A vehicle identification and tracking method based on deep learning

Abstract

The application discloses a vehicle identification and tracking method based on deep learning, which is characterized by the following steps: enabling a deep learning computer vision algorithm for a monitoring camera or a camera of a unmanned aerial vehicle, analyzing a real-time video stream, detecting and tracking a vehicle appearing in the video stream, filtering a specific vehicle by input information, capturing the specific vehicle, uploading a captured image to a background system for recording, and fully utilizing an existing hardware camera with high identification precision.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of deep learning and machine learning, and particularly relates to a vehicle identification and tracking method based on deep learning. BACKGROUND

[0002] Vehicle identification and tracking is of great significance for the pursuit of illegal and criminal vehicles. With the rapid development of computer technology, the intelligent application of video images is promoted. One of the current research hotspots of video images is motion target tracking. Current research on motion targets mainly focuses on tracking by a single camera. However, when tracking by a single camera, there are easy occlusions between vehicles, and for vehicle plate tracking, there is a problem of small monitoring area, which can easily cause low motion target tracking accuracy or inaccurate identification, thereby causing many disadvantages in actual application. SUMMARY

[0003] The application can realize cross-camera target vehicle detection and tracking by analyzing the real-time video stream returned by the deep learning algorithm based on the real-time image acquisition of the existing monitoring probe, and can send alarm information to the management personnel, thereby meeting the actual needs in the monitoring area and time, improving the work efficiency of the staff, and fighting illegal and criminal behavior to protect the safety of the country and the people.

[0004] The purpose of the application is achieved by the following technical solutions:

[0005] A vehicle identification and tracking method based on deep learning, comprising the following steps:

[0006] Step 1: Collect vehicle pictures in real traffic scenes, and label the vehicles, vehicle plate positions on the vehicles, and vehicle plate text information in the pictures to construct a vehicle detection dataset, a vehicle plate detection dataset, and a vehicle plate recognition dataset;

[0007] Step 2: Construct and train a vehicle detection network model; the input of the vehicle detection network model is a picture in a real traffic scene, the output is a prediction box coordinate of a vehicle and category information, and the prediction box coordinate of the vehicle is saved, and then a vehicle image is cropped according to the prediction box coordinate of the vehicle;

[0008] Construct and train a vehicle plate detection network model; the input of the vehicle plate detection network model is the cropped vehicle image, the output is a prediction box coordinate of a vehicle plate, and the prediction box coordinate of the vehicle plate is saved, and a vehicle plate image is cropped according to the prediction box coordinate of the vehicle plate;

[0009] Construct and train a vehicle plate recognition network model; the input of the vehicle plate recognition network model is the cropped vehicle plate image, and the output is character information in the vehicle plate;

[0010] Step three: match the recognition result of the license plate recognition network model with the to-be-tracked license plate, if the matching is successful, consider that the vehicle is the to-be-tracked vehicle, capture the image containing the to-be-tracked vehicle, determine the position information of the corresponding vehicle according to the output result of the vehicle detection network model in step two, frame mark the to-be-tracked vehicle and alarm.

[0011] Further, the vehicle detection network model comprises a backbone network and a detection head; the backbone network is based on resnet, and the input image is subjected to five times of down-sampling through convolution to reserve the feature maps of the last three scales; an SPPF module is further inserted into the backbone network, which is used for processing the input features by serially passing through a plurality of 5x5 size maximum pooling layers; the detection head adopts the FPN+PAN mode to fuse the three scale feature maps, so that the shallow features are firstly passed through the FPN, then combined with the low-level features through up-sampling to generate higher-level features, and more accurate position information is transmitted.

[0012] Further, all activation functions of the backbone network of the vehicle detection network model adopt the SILU activation function.

[0013] Further, the detection head of the vehicle detection network model adopts the idea of bounding box regression, and generates the corresponding prediction box through the anchor box with a fixed size and aspect ratio preset in advance; the center coordinates of the prediction box are calculated through the center coordinates of the anchor box, and the position information of the prediction box is calculated through the scaling ratio of the anchor box.

[0014] Further, the license plate detection network model comprises a backbone network and a detection head, and the backbone network of the license plate detection network model is based on DarkNet, detects on four scale feature maps, and introduces an attention mechanism in convolution.

[0015] Further, the license plate recognition network model comprises a visual model, a text model and a fusion model.

[0016] The visual model first extracts features of the input license plate image through the ResNet+Transformer structure to obtain an output feature map; the text model is based on RNN and Transformer, takes a probability vector of a character as input, and outputs a probability distribution of an expected character; and the fusion model splices the results of the visual model and the text model together, and then learns a weight value to adjust the influence of the visual model and the text model on the final prediction result.

[0017] Further, the fusion model outputs a character position sequence number code and a visual model result together to a text model for correction, and then iteratively corrects the language model for multiple rounds to gradually correct the recognition result and obtain the final output result.

[0018] Further, mosaic data enhancement is used for vehicle pictures in a real traffic scene, four pictures are randomly scaled, cropped and arranged to form a new picture, and the annotation frame of the new picture is obtained according to the annotation frames of the original four pictures; the vehicle recognition and tracking model is trained using the new picture.

[0019] An electronic device comprises:

[0020] One or more processors;

[0021] A storage device for storing one or more programs, when the one or more programs are executed by the electronic device, the electronic device implements a deep learning-based vehicle recognition and tracking method.

[0022] A computer-readable storage medium having a program stored thereon, the program being executed by a processor to implement a deep learning-based vehicle recognition and tracking method.

[0023] The beneficial effects of the present application are as follows:

[0024] (1) The vehicle detection network model, license plate detection network model and license plate recognition network model are connected in series in the present application, SPPF is used for multi-scale fusion of pictures, and compared with SPP which needs to specify three convolution kernel sizes, the calculation speed is faster through the convolution module to perform three times of pooling and splicing operations on the data; the attention mechanism is introduced, so that the network can better extract features, and for small targets such as license plates, one more scale of feature maps is added to improve the detection accuracy. In the text recognition part, through the multi-modal model, bidirectional expression can be realized, based on the complete fill-in-the-blank idea, spelling correction language model is made, and then the language model is repeatedly executed for multiple rounds to further correct the recognition result, thereby improving the character recognition accuracy.

[0025] (2) In the present application, when detecting vehicles, the multi-scale fusion method can also well detect vehicle targets in the mutual occlusion scene between vehicles; for small targets such as license plates, the small target features are better extracted through the way of increasing feature maps, thereby improving the detection accuracy; in addition, for license plate recognition, for the problem of Chinese recognition difficulty, the language model is repeatedly executed for multiple rounds to gradually correct the recognition result and obtain the final output result, thereby improving the recognition accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 Flow chart of the vehicle identification tracking method of the present application.

[0027] Figure 2 Schematic diagram of CarNet backbone network.

[0028] Figure 3 Schematic diagram of LPDNet network structure.

[0029] Figure 4 Specific implementation flow chart of attention mechanism.

[0030] Figure 5 Schematic diagram of LPRNet network structure.

[0031] Figure 6 Picture of input CarNet of the embodiment.

[0032] Figure 7 Schematic diagram of detection frame output by CarNet.

[0033] Figure 8 Vehicle image cropped according to the prediction frame output by CarNet.

[0034] Figure 9 Schematic diagram of detection frame output by LPDNet. DETAILED DESCRIPTION

[0035] The exemplary embodiments will be described in detail herein with reference to the attached drawings. In the following description, the same numbers are used to indicate the same or similar components. The embodiments described in the following exemplary embodiments do not represent all the implementations consistent with the present application. Instead, they only represent examples of devices and methods consistent with some aspects of the present application, as detailed in the appended claims.

[0036] The terminology used in the present application is for the purpose of describing particular embodiments only and is not intended to be limiting of the present application. As used in the present application and the appended claims, the singular forms "a," "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0037] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0038] like Figure 1 As shown, the vehicle identification and tracking method based on deep learning of the present invention empowers surveillance cameras or drone cameras with deep learning computer vision algorithms, enabling analysis of real-time video streams, detection and tracking of vehicles appearing in the video, filtering out specific vehicles through input information and capturing images of those vehicles, while simultaneously uploading the captured images to a background system for recording. The method includes the following steps:

[0039] Step 1: Collect vehicle images from real traffic scenes, and annotate the vehicles, license plate locations, and license plate text information to construct vehicle detection datasets, license plate detection datasets, and license plate recognition datasets.

[0040] Step 2: Construct and train a vehicle detection network model; the input of the vehicle detection network model is an image from a real traffic scene, and the output is the predicted bounding box coordinates of the vehicle and category information (including car, bus, truck, etc.), which are then saved. The vehicle image is then cropped based on the predicted bounding box coordinates of the vehicle.

[0041] A license plate detection network model is constructed and trained. The input of the license plate detection network model is a cropped vehicle image, and the output is the predicted bounding box coordinates of the license plate. The model is saved, and the license plate image is cropped according to the predicted bounding box coordinates of the license plate.

[0042] A license plate recognition network model is constructed and trained. The input of the license plate recognition network model is a cropped license plate image, and the output is the character information in the license plate.

[0043] Step 3: Match the recognition results of the license plate recognition network model with the license plate to be tracked. If the match is successful, the vehicle is considered to be the vehicle to be tracked. Capture the image containing the vehicle to be tracked. Based on the output results of the vehicle detection network model in Step 2, determine the location information of the corresponding vehicle, mark the tracked vehicle with a bounding box and issue an alarm.

[0044] First, the mosaic data enhancement method is used to randomly scale, randomly crop and randomly arrange the input four pictures. Each picture has a corresponding label box. The four pictures are spliced together to obtain a new picture, and the label box of the picture can be obtained according to the original four picture label boxes. The generated picture is transmitted to the vehicle recognition and tracking model for training, which is like receiving four images at the same time, increasing the model's ability to detect the background, effectively solving the problem of poor detection effect of small targets in model training, and improving the robustness of the model. And when performing BN operation, four images can also be processed at the same time to achieve the purpose of reducing training resources.

[0045] (1) Vehicle detection network model CarNet

[0046] The CarNet network includes a backbone network (Backbone) and a detection head (Head) two parts, and the network diagram is as shown in the figure Figure 2 The backbone network is based on resnet, and five times of convolution are performed for downsampling, and the feature maps of the last three scales are reserved. For an image with an input size of 640x640, three scale feature maps with sizes of 80x80, 40x40 and 20x20 will be obtained finally. All activation functions in the backbone network use the SILU activation function, which has the characteristics of no upper bound, smoothness and non-monotonicity, and performs better than the Leaky_relu activation function in the network model. As shown in Figure 2 The basic CBS module represents the convolution, BN and SILU activation operations; the Basic1 module represents the serial connection of two CBS modules; the Basic2 module represents the operation of short-circuiting the output of the Basic1 module and the input through a shortcut; and the Layer module represents the serial connection of two Basic2 modules. The Focus module divides the data into four parts, each part of the data is 2 times down-sampling, then splicing in the channel, and finally convolution operation. Its greatest advantage is that when performing down-sampling, it can minimize information loss.

[0047] In order to better extract the fusion features, the SPPF module is inserted in the Backbone. The traditional SPP (Spatial Pyramid Pooling) module adopts the maximum pooling method of 1x1, 5x5, 9x9 and 13x13 to perform multi-scale fusion on the picture. The SPPF structure is to serially pass the input features through multiple 5x5 MaxPool layers for processing. The serial passing through two 5x5 MaxPool layers is the same as the calculation result of a 9x9 MaxPool layer, and the serial passing through three 5x5 MaxPool layers is the same as the calculation result of a 13x13 MaxPool layer. SPPF only needs to specify one convolution kernel, and the output after each pooling will become the input of the next pooling. Compared with SPP which needs to specify the size of the convolution kernel three times and then performs three times of pooling and splicing operations on the data through the convolution module, the calculation speed is faster. SPPF aims to enhance the feature expression ability of the feature map, which can effectively avoid the image distortion caused by the cropping and scaling operations, and also can well solve the problem of repeated feature extraction of the convolution network, thereby speeding up the generation of the candidate frame and saving the calculation cost.

[0048] In order to improve the detection accuracy, the FPN+PAN method is used to fuse the three scale feature maps obtained. FPN uses a top-down method to combine the feature maps obtained by upsampling with low-level features to generate higher-level features and transmit more accurate position information. The structure of PAN includes bottom-up link and top-down link. Because the structure of FPN transmits shallow features to the upper layer, it needs dozens of layers or even hundreds of layers, so the shallow information will be lost more. However, the path from bottom to top is generally less, and the shallow features are first transmitted through FPN and then connected with the upper layer through upsampling to prevent excessive information loss. Therefore, the high-level feature map is combined with the low-level feature map to obtain a new feature map. The feature map not only contains a large amount of semantic information, but also has a large number of pixels, so it can better detect the image.

[0049] The detection head adopts the idea of bounding box regression, and the anchor frame with a fixed size and aspect ratio is used to form the prediction frame. The prediction frame can be regarded as a fine tuning based on the anchor frame. Each anchor frame has a corresponding prediction frame. The center coordinates of the prediction frame can be calculated through the center coordinates of the anchor frame, and the position information of the prediction frame can be calculated through the scaling ratio of the anchor frame.

[0050] (II) License plate detection network model LPDNet

[0051] Collect vehicle pictures in real traffic scenes, label the license plates in the images, and the label information includes not only the position information but also the license plate text information. Then, a license plate detection database is established according to the labeled position information.

[0052] The schematic diagram of the license plate detection network model LPDNet is shown in FIG. 3. The LPDNet network is also composed of two parts of Backbone and Head. Different from the CarNet network, since the license plate is a small target under the monitoring angle, the backbone network and the detection head are improved to improve the detection accuracy. The backbone network of the LPDNet is based on DarkNet, and detection is performed on four scale feature maps to improve the detection effect of small target features. At the same time, an attention mechanism is introduced for the detection system to improve the detection performance of the system.

[0053] Among them, the basic CBS module represents the convolution, BN and SILU activation operation; the Bottleneck_F module represents the serial connection of two CBS modules; and the Bottleneck module represents the operation of splicing the output of the Bottleneck_F module and the input through a shortcut.

[0054] The C3_x module is divided into two parts after the CBS module, one part is processed through x Bottleneck modules, and the other part is processed through another CBS module, and then the two part results are spliced together and finally adjusted through the CBS module; the structure of the C3_x_F module is the same as that of the C3_x module, except that the Bottleneck module in the structure is replaced by the Bottleneck_F module. The SPPF structure is to process the input features through multiple 5x5 MaxPool layers in series, and to perform feature fusion on the picture.

[0055] The attention mechanism can be used as an additional network applied to the conv in the network, which hard selects specific inputs or gives different weights to features. The attention mechanism can be used as an additional network applied to the conv in the network, which hard selects specific inputs or gives different weights to features. The attention mechanism here is mainly the attention mechanism in the neural network. The neural network can not only learn by itself according to the attention mechanism, but also understand the characteristics of the network through the attention mechanism, so as to better improve the performance of the model.

[0056] The basic idea of the attention mechanism is to identify the key features in the image with a new weighting method, and then through learning and training, the network can identify the part of each image that needs attention, so as to form attention. The specific implementation process is as follows Figure 4As shown, first, the input feature map is globally averaged and pooled, and then passed through a series of fully connected layers and a sigmoid activation function to obtain the channel attention weight. Finally, the channel attention weight and the input feature map are multiplied.

[0057] (III) License plate recognition network LPRNet

[0058] A dictionary is established for the character information in the license plate, and the dictionary content includes Chinese characters, uppercase English letters, and numbers. The license plate recognition network LPRNet is mainly responsible for recognizing the character information in the license plate and saving the output character information. The input of the license plate recognition network LPRNet is the license plate image detected by the license plate detection network model LPDNet. The license plate image is cut from the original image and used as the input of the license plate recognition network LPRNet, while the position information of the license plate is retained. The label is converted into the index of the corresponding dictionary for each character. The output result of the license plate recognition network LPRNet is mapped to the dictionary to obtain the corresponding character. The structure diagram of the LPDNet is as shown in Figure 5 The LPDNet mainly includes a visual model, a text model, and a fusion model. The visual model extracts features, the text model corrects, and the fusion model fuses the results of the two models together and calculates the final recognition result according to the adaptive weight.

[0059] The Backbone of the visual model first extracts features from the input text picture through the ResNet+Transformer structure to obtain the output feature map. The Position Attention of the Transformer is different from the traditional SelfAttention. It directly uses Attention. Q is the position encoding generated by itself. The initial value is similar to the sine coding plus a linear transformation. K represents the vector of the relevance of information to other information, and V represents the vector of information content. The purpose of doing such Position Attention is to fix the information of each letter position.

[0060] For general ocr text recognition, the output of the visual module is directly connected to the multi-classification loss function, which can be realized. However, by adding a text model to form a multi-modal model, the recognition accuracy can be effectively improved. The network architecture of the text model is based on RNN and Transformer, taking the output result of the visual model and the character position sequence encoding as input, and outputting the index of the predicted character. The purpose is to correct, check each letter one by one, and iterate n times to improve the model accuracy.

[0061] The input of the text model is the output of the visual model and the character position sequence code, and then a softmax function is connected, without changing the dimension of the input; then the dimension is increased to obtain more information; and the current position is inferred according to the context, and the information of the current position is shielded when calculating the attention, so as to realize the rectification function by updating the input.

[0062] The fusion model splices the results of the visual model and the text model, and then learns a weight value to adjust the influence of the visual model and the text model on the final prediction result. The fusion model outputs a character position sequence code and a visual model result, which are sent to the text model for correction. Then, the iterative correction idea is used to execute the language model repeatedly for multiple rounds, so that the recognition effect is gradually corrected, and the final output result is obtained.

[0063] The loss function of the license plate recognition network LPRNet is a general multi-classification cross-entropy loss, and the corresponding loss of three parts should be considered, that is, the visual model loss, the text model loss and the fusion model loss.

[0064] Step three: matching the recognition result of the license plate recognition network model with the to-be-tracked license plate, if the matching is successful, considering that the vehicle is the to-be-tracked vehicle, the image containing the to-be-tracked vehicle is captured, the position information of the corresponding vehicle is determined according to the output result of the vehicle detection network model in step two, the to-be-tracked vehicle is framed and labeled, and an alarm is given.

[0065] Corresponding to the foregoing embodiments of the vehicle recognition tracking method, the present application also provides an electronic device corresponding thereto. The electronic device comprises one or more processors and a storage device for storing one or more programs, when the one or more programs are executed by the electronic device, the electronic device implements the vehicle recognition tracking method based on deep learning. The electronic device can be applied to any device with data processing capability, which can be a device such as a computer or a device. The device embodiment can be realized by software, or realized by hardware or a combination of software and hardware. Taking software implementation as an example, as a logical device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory and running by the processor of the electronic device.

[0066] For the apparatus embodiment, since it basically corresponds to the method embodiment, the relevant part is described in the part of the method embodiment. The apparatus embodiment described above is only illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place or distributed on multiple network units. Some or all of the modules can be selected to achieve the purpose of the present application according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0067] The embodiment of the present application also provides a computer readable storage medium, which stores a program, and the program is executed by a processor to realize the deep learning based vehicle identification and tracking method in the above embodiment.

[0068] The computer readable storage medium can be an internal storage unit of any device with data processing capability, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device, such as a plug-in hard disk, a smart media card (SMC), an SD card, a flash card, etc. Further, the computer readable storage medium can include both the internal storage unit of any device with data processing capability and the external storage device. The computer readable storage medium is used to store the computer program and other programs and data required by the device with data processing capability, and can also be used to temporarily store data that has been output or will be output.

[0069] The following gives a specific application example of the vehicle identification and tracking method of the present application. A picture as shown in FIG. 1 is input into the vehicle detection network model CarNet, and the model output is as shown in FIG. 2, which is the prediction box coordinates, category information and confidence of each vehicle. Figure 6 As can be seen from FIG. 2, the vehicle detection network model can accurately identify the category of the vehicle and output the prediction box with accurate size and position. Figure 7 Figure 7 As can be seen from FIG. 2, the vehicle detection network model can accurately identify the category of the vehicle and output the prediction box with accurate size and position. Figure 8 A vehicle image cropped according to one of the prediction boxes in FIG. 2 is input into the license plate detection network model, and the model output is as shown in FIG. 3, which is the license plate detection box and confidence. Figure 7 Figure 9 As can be seen from FIG. 3, the license plate detection network model can accurately identify the category of the vehicle and output the prediction box with accurate size and position. Figure 9 ​​The license plate image cropped by the detection frame is input into a license plate recognition network model, OCR recognition is performed, and the result is mapped into a dictionary to obtain corresponding characters. Other embodiments of the application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. The specification and examples given are considered exemplary only, and the true scope and spirit of the application is indicated by the claims. The following examples are provided to further illustrate the application.

[0070] It is to be understood that the application is not limited to the precise construction hereinafter described and as shown in the attached drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application is limited only by the claims that follow.

Claims

1. A deep learning-based vehicle recognition tracking method, characterized by, The method comprises the following steps: Step one: collect vehicle pictures in real traffic scenes, and label the vehicles, license plate positions on the vehicles and license plate text information therein to construct a vehicle detection dataset, a license plate detection dataset and a license plate recognition dataset; Step two: construct and train a vehicle detection network model; the input of the vehicle detection network model is a picture in a real traffic scene, the output is prediction box coordinates and category information of a vehicle, and is saved, and then a vehicle image is cropped according to the prediction box coordinates of the vehicle; construct and train a license plate detection network model; the input of the license plate detection network model is the cropped vehicle image, the output is prediction box coordinates of a license plate, and is saved, and a license plate image is cropped according to the prediction box coordinates of the license plate; construct and train a license plate recognition network model; the input of the license plate recognition network model is the cropped license plate image, and the output is character information in the license plate; Step three: match the recognition result of the license plate recognition network model with a to-be-tracked license plate; if the matching is successful, the vehicle is considered to be the to-be-tracked vehicle, an image containing the to-be-tracked vehicle is captured, the position information of the corresponding vehicle is determined according to the output result of the vehicle detection network model in step two, the to-be-tracked vehicle is framed and labeled, and an alarm is given; The vehicle detection network model comprises a backbone network and a detection head; the backbone network is based on resnet, and the input image is convoluted and down-sampled five times to retain the feature maps of the last three scales; an SPPF module is also inserted into the backbone network to serially process the input features through multiple 5x5 size maximum pooling layers; the detection head adopts an FPN+PAN mode to fuse the three scale feature maps, so that the features of the shallow layer are first passed through the FPN, then combined with the low-level features through up-sampling to generate higher-level features and pass more accurate position information; The license plate recognition network model comprises a vision model, a text model and a fusion model; The vision model first extracts features of the input license plate image through a ResNet+Transformer structure to obtain an output feature map; the text model takes a probability vector of a character as input and outputs a probability distribution of an expected character based on RNN and Transformer; and the fusion model splices the results of the vision model and the text model together, and then learns a weight value to adjust the influence of the vision model and the text model on the final prediction result. 2.The deep learning-based vehicle identification tracking method of claim 1, wherein, All activation functions of the backbone network of the vehicle detection network model adopt a SILU activation function. 3.The deep learning-based vehicle identification tracking method of claim 2, wherein, The detection head of the vehicle detection network model adopts the idea of bounding box regression, generates corresponding prediction boxes through anchor boxes of a fixed size and aspect ratio, calculates the center coordinates of the prediction boxes through the center coordinates of the anchor boxes, and calculates the position information of the prediction boxes through the scaling ratios of the anchor boxes. 4.The deep learning-based vehicle identification tracking method of claim 1, wherein, The license plate detection network model comprises a backbone network and a detection head, the backbone network of the license plate detection network model is based on DarkNet, detection is performed on four scale feature maps, and an attention mechanism is introduced in convolution. 5.The deep learning-based vehicle identification tracking method of claim 1, wherein, The fusion model outputs a character position sequence number code and a visual model result, and sends them into a text model for correction, and then iteratively executes a language model for multiple rounds, so that the recognition effect is gradually corrected, and finally an output result is obtained. 6.The deep learning-based vehicle identification tracking method of claim 1, wherein, The vehicle picture in a real traffic scene is subjected to mosaic data enhancement, four pictures are randomly scaled, randomly cropped and randomly arranged to form a new picture, and the label frame of the new picture is obtained according to the label frames of the original four pictures; and the vehicle recognition and tracking model is trained by using the new picture.

7. An electronic device, comprising: Comprise: One or more processors; A storage device for storing one or more programs, which, when executed by the electronic device, cause the electronic device to implement the deep learning-based vehicle recognition and tracking method according to any one of claims 1 to 6.

8. A computer-readable storage medium, characterized in that, A storage device for storing one or more programs, which, when executed by the electronic device, cause the electronic device to implement the deep learning-based vehicle recognition and tracking method according to any one of claims 1 to 6.

Images