameras, divides the vehicle images into a training image set and a test image set, and characterizes the images. Extract, respectively form a training data set and a test data set. On the training data set, learn the calculation method of the data space adaptive sparse metric projection matrix, and on the test data set, use the data space adaptive sparse metric projection matrix to perform image feature transformation , and calculate the distance based on the transformed image features to complete the vehicle re-identification.

This application considers that in the traffic monitoring network, the shooting environment (including factors such as illumination, camera angle, camera parameters, occlusion, etc.) of each vehicle image is different, resulting in unique data distribution of corresponding features. The data space is overcomplete with dictionary and meta-projection matrix, and an adaptive sparse projection matrix is learned for each image feature, so that the projected feature samples are all in the same feature space, thus ensuring the effectiveness of the nearest neighbor comparison.

The present application belongs to a method based on metric learning, and the goal is to project all feature vectors into a unified feature space, so that the features of the same car are closer, and the features of different cars are farther away.

Embodiment 2

This embodiment provides a vehicle re-identification system based on depth feature and sparse metric projection;

It should be noted here that the above-mentioned acquisition module, feature extraction module, projection matrix calculation module, distance calculation module and output module correspond to steps S101 to S105 in the first embodiment, and the examples and The application scenarios are the same, but are not limited to the content disclosed in the first embodiment. It should be noted that the above modules can be executed in a computer system such as a set of computer-executable instructions as part of the system.

The description of each embodiment in the foregoing embodiments has its own emphasis. For the part that is not described in detail in a certain embodiment, reference may be made to the relevant description of other embodiments.

The proposed system can be implemented in other ways. For example, the system embodiments described above are only illustrative. For example, the division of the above modules is only a logical function division. In actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into other A system, or some feature, can be ignored, or not implemented.

Embodiment 3

This embodiment also provides an electronic device, including: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are Stored in the memory, when the electronic device runs, the processor executes one or more computer programs stored in the memory, so that the electronic device executes the method described in the first embodiment.

It should be understood that, in this embodiment, the processor may be a central processing unit (CPU), and the processor may also be other general-purpose processors, digital signal processors, DSPs, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices. , discrete gate or transistor logic devices, discrete hardware components, etc. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The memory may include read-only memory and random access memory and provide instructions and data to the processor, and a portion of the memory may also include non-volatile random access memory. For example, the memory may also store device type information.

In the implementation process, each step of the above-mentioned method can be completed by a hardware integrated logic circuit in a processor or an instruction in the form of software.

The method in the first embodiment may be directly embodied as being executed by a hardware processor, or executed by a combination of hardware and software modules in the processor. The software modules may be located in random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, registers and other storage media mature in the art. The storage medium is located in the memory, and the processor reads the information in the memory, and completes the steps of the above method in combination with its hardware. To avoid repetition, detailed description is omitted here.

Those skilled in the art can realize that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of this application.

Embodiment 4

This embodiment also provides a computer-readable storage medium for storing computer instructions, and when the computer instructions are executed by a processor, the method described in the first embodiment is completed.

The above descriptions are only preferred embodiments of the present application, and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this application shall be included within the protection scope of this application.

Claims

The vehicle re-identification method based on deep feature and sparse metric projection is characterized by including:

Obtain the target vehicle image; obtain the image set to be re-identified;

Perform depth feature extraction on the target vehicle image and each image in the set of images to be re-identified to obtain the depth feature of each image;

Calculate the adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image; at the same time, calculate the adaptive sparse projection matrix corresponding to the depth feature of each image in the image set to be re-identified;

Based on the depth feature and the adaptive sparse projection matrix corresponding to the depth feature, the distance between the depth feature of the target vehicle image and the depth feature of each image in the set of images to be re-identified is calculated;

Repeat the steps of the previous step until the distance between the depth feature of the target image and the depth features of all images in the set of images to be re-identified is calculated; the image corresponding to the minimum distance is selected as the re-identified image of the target vehicle.
The vehicle re-identification method based on depth feature and sparse metric projection as claimed in claim 1, wherein depth feature extraction is performed on each image in the target vehicle image and the image set to be re-identified to obtain each image. Depth features of an image; specifically include:

For each image in the target vehicle image and the image set to be re-identified, the improved VGG-19 network is used to extract the depth feature, and the depth feature of each image is obtained;

In the improved VGG-19 network, in order to remove the last two fully connected layers of the VGG-19 network, only the first 16 convolutional layers and the first fully connected layer are retained.
The vehicle re-identification method based on depth feature and sparse metric projection as claimed in claim 1, wherein the adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image is calculated; specifically, it includes:

Calculate the sparse coefficient corresponding to the depth feature of the target vehicle image according to the overcomplete dictionary;

The sparse coefficients corresponding to the depth features of the target vehicle image are regarded as weights, and the meta-projection matrix is weighted and summed to obtain the adaptive sparse projection matrix corresponding to the depth features of the target vehicle image.
The vehicle re-identification method based on depth feature and sparse metric projection as claimed in claim 3, wherein, in the overcomplete dictionary, the obtaining step comprises:

Initialize the overcomplete dictionary D as the K cluster center in the training data space; initialize each element of the sparse coefficient matrix; the training data includes: known target vehicle images and re-identified images of known target vehicles;

Calculate the feature sparse coding loss function;

According to the feature sparse coding loss function, an iterative training strategy is adopted. First, the overcomplete dictionary D is fixed, and the sparse coefficient matrix is updated by the gradient descent method, and then the sparse coefficient matrix is fixed, and the overcomplete dictionary D is updated by the gradient descent method.
The vehicle re-identification method based on depth feature and sparse metric projection as claimed in claim 3, wherein, in the element projection matrix, the obtaining step comprises:

Using the joint training strategy, splicing the element projection matrix to construct the composite projection matrix;

Calculate the loss function of the composite projection matrix;

According to the loss function of the composite projection matrix, using the gradient descent strategy, the gradient value of the composite projection matrix is calculated, and the gradient value of the composite projection matrix is updated, that is, each element projection matrix is obtained.
The vehicle re-identification method based on depth feature and sparse metric projection according to claim 1, wherein the depth feature of the target vehicle image and the to-be-re-identified vehicle image are calculated based on the depth feature and the adaptive sparse projection matrix corresponding to the depth feature. The distance between the depth features of each image in the image set; the specific steps include:

Multiply the depth feature of the target vehicle image and the adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image to obtain the first product;

Multiplying the depth feature of a certain image in the set of images to be re-identified with the adaptive sparse projection matrix corresponding to the depth feature of the image in the set of images to be re-identified to obtain a second product;

Calculate the distance between the first product and the second product;

The distance between the first product and the second product is the distance between the depth feature of the target vehicle image and the depth feature of each image in the image set to be re-identified.
The method for vehicle re-identification based on deep feature and sparse metric projection according to claim 2, wherein the improved VGG-19 network is pre-trained with ImageNet dataset.
The vehicle re-identification system based on deep feature and sparse metric projection is characterized by including:

an acquisition module, which is configured to: acquire an image of the target vehicle; acquire a set of images to be re-identified;

a feature extraction module, which is configured to: perform depth feature extraction on each image in the target vehicle image and the image set to be re-identified to obtain the depth feature of each image;

The projection matrix calculation module is configured to: calculate the adaptive sparse projection matrix corresponding to the depth feature of the target vehicle image; at the same time, calculate the adaptive sparse projection matrix corresponding to the depth feature of each image in the set of images to be re-identified ;

a distance calculation module, which is configured to: based on the depth feature and the adaptive sparse projection matrix corresponding to the depth feature, calculate the distance between the depth feature of the target vehicle image and the depth feature of each image in the image set to be re-identified;

The output module is configured to: repeat the steps of the distance calculation module until the distance between the depth feature of the target image and the depth features of all images in the set of images to be re-identified is calculated; select the image corresponding to the minimum distance as the target vehicle the re-identified image.
An electronic device is characterized by comprising: one or more processors, one or more memories, and one or more computer programs; wherein the processor is connected to the memory, and the one or more computer programs are stored in In the memory, when the electronic device is running, the processor executes one or more computer programs stored in the memory, so that the electronic device executes the method of any one of the above claims 1-7.
A computer-readable storage medium, characterized in that it is used for storing computer instructions, and when the computer instructions are executed by a processor, the method according to any one of claims 1-7 is completed.