Method and system for vehicle re-identification based on time-series network multi-camera

Through the multi-camera vehicle re-identification method based on timing network, combined with memory network and query model, the static, dynamic and scene information of the vehicle is extracted, and the vehicle re-identification challenge in multi-camera scenarios is solved, improving the recognition accuracy and robustness of the recognition sex.

CN118736385BActive Publication Date: 2024-11-15HEFEI ARTIFICIAL INTELLIGENCE & BIG DATA RES INST CO LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411226171.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-03
Publication Date
2024-11-15
Estimated Expiration
2044-09-03

AI Technical Summary

Technical Problem

The existing vehicle re-identification technology has problems such as camera viewing angle, lighting changes, occlusion and vehicle appearance similarity in multi-camera scenarios, and there are limitations to relying solely on appearance feature recognition.

Method used

Using a multi-camera vehicle re-identification method based on a timing network, the image data of the vehicle under a multi-camera is obtained, preprocessing and random transformation are performed, and the memory network model and query model are constructed, and the vehicle is extracted by combining the timing convolution network and the ResNet18 network. The static, dynamic and scene information of the three-tuple loss function is used to optimize the model.

Benefits of technology

It improves the accuracy and robustness of vehicle re-identification, can better handle different camera viewing angles and lighting changes, enhances the recognition ability of vehicle appearance similarity, and reduces resource consumption.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118736385B_ABST
    Figure CN118736385B_ABST
Patent Text Reader

Abstract

The invention discloses a method and system for vehicle re-identification based on a time series network with multiple cameras, relates to the technical field of vehicle re-identification, and solves the technical problem that the prior art has certain limitations in a multi-camera scenario because it simply relies on the appearance features of a vehicle for re-identification. The invention obtains image data of a vehicle under several cameras, pre-processes the image data to obtain an image data set, increases the diversity of the data set, and solves the problems of camera viewing angle, illumination change, occlusion, and high vehicle appearance similarity that still exist in the prior art when processing a multi-camera scenario. A vehicle re-identification model is formed based on a memory network model and an inquiry model. The vehicle re-identification model is trained based on the image data set to obtain a target vehicle re-identification model. The invention solves the technical problem that the prior art has certain limitations in a multi-camera scenario because it simply relies on the appearance features of a vehicle for re-identification.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of vehicle re-identification, and specifically relates to a method and system for vehicle re-identification based on a time-series network with multiple cameras. Background Art

[0002] Vehicle re-identification is an important technology in intelligent transportation systems, which aims to determine whether the vehicles captured by different cameras are the same vehicle by analyzing the image or video data of the vehicle.

[0003] Existing vehicle re-identification technologies mainly rely on deep learning models to map vehicle images to a high-dimensional feature space and compare them through the similarity between feature vectors; and draw on face recognition technology to extract vehicle features through convolutional neural networks and map these features to a high-dimensional feature vector for vehicle re-identification; however, existing technologies still face many challenges when dealing with multi-camera scenarios, such as camera viewing angles, lighting changes, occlusions, and high vehicle appearance similarity; at the same time, re-identification based solely on vehicle appearance features has certain limitations in multi-camera scenarios.

[0004] Therefore, the present invention proposes a method and system for vehicle re-identification based on a temporal network with multiple cameras to solve the above problems. Summary of the invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes a method and system for vehicle re-identification based on a temporal network multi-camera, which are used to solve the many challenges that still exist in the prior art when processing multi-camera scenes, such as camera viewing angle, lighting changes, occlusion, and high similarity of vehicle appearance; at the same time, the technical problem that re-identification relying solely on the appearance features of the vehicle has certain limitations in multi-camera scenarios.

[0006] To achieve the above object, a first aspect of the present invention provides a method for vehicle re-identification based on a time series network with multiple cameras, comprising:

[0007] Step 1: Obtain image data of the vehicle under several cameras;

[0008] Step 2: Obtain an image data set by preprocessing the image data;

[0009] Step 3: Construct a vehicle re-identification model based on the memory network model and the query model;

[0010] Step 4: Train the vehicle re-identification model based on the image dataset to obtain the target vehicle re-identification model;

[0011] Step 5: Re-identify the vehicle based on the target vehicle re-identification model.

[0012] Preferably, the step of obtaining image data of the vehicle under a plurality of cameras includes:

[0013] The image data of the vehicle is collected by several cameras installed in the target area; wherein the image collected by each camera includes: a scene picture and several vehicle pictures; the target area refers to the area where the vehicle needs to be re-identified.

[0014] Preferably, the preprocessing of the image data includes:

[0015] An image data set is obtained by performing random transformation processing on the image data; wherein the random transformation processing includes: flipping and rotating, scaling and translating, cropping, color changing or adding random noise; wherein the data in the image data set are in units of bars, and each piece of data includes: several scene pictures and continuous frames of the same vehicle under different cameras.

[0016] It should be noted that each piece of data contains images captured by multiple cameras; continuous frames refer to a sequence of images taken continuously within a certain time interval.

[0017] Preferably, the vehicle re-identification model composed of the memory network model and the query model includes:

[0018] Build a memory network model based on the temporal convolutional network, and build a query model.

[0019] Preferably, the step of constructing a memory network model based on a temporal convolutional network includes:

[0020] Three causal convolutional layers and one dilated convolutional layer form a module, and several modules are stacked together to form a memory network model;

[0021] The image dataset is labeled as x=[x1,x2,x3,…,xT]; where T is the total number of data in the image dataset;

[0022] By inputting x into the memory network model, applying causal convolution and dilated convolution to obtain the output sequence y; where y=[y1,y2,y3,…,yT], and y is a 512-dimensional vector;

[0023] The whole process is specifically expressed as:

[0024] Causal Convolutional Layer:

[0025] ;

[0026] Where y[t] is the value of the output sequence at time step t, w[i] is the weight of the i-th position in the convolution kernel, and x[ti] is the value of the input sequence at time step ti;

[0027] Expanded convolutional layer:

[0028] ;

[0029] Among them, y[t] is the value of the output sequence at time step t, w[i] is the weight of the i-th position in the convolution kernel, x[td×i] is the value of the input sequence at time step td×i, and d is the expansion rate, which controls the spacing of elements in the convolution kernel.

[0030] It should be noted that the output sequence y corresponds one-to-one with the input x. If the input is x1, the output is y1;

[0031] The memory network model adopts a temporal convolutional network, which combines the advantages of convolutional neural networks and recurrent neural networks. It can process sequence data and maintain long-term dependencies.

[0032] Preferably, the constructing of the inquiry model includes:

[0033] The query model uses the Resnet18 network to extract the features of the target vehicle image; Resnet18 is composed of multiple residual modules stacked together; the target vehicle image refers to the vehicle image that needs to be re-identified;

[0034] The structure of the residual module is as follows:

[0035] Convolutional layer 1: F(X)=σ(W1×X+b1); where W1 is the weight of the first convolutional layer, b1 is the bias term, and σ is the ReLU activation function;

[0036] Convolutional layer 2: F(F(x))=σ(W2×F(x)+b2); where F(x) is the output of the first convolution layer, W2 is the weight of the second convolution layer, and b2 is the bias term;

[0037] Identity mapping: Y=F(F(x))+W id ×X; where W id is a 1×1 convolution kernel for dimension conversion, and Y is the output 512-dimensional vector.

[0038] It should be noted that Resnet18 is composed of multiple stacked residual modules. The overall structure is relatively shallow and can effectively extract target features with less resource consumption. The residual module can train deeper networks and is easier to optimize the model, which significantly reduces the number of parameters while increasing the depth.

[0039] Preferably, the training of the vehicle re-identification model based on the image dataset includes:

[0040] S1: Input each data in the image dataset and the target vehicle image into the vehicle re-identification model;

[0041] S2: extracting feature data from each piece of data through a memory network model, and encoding the feature data to output a feature vector; wherein the feature data includes: static features, dynamic features and scene information of the vehicle;

[0042] S3: extracting target feature data of the target vehicle image through the query model, and converting the target feature data into a target feature vector of the same dimension as the output of the memory network;

[0043] S4: judging whether the feature vector originates from the target vehicle based on the Euclidean distance;

[0044] S5: Integrate each piece of data in the image data set and the target vehicle image as standard input data, and integrate the image corresponding to the feature vector from the target vehicle as standard output data;

[0045] S6: Train a vehicle re-identification model based on the standard input data and the standard output data, and optimize the vehicle re-identification model through a triplet loss function to obtain a target vehicle re-identification model.

[0046] It should be noted that the reason why the memory network model can obtain the dynamic, static and scene information of the vehicle is mainly because it uses a temporal convolutional network (TCN), which combines the advantages of convolutional neural networks (CNN) and recurrent neural networks (RNN), and can effectively process sequence data and maintain long-term dependencies; by using causal convolution and dilated convolution, TCN can process dynamic features in sequence data; because the model receives continuous frames from different cameras, even on static images, it can obtain more comprehensive static features of the vehicle by comparing vehicle features under different perspectives; for scene information, each piece of data contains several scene pictures, which helps the model understand the static features of the vehicle's environment; among them, dynamic features include: vehicle speed, direction and acceleration, etc.; static features include: vehicle brand, color, model, logo and damage marks, etc.; scene information includes: road conditions, weather conditions, surrounding buildings and signs, etc.

[0047] The triplet loss function can optimize the model by maximizing the similarity between positive examples (samples belonging to the same category) and minimizing the similarity between negative examples (samples not belonging to the same category). Using the triplet loss function to optimize the model during training can enhance the learning ability of the model, that is, shorten the mapping distance of the same vehicle and lengthen the mapping distance of different vehicles.

[0048] Preferably, judging whether the feature vector originates from the target vehicle based on the Euclidean distance includes:

[0049] S1: The Euclidean distance between the feature vector and the target feature vector is calculated based on the Euclidean distance calculation formula;

[0050] S2: Determine whether the Euclidean distance is less than a preset distance threshold; if yes, it is determined that the feature vector originates from the target vehicle; if no, it is determined that the feature vector does not originate from the target vehicle; wherein the preset distance threshold is set according to the actual vehicle re-identification scenario.

[0051] A second aspect of the present invention provides a vehicle re-identification system based on a time-series network with multiple cameras, comprising: a data acquisition module, a data analysis module and a vehicle re-identification model building module;

[0052] Data acquisition module: used to obtain image data of the vehicle under several cameras;

[0053] Data analysis module: used to pre-process image data to obtain image data sets;

[0054] Vehicle re-identification model building module: a vehicle re-identification model is constructed based on a memory network model and a query model; and,

[0055] A vehicle re-identification model is trained based on the image data set to obtain a target vehicle re-identification model; and a vehicle re-identification is performed on the vehicle based on the target vehicle re-identification model.

[0056] Preferably, the data analysis module is in communication and / or electrically connected with the data acquisition module and the vehicle re-identification model building module respectively.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. The existing technology still faces many challenges when processing multi-camera scenes, such as camera perspective, lighting changes, occlusion, and high vehicle appearance similarity. The present invention obtains image data of the vehicle under several cameras and obtains an image data set by preprocessing the image data to solve the above problems. The training images are randomly transformed, such as flipping, rotating, scaling, translating, cropping, color changing, adding random noise, etc., to increase the diversity of the data set, so that the model can better cope with problems such as camera perspective changes, lighting changes, and occlusion.

[0059] 2. The prior art relies solely on the appearance features of the vehicle for re-identification, which has certain limitations in multi-camera scenarios. The present invention forms a vehicle re-identification model based on a memory network model and an inquiry model; trains the vehicle re-identification model based on an image data set to obtain a target vehicle re-identification model; and performs vehicle re-identification on the vehicle based on the target vehicle re-identification model, thereby solving the above problems. The memory network model adopts a temporal convolutional network (TCN), which combines the advantages of a convolutional neural network (CNN) and a recurrent neural network (RNN), and is capable of processing sequence data and maintaining long-term dependencies, thereby improving recognition accuracy by capturing the dynamic features, static features, and scene information of the vehicle. A lightweight ResNet18 network is used to extract target image features. This structure is simple and efficient, and can effectively extract target features, thereby improving the overall efficiency of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0061] Figure 1 A schematic diagram of the method steps of an embodiment of the present invention;

[0062] Figure 2 A schematic diagram of vehicle re-identification model training according to an embodiment of the present invention;

[0063] Figure 3 Schematic diagram of system modules according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] See also Figure 1 The first embodiment of the present invention provides a method and system for vehicle re-identification based on a time series network with multiple cameras, including:

[0066] Step 1: Obtain image data of the vehicle under several cameras;

[0067] Step 2: Obtain an image data set by preprocessing the image data;

[0068] Step 3: Construct a vehicle re-identification model based on the memory network model and the query model;

[0069] Step 4: Train the vehicle re-identification model based on the image dataset to obtain the target vehicle re-identification model;

[0070] Step 5: Re-identify the vehicle based on the target vehicle re-identification model.

[0071] Obtain image data of the vehicle under several cameras, including:

[0072] The image data of the vehicle is collected by several cameras installed in the target area; wherein the image collected by each camera includes: a scene picture and several vehicle pictures; the target area refers to the area where the vehicle needs to be re-identified.

[0073] By preprocessing the image data, including:

[0074] An image data set is obtained by performing random transformation processing on the image data; wherein the random transformation processing includes: flipping and rotating, scaling and translating, cropping, color changing or adding random noise; wherein the data in the image data set are in units of bars, and each piece of data includes: several scene pictures and continuous frames of the same vehicle under different cameras.

[0075] The vehicle re-identification model is composed of the memory network model and the query model, including:

[0076] Build a memory network model based on the temporal convolutional network, and build a query model.

[0077] Build a memory network model based on the temporal convolutional network, including:

[0078] Three causal convolutional layers and one dilated convolutional layer form a module, and several modules are stacked together to form a memory network model;

[0079] The image dataset is labeled as x=[x1,x2,x3,…,xT]; where T is the total number of data in the image dataset;

[0080] By inputting x into the memory network model, applying causal convolution and dilated convolution to obtain the output sequence y; where y=[y1,y2,y3,…,yT], and y is a 512-dimensional vector;

[0081] The whole process is specifically expressed as:

[0082] Causal Convolutional Layer:

[0083] ;

[0084] Where y[t] is the value of the output sequence at time step t, w[i] is the weight of the i-th position in the convolution kernel, and x[ti] is the value of the input sequence at time step ti;

[0085] Expanded convolutional layer:

[0086] ;

[0087] Among them, y[t] is the value of the output sequence at time step t, w[i] is the weight of the i-th position in the convolution kernel, x[td×i] is the value of the input sequence at time step td×i, and d is the expansion rate, which controls the spacing of elements in the convolution kernel.

[0088] Build an inquiry model, including:

[0089] The query model uses the Resnet18 network to extract the features of the target vehicle image; Resnet18 is composed of multiple residual modules stacked together; the target vehicle image refers to the vehicle image that needs to be re-identified;

[0090] The structure of the residual module is as follows:

[0091] Convolutional layer 1: F(X)=σ(W1×X+b1); where W1 is the weight of the first convolutional layer, b1 is the bias term, and σ is the ReLU activation function;

[0092] Convolutional layer 2: F(F(x))=σ(W2×F(x)+b2); where F(x) is the output of the first convolution layer, W2 is the weight of the second convolution layer, and b2 is the bias term;

[0093] Identity mapping: Y=F(F(x))+W id ×X; where W id is a 1×1 convolution kernel for dimension conversion, and Y is the output 512-dimensional vector.

[0094] See also Figure 2 , train the vehicle re-identification model based on the image dataset, including:

[0095] S1: Input each data in the image dataset and the target vehicle image into the vehicle re-identification model;

[0096] S2: extracting feature data from each piece of data through a memory network model, and encoding the feature data to output a feature vector; wherein the feature data includes: static features, dynamic features and scene information of the vehicle;

[0097] S3: extracting target feature data of the target vehicle image through the query model, and converting the target feature data into a target feature vector with the same dimension as the output of the memory network;

[0098] S4: judging whether the feature vector originates from the target vehicle based on the Euclidean distance;

[0099] S5: Integrate each piece of data in the image data set and the target vehicle image as standard input data, and integrate the image corresponding to the feature vector from the target vehicle as standard output data;

[0100] S6: Train a vehicle re-identification model based on the standard input data and the standard output data, and optimize the vehicle re-identification model through a triplet loss function to obtain a target vehicle re-identification model.

[0101] Determine whether the feature vector originates from the target vehicle based on the Euclidean distance, including:

[0102] S1: The Euclidean distance between the feature vector and the target feature vector is calculated based on the Euclidean distance calculation formula;

[0103] S2: Determine whether the Euclidean distance is less than a preset distance threshold; if yes, it is determined that the feature vector originates from the target vehicle; if no, it is determined that the feature vector does not originate from the target vehicle; wherein the preset distance threshold is set according to the actual vehicle re-identification scenario.

[0104] For example, suppose three high-definition cameras are set up in a parking lot, each camera covers a different direction to capture all vehicles passing through the intersection. Now the vehicles need to be re-identified as follows:

[0105] Step 1: Obtain image data of the vehicle under several cameras;

[0106] The entrance to the parking lot in the target area is used as the target area.

[0107] Camera Setup: Three different cameras are installed in the area, numbered A, B and C.

[0108] Image data: Each camera takes an image every 1 second. The images taken by each camera include: a scene picture and several vehicle pictures.

[0109] Step 2: Preprocess the image data;

[0110] Random transformation processing: randomly flip, rotate, scale, translate, crop, change color, or add random noise to image data.

[0111] Image dataset: Each data includes a scene picture and consecutive frames of the same vehicle under different cameras.

[0112] Number of data items: Assume that there are 100 data items in the dataset, and each item contains 50 images.

[0113] Step 3: Construct a vehicle re-identification model based on the memory network model and the query model;

[0114] Memory network model: Build a temporal convolutional network (TCN) consisting of multiple stacked modules, each of which consists of three causal convolutional layers and one dilated convolutional layer.

[0115] Input sequence: x=[x1,x2,x3,…,xT], where T is the total number of data in the image dataset.

[0116] Output sequence: y=[y1,y2,y3,…,yT], each output y is a 512-dimensional feature vector.

[0117] The calculation formulas for the causal convolution layer and the dilated convolution layer are as follows:

[0118] Causal Convolutional Layer:

[0119] ;

[0120] Where y[t] is the value of the output sequence at time step t, w[i] is the weight of the i-th position in the convolution kernel, and x[ti] is the value of the input sequence at time step ti;

[0121] Expanded convolutional layer:

[0122] ;

[0123] Among them, y[t] is the value of the output sequence at time step t, w[i] is the weight of the i-th position in the convolution kernel, x[td×i] is the value of the input sequence at time step td×i, and d is the expansion rate, which controls the spacing of elements in the convolution kernel.

[0124] Query model: ResNet18 network is used to extract the features of the target vehicle image. The structure of the residual module is as follows:

[0125] Convolutional layer 1: F(X)=σ(W1×X+b1); where W1 is the weight of the first convolutional layer, b1 is the bias term, and σ is the ReLU activation function;

[0126] Convolutional layer 2: F(F(x))=σ(W2×F(x)+b2); where F(x) is the output of the first convolution layer, W2 is the weight of the second convolution layer, and b2 is the bias term;

[0127] Identity mapping: Y=F(F(x))+W id ×X; where W id is a 1×1 convolution kernel for dimension conversion, and Y is the output 512-dimensional vector.

[0128] ResNet18 is composed of multiple stacked residual modules, and the output is a 512-dimensional feature vector.

[0129] Step 4: Train the vehicle re-identification model based on the image dataset;

[0130] Training process:

[0131] S1. Input data: Input each data in the dataset into the model.

[0132] S2. Feature extraction: The memory network processes the input data to extract the static features, dynamic features and scene information of the vehicle.

[0133] S3. Query model processing: The query model receives the target vehicle image and extracts its feature vector.

[0134] S4. Euclidean distance calculation: Determine whether the feature vector originates from the target vehicle based on the Euclidean distance.

[0135] S5. Standard input / output data: Integrate each data in the dataset and the target vehicle image, as well as the image corresponding to the feature vector derived from the target vehicle.

[0136] S6. Model training: Optimize the model using the triplet loss function.

[0137] Step 5: Re-identify the vehicle based on the target vehicle re-identification model;

[0138] Suppose there is an image of a target vehicle, and you need to find all occurrences of the vehicle in the data sets of cameras A, B, and C.

[0139] 1. Data preprocessing:

[0140] Collect 50 consecutive vehicle images from each camera, including a scene picture.

[0141] The image data is randomly transformed to obtain a preprocessed image data set.

[0142] 2. Model training:

[0143] Each piece of data in the image dataset is fed into the memory network model.

[0144] The memory network model extracts feature data from each piece of data and outputs a 512-dimensional feature vector.

[0145] The query model receives the target vehicle image, extracts the target feature data, and outputs a 512-dimensional target feature vector.

[0146] The Euclidean distance is calculated to determine whether the feature vector originates from the target vehicle.

[0147] The model is optimized using the triplet loss function.

[0148] 3. Model application:

[0149] Assume there is an image of a target vehicle and ask the model to extract its feature vector.

[0150] Calculate the Euclidean distance between this feature vector and the feature vectors of all vehicle images in the dataset.

[0151] Set a preset distance threshold, such as 0.5.

[0152] If the distance between the feature vector of a vehicle image and the target feature vector is less than 0.5, the image is considered to be from the target vehicle.

[0153] Repeat this process to find all images that match the target vehicle.

[0154] By using a temporal convolutional network, the model can effectively capture the dynamic characteristics of the vehicle and improve the accuracy of recognition. The model can process images from different camera perspectives, which improves the robustness in actual scenes. By processing scene images, the model can better understand the environment in which the vehicle is located and improve its adaptability to different scenes. By using a triplet loss function, the model can converge faster, reducing the training time and computing resource requirements. The model design is flexible and can be easily extended to other types of vehicle re-identification tasks.

[0155] See also Figure 3 , the second aspect of the present invention provides a vehicle re-identification system based on a time series network multi-camera, including: a data acquisition module, a data analysis module and a vehicle re-identification model building module;

[0156] Data acquisition module: used to obtain image data of the vehicle under several cameras;

[0157] Data analysis module: used to pre-process image data to obtain image data sets;

[0158] Vehicle re-identification model building module: a vehicle re-identification model is constructed based on a memory network model and a query model; and,

[0159] A vehicle re-identification model is trained based on the image data set to obtain a target vehicle re-identification model; and a vehicle re-identification is performed on the vehicle based on the target vehicle re-identification model.

[0160] The data analysis module is communicated and / or electrically connected with the data acquisition module and the vehicle re-identification model building module respectively.

[0161] Part of the data in the above formula is calculated by removing the dimension and taking its numerical value. The formula is a formula closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0162] Working principle of the present invention:

[0163] The present invention obtains image data of a vehicle under several cameras; pre-processes the image data to obtain an image data set; forms a vehicle re-identification model based on a memory network model and a query model; trains the vehicle re-identification model based on the image data set to obtain a target vehicle re-identification model; and re-identifies the vehicle based on the target vehicle re-identification model, thereby solving the problem that the prior art has certain limitations in multi-camera scenarios when re-identifying solely by relying on the appearance features of the vehicle.

[0164] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A method for vehicle re-identification based on a time series network with multiple cameras, characterized in that: include: Step 1: Obtain image data of the vehicle under several cameras; Step 2: Obtain an image data set by preprocessing the image data; Step 3: Construct a vehicle re-identification model based on the memory network model and the query model; Step 4: Train the vehicle re-identification model based on the image dataset to obtain the target vehicle re-identification model; Step 5: Re-identify the vehicle based on the target vehicle re-identification model; The vehicle re-identification model composed of the memory network model and the query model includes: Construct a memory network model based on a temporal convolutional network, and construct a query model; The memory network model is constructed based on the temporal convolutional network, including: Three causal convolutional layers and one dilated convolutional layer form a module, and several modules are stacked together to form a memory network model; The image dataset is labeled as x=[x1,x2,x3,…,xT]; where T is the total number of data in the image dataset; By inputting x into the memory network model, applying causal convolution and dilated convolution to obtain the output sequence y; where y=[y1,y2,y3,…,yT], and y is a 512-dimensional vector; The whole process is specifically expressed as: Causal Convolutional Layer: ; Where y[t] is the value of the output sequence at time step t, w[i] is the weight of the i-th position in the convolution kernel, and x[ti] is the value of the input sequence at time step ti; Expanded convolutional layer: ; Among them, y[t] is the value of the output sequence at time step t, w[i] is the weight of the i-th position in the convolution kernel, x[td×i] is the value of the input sequence at time step td×i, and d is the expansion rate, which controls the spacing of elements in the convolution kernel; The construction of the inquiry model includes: The query model uses the Resnet18 network to extract the features of the target vehicle image; Resnet18 is composed of multiple residual modules stacked together; the target vehicle image refers to the vehicle image that needs to be re-identified; The structure of the residual module is as follows: Convolutional layer 1: F(X)=σ(W1×X+b1); where W1 is the weight of the first convolutional layer, b1 is the bias term, σ is the ReLU activation function, and X is the input data of the query model; Convolutional layer 2: F(F(x))=σ(W2×F(x)+b2); where F(x) is the output of the first convolution layer, W2 is the weight of the second convolution layer, and b2 is the bias term; Identity mapping: Y=F(F(x))+W id ×X; where W id is a 1×1 convolution kernel for dimension conversion, and Y is the output 512-dimensional vector.

2. The method for vehicle re-identification based on time series network multi-camera according to claim 1 is characterized in that: The step of obtaining image data of the vehicle under a plurality of cameras includes: The image data of the vehicle is collected by several cameras installed in the target area; wherein the image collected by each camera includes: a scene picture and several vehicle pictures; the target area refers to the area where the vehicle needs to be re-identified.

3. The method for vehicle re-identification based on time series network multi-camera according to claim 1 is characterized in that: The preprocessing of the image data includes: An image data set is obtained by performing random transformation processing on the image data; wherein the random transformation processing includes: flipping and rotating, scaling and translating, cropping, color changing or adding random noise; wherein the data in the image data set are in units of bars, and each piece of data includes: several scene pictures and continuous frames of the same vehicle under different cameras.

4. The method for vehicle re-identification based on time series network multi-camera according to claim 1 is characterized in that: The training of the vehicle re-identification model based on the image dataset includes: S1: Input each data in the image dataset and the target vehicle image into the vehicle re-identification model; S2: extracting feature data from each piece of data through a memory network model, and encoding the feature data to output a feature vector; wherein the feature data includes: static features, dynamic features and scene information of the vehicle; S3: extracting target feature data of the target vehicle image through the query model, and converting the target feature data into a target feature vector with the same dimension as the output of the memory network; S4: judging whether the feature vector originates from the target vehicle based on the Euclidean distance; S5: Integrate each piece of data in the image data set and the target vehicle image as standard input data, and integrate the image corresponding to the feature vector from the target vehicle as standard output data; S6: Train a vehicle re-identification model based on the standard input data and the standard output data, and optimize the vehicle re-identification model through a triplet loss function to obtain a target vehicle re-identification model.

5. The method for vehicle re-identification based on time series network multi-camera according to claim 4 is characterized in that: The determining whether the feature vector is derived from the target vehicle based on the Euclidean distance includes: S1: The Euclidean distance between the feature vector and the target feature vector is calculated based on the Euclidean distance calculation formula; S2: Determine whether the Euclidean distance is less than a preset distance threshold; if yes, it is determined that the feature vector originates from the target vehicle; if no, it is determined that the feature vector does not originate from the target vehicle; wherein the preset distance threshold is set according to the actual vehicle re-identification scenario.

6. A vehicle re-identification system based on a time-series network with multiple cameras, used to execute the vehicle re-identification method based on a time-series network with multiple cameras according to any one of claims 1 to 5, characterized in that: include: Data collection module, data analysis module and vehicle re-identification model building module; Data acquisition module: used to obtain image data of the vehicle under several cameras; Data analysis module: used to pre-process image data to obtain image data sets; Vehicle re-identification model building module: a vehicle re-identification model is constructed based on a memory network model and a query model; and, A vehicle re-identification model is trained based on the image data set to obtain a target vehicle re-identification model; and a vehicle re-identification is performed on the vehicle based on the target vehicle re-identification model.