Track area recognition method based on large target semantic segmentation model in limited scenes

By adopting a semantic segmentation model of encoder-decoder structure in rail transit, combining ResNet and hollow convolution technology, the problem of inaccurate obstacle recognition in traditional models in rail transit scenarios is solved, and the accurate identification of rail areas is achieved and efficient identification, suitable for complex environments of rail transit.

CN114140764BActive Publication Date: 2024-11-01CASCO SIGNAL LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111312558.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-08
Publication Date
2024-11-01
Estimated Expiration
2041-11-08

AI Technical Summary

Technical Problem

In the existing rail transit obstacle detection system, the traditional vehicle driving area identification model algorithm in the field of automobile autonomous driving is not suitable for limited scenarios of rail transit, resulting in inaccurate obstacle identification and increasing safety risks.

Method used

Using a semantic segmentation model based on the encoder-decoder structure, using the ResNet residual network as the backbone network, adding convolution network and Block network, introducing hollow convolution and Batch Normalization, designing three-branch encoder structure and shallow feature stitching. Achieve accurate identification of track areas.

Benefits of technology

It improves the effectiveness and recognition accuracy of feature extraction in rail area, reduces the number of model parameters, improves the computing speed, is suitable for different scenarios, has strong generalization ability and high operating efficiency, and is suitable for complex environments of rail transit.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114140764B_ABST
    Figure CN114140764B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for identifying a track area based on a large target semantic segmentation model in a limited scene, the method comprising: step S 1 , determine the track area recognition algorithm, and use the encoder-decoder structure model to perform semantic segmentation on the image; step S 2 , select the feature backbone network model and select the ResNet residual network as the backbone network prototype architecture; Step S 3 , reconstruct the backbone network architecture, add the first layer of convolutional network and three groups of Block networks; Step S 4 , the number of layers of the new backbone network model is 40; step S 5 , add activation layer ReLU after each Block; step S 6 , adopting the hole convolution method to design the encoder structure, etc. Compared with the prior art, the present invention has the advantages of improving the operation speed and being more effective in extracting the features of the track area.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to an obstacle detection technology for rail transit, and in particular to a track area recognition method based on a large target semantic segmentation model in a limited scene. Background Art

[0002] With the rapid development of the rail transit industry, the operating environment of national railways, subways, and trams has become increasingly complex. Among the various safety accidents that have occurred in recent years, a considerable proportion is caused by the front obstacle invading the limit and colliding with the train. Since the braking distance of the train is long, it is very necessary to install an active obstacle detection system on the train. The convolutional neural network algorithm is used as the core technology of the obstacle detection system. In the obstacle detection system based on video recognition technology, in addition to the identification of obstacles, the most critical thing is to accurately identify the driving area of ​​the train, that is, the "track area". On the basis of the identification of the track area, the vehicle spacing limit is defined, which is an important first step in the obstacle identification and processing in the rail transit field, and it is also a big difference from the field of automatic driving. When the camera detects that there is an obstacle in front of the train, it is necessary to determine whether the position of the obstacle is located in the track area. Only when the obstacle is in the track area will it pose a safety threat to the vehicle. Therefore, it is necessary to design a model algorithm to identify the driving area of ​​the train in the image, that is, the track area. The traditional road detection algorithm for autonomous driving is based on lane lines, but train vehicles use rails as their driving area, and the body of the vehicle is slightly wider than the distance between the two rails. The "driving limit" is calculated based on the body shape and the lateral swing during operation. There are multiple rails intersecting in the turnout area, which brings more complexity. The route of the train is a fixed, limited scene with repeatability, which is quite different from the traditional car driving environment. Therefore, the current common vehicle driving area recognition model algorithm in the field of autonomous driving is not applicable to this scenario. Summary of the invention

[0003] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a track area recognition method based on a large target semantic segmentation model in a limited scene.

[0004] The purpose of the present invention can be achieved by the following technical solutions:

[0005] According to a first aspect of the present invention, a method for identifying a track area based on a large target semantic segmentation model in a limited scene is provided, the method comprising:

[0006] Step S1, determining a track area recognition algorithm, and using an encoder-decoder structure model to perform semantic segmentation on the image;

[0007] Step S2, screening the feature backbone network model, and selecting the ResNet residual network as the backbone network prototype architecture;

[0008] Step S3, reconstruct the backbone network architecture, add the first layer of convolutional network and three groups of Block networks;

[0009] Step S4: the number of layers of the new backbone network model is 40;

[0010] Step S5, add an activation layer ReLU after each Block;

[0011] Step S6, using a dilated convolution method to design the encoder structure;

[0012] Step S7: Under each branch of the coding region, a Batch Normalization is designed to normalize the outputs of different dilation rate layers;

[0013] Step S8, after normalizing the data of the three different branches, three activation layers ReLU are added respectively;

[0014] Step S9, upsampling after adding an activation layer to the first pooling layer branch of the encoder;

[0015] Step S 10 , connect the outputs of the three branches of the encoder, and then add a convolutional layer as the output of the encoder;

[0016] Step S 11 , decoder structure design, introducing shallow features in the backbone feature network for splicing;

[0017] Step S 12 , the shallow features introduced are used as the last set of outputs of the first Block of the backbone feature network, and the output is convolved and parameter normalized;

[0018] Step S 13 , add activation layer ReLU to the output layer of the shallow network;

[0019] Step S 14 , the encoder output layer needs to be upsampled after adding the activation layer ReLU;

[0020] Step S 15 , step S 13 and S 14 The output layer of is connected and an upsampling layer is added, the size of which is the same as the input image, thus restoring the full image size;

[0021] Step S 16 , step S 13 A convolutional layer is added after the output layer to build a model, which is used to determine whether a pixel is a track area.

[0022] As a preferred technical solution, the first separate convolutional layer in step S4 has name=conv1_conv;

[0023] The first Block contains three groups;

[0024] The second Block contains four groups;

[0025] The third Block contains six groups.

[0026] As a preferred technical solution, each group of the first Block includes the following three layers:

[0027] kernel_size=1*1, filters=64, strides=1*1, bias=64, name=conv2_block1_X_1;

[0028] kernel_size=3*3, filters=256, strides=1*1, bias=64, name=conv2_block1_X_2;

[0029] kernel_size=1*1, filters=256, strides=1*1, bias=256, name=conv2_block1_X_3.

[0030] As a preferred technical solution, each group of the second Block includes the following three layers:

[0031] kernel_size=1*1, filters=128, strides=2*2, bias=128, name=conv2_block2_X_1;

[0032] kernel_size=3*3, filters=128, strides=1*1, bias=128, name=conv2_block2_X_2;

[0033] kernel_size=1*1, filters=512, strides=1*1, bias=512, name=conv2_block2_X_3.

[0034] As a preferred technical solution, each group of the third Block includes the following three layers:

[0035] kernel_size=1*1, filters=256, strides=2*2, bias=256, name=conv2_block3_X_1;

[0036] kernel_size=3*3, filters=256, strides=1*1, bias=256, name=conv2_block3_X_1;

[0037] kernel_size=1*1, filters=1024, strides=1*1, bias=1024, name=conv2_block3_X_1.

[0038] As a preferred technical solution, the step S6 designs the encoder into a three-branch structure according to the usage environment of the semantic segmentation model, the characteristics of limited scenes in the tunnel and large targets in the track area.

[0039] As a preferred technical solution, the three branch structures are as follows:

[0040] 1) The first branch, the only pooling layer of the encoder module, uses an average pooling layer to take the average value of the image;

[0041] Pool_size=27*50, strides=27*50, name=Encoder_Average_Pooling;

[0042] kernel_size=1*1, filters=36, strides=1*1, bias=False, name=Pooling_1x1conv2d;

[0043] 2) The second branch uses a dilated convolutional layer with a dilation rate of 1;

[0044] kernel_size=1*1, filters=36, strides=1*1, bias=False, name=Encoder_conv2d_d1;

[0045] 3) The third branch uses a dilated convolutional layer with a dilation rate of 6;

[0046] kernel_size=3*3, filters=36, strides=1*1, bias=False, name=Encoder_conv2d_d6.

[0047] As a preferred technical solution, the size of the up-sampled data in step S9 is: input image height / 16 and padded to an integer, input image width / 16 and padded to an integer.

[0048] As a preferred technical solution, the step S 10 The convolutional layer in is designed as follows:

[0049] kernel_size=1*1, filters=36, strides=1*1, bias=False, name=Encoder_conv2d_Final.

[0050] As a preferred technical solution, the step S 12 The convolutional layer in is designed as follows:

[0051] kernel_size=1*1, filters=48, strides=1*1, bias=False, name=Decoder_Low_Pre.

[0052] As a preferred technical solution, the step S 14 The size of the upsampled data in is: input image height / 4 and padded to an integer, input image width / 4 and padded to an integer.

[0053] As a preferred technical solution, the step S 16 The convolutional layer in is designed as follows:

[0054] kernel_size=1*1, filters=2, strides=1*1, bias=2, name=Con_Deocoder_Pre.

[0055] According to a second aspect of the present invention, there is provided an electronic device, comprising a memory and a processor, wherein a computer program is stored in the memory, and the method described above is implemented when the processor executes the program.

[0056] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, wherein the program implements the method described above when executed by a processor.

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

[0058] 1. A more streamlined backbone feature network is designed using the common ResNet residual network, which retains the shallow modules of the classic ResNet50 network, removes unnecessary redundant layers and deeper modules, and has only 40 layers of convolution in the backbone network model, which greatly reduces the number of parameters, improves the calculation speed, and is more effective in extracting features from the track area.

[0059] 2. In the encoder part, a double-hole convolution layer and a pooling layer with three branches in parallel are innovatively adopted. Considering the characteristics of large targets and limited scenes in the track area, hole convolution layers with expansion rates of 1 and 6 are used to achieve good feature encoding and extraction.

[0060] 3. Due to the singularity of the track area features, the shallowest feature layer of the backbone feature network is introduced in the decoder. Compared with other feature layers, the introduction of this layer can simultaneously take into account the restoration of track detail edges in the upsampled image and the collection of large feature information.

[0061] 4. It has good generalization ability in different scenarios. Traditional image processing algorithms require fine-tuning of parameters or even modification of functions for different environments. This model only needs to add images with poor effects and retrain them without changing the model. Model training takes less time and the effect is better.

[0062] 5. The designed model is very light and only 36M in size. It can easily reach 20FPS when running on a 30TOPS platform, and the running effect is good. It can accurately identify the switch area, platform area and track area of ​​the elevated section including multiple tracks. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a detailed design drawing of the model of the present invention;

[0064] Figure 2 A flow chart for training the model of the present invention;

[0065] Figure 3(a) to Figure 3(d) This is the usage effect of the model of the present invention in different scenarios. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not 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 should fall within the scope of protection of the present invention.

[0067] The present invention provides a track area recognition method based on a large target semantic segmentation model in a limited scene, which specifically includes:

[0068] Step S1: determine the track area recognition algorithm scheme, and use the encoder-decoder structure model to perform semantic segmentation on the image.

[0069] Step S2: Screen the feature backbone network model, compare the parameters and architecture design of the current mainstream backbone network models, and finally select the ResNet residual network as the backbone network prototype architecture.

[0070] Step S3: Reconstruct the backbone network architecture, add the first layer of convolutional network, and three groups of Block networks.

[0071] Step S4: The new backbone network model has 40 layers and the structure is as follows:

[0072] The first single convolutional layer, name=conv1_conv

[0073] The first Block contains three groups, each of which includes the following three layers:

[0074] kernel_size=1*1, filters=64, strides=1*1, bias=64, name=conv2_block1_X_1;

[0075] kernel_size=3*3, filters=256, strides=1*1, bias=64, name=conv2_block1_X_2;

[0076] kernel_size=1*1, filters=256, strides=1*1, bias=256, name=conv2_block1_X_3;

[0077] The second Block contains four groups, each of which includes the following three layers:

[0078] kernel_size=1*1, filters=128, strides=2*2, bias=128, name=conv2_block2_X_1;

[0079] kernel_size=3*3, filters=128, strides=1*1, bias=128, name=conv2_block2_X_2;

[0080] kernel_size=1*1, filters=512, strides=1*1, bias=512, name=conv2_block2_X_3;

[0081] The third Block contains six groups, each of which includes the following three layers:

[0082] kernel_size=1*1, filters=256, strides=2*2, bias=256, name=conv2_block3_X_1;

[0083] kernel_size=3*3, filters=256, strides=1*1, bias=256, name=conv2_block3_X_1;

[0084] kernel_size=1*1, filters=1024, strides=1*1, bias=1024, name=conv2_block3_X_1;

[0085] Step S5: Add an activation layer ReLU after each Block:

[0086] Step S6, encoder structure design, using the dilated convolution method. According to the use environment of the model, the scene in the tunnel is limited and the target in the track area is large. Therefore, the encoder layer is designed as a three-branch structure as follows:

[0087] The first branch, the only pooling layer in the encoder module, uses an average pooling layer to take the average value of the image

[0088] Pool_size=27*50, strides=27*50, name=Encoder_Average_Pooling

[0089] kernel_size=1*1, filters=36, strides=1*1, bias=False, name=Pooling_1x1conv2d;

[0090] The second branch uses a dilated convolutional layer with a dilation rate of 1.

[0091] kernel_size=1*1, filters=36, strides=1*1, bias=False, name=Encoder_conv2d_d1;

[0092] The third branch uses a dilated convolutional layer with a dilation rate of 6.

[0093] kernel_size=3*3, filters=36, strides=1*1, bias=False, name=Encoder_conv2d_d6;

[0094] Step S7: Under each branch of the coding region, a Batch Normalization is designed to normalize the outputs of different dilation rate layers.

[0095] Step S8: After normalizing the data of the three different branches, three activation layers ReLU are added respectively.

[0096] Step S9: Upsample after the activation layer of the first pooling layer branch of the encoder, with a size of (input image height / 16 and padded to an integer, input image width / 16 and padded to an integer).

[0097] Step S 10 , connect the outputs of the three branches of the encoder, and then add a convolutional layer as the output of the encoder. The convolutional layer is designed as shown below.

[0098] kernel_size=1*1, filters=36, strides=1*1, bias=False, name=Encoder_conv2d_Final;

[0099] Step S 11 , decoder structure design. Since the features of the track area in the tunnel are simple and relatively single, it is necessary to introduce shallow features in the backbone feature network for splicing in the design process of this model.

[0100] Step S 12 The shallow features introduced are the last set of outputs of the first block of the backbone network, that is, the output of conv2_block1_3_3. The output is convolved and parameter normalized. The convolution layer is designed as follows:

[0101] kernel_size=1*1, filters=48, strides=1*1, bias=False, name=Decoder_Low_Pre;

[0102] Step S 13 , add activation layer ReLU to the output layer of the shallow network, name = Low_level_activation.

[0103] Step S 14, after adding the activation layer ReLU to the encoder output layer, it needs to be upsampled. The size is the input image height / 4 and padded to an integer, the input image width / 4 and padded to an integer), name=Decoder_upsample_4x4.

[0104] Step S 15 , step S 13 and S 14 The output layer of is connected and an upsampling layer is added. The layer name = Decoder_upsample_WH has the same size as the input image, thus restoring the full image size.

[0105] Step S 16 , step S 13 A convolutional layer (output layer) is added after the output layer of . Since the model is designed to determine whether a pixel is a track area, that is, there are only two categories, the design is as follows:

[0106] kernel_size=1*1, filters=2, strides=1*1, bias=2, name=Con_Deocoder_Pre. Specific embodiments

[0108] The present invention designs a neural network model algorithm which can identify the track area where a train is running from an image with pixel-level accuracy.

[0109] First reference Figure 1 That is, the detailed design of the neural network, the encoder-decoder architecture.

[0110] Step 100 designs a backbone network based on the residual ResNet model. After adding the first convolutional layer and pooling layer, three residual modules Block1 to Block3 are added.

[0111] Step 101 designs the encoder as a three-branch architecture.

[0112] Step 102 adds a pooling layer to the encoder module.

[0113] Step 103 connects an upsampling layer after step 102, whose size is one sixteenth of the image.

[0114] Step 104 adds two dilated convolutional layers with dilation rates of 1 and 6 respectively to the encoder module.

[0115] Step 105 connects the three branches of the encoder to form an encoder module.

[0116] Step 106 designs the decoder module and introduces the shallow features of the backbone network, that is, the output of Block 1.

[0117] Step 107 adds a convolution layer to the output of Block 1, and adds a normalization layer and an activation layer.

[0118] Step 108 adds a normalization layer and an activation layer to the output of the encoder in step 105

[0119] Step 109 upsamples the output of step 108 to a size one-quarter of the image.

[0120] Step 110 connects the outputs of steps 109 and 107 .

[0121] Step 111 upsamples the output of step 110 to the same size as the input image.

[0122] Step 112 adds a convolutional layer after the output of step 111, and the model design is now complete.

[0123] And refer to Figure 2 How to train and use this neural network model.

[0124] Step 113 first performs data collection, and collects video data during the operation of the train through cameras deployed on the train.

[0125] Step 114 converts the video stream into images through a script program.

[0126] Step 115 filters the image generated in step 114 and selects representative areas, such as elevated roads, entry and exit areas of platforms, and switch areas with multiple tracks.

[0127] Step 116 labels the filtered images, using the labeling tool labelme to depict the track area on the image with polygons to form a data set.

[0128] Step 117 converts the dataset format into a single-channel image with a value of 0 / 1.

[0129] Step 118 randomly classifies the labeled data set into a training set and a validation set. Usually, the number of validation sets is one third of the number of training sets.

[0130] Step 119 designs training parameters, including input image size, training steps, learning rate, and BatchSize. The image size and BatchSize need to be set according to the video memory of the server graphics card. Generally, the larger the video memory, the larger the image size and BatchSize can be set. The training steps need to be adjusted according to the convergence of the loss. The initial learning rate is the default 0.001, and the learning rate will automatically adjust and decrease as the training progresses.

[0131] Step 120 Load Figure 1 Designed model.

[0132] Step 121 loads the pre-trained model of ResNet, but only imports its valid parameters, and does not import the parameters of the redundant layers.

[0133] Step 122 starts training.

[0134] In step 123, as the loss no longer decreases, you can terminate the training at any time or wait until the training is completed and keep the last model.

[0135] Step 124 randomly selects a video and imports it into the model for testing. If the effect is satisfactory, the model is retained.

[0136] If the result is not satisfactory in step 125 , the unsatisfactory segment images are captured as a new data set and steps 115 to 119 are repeated.

[0137] Step 126 loads the model generated in step 123 as a pre-trained model and continues training.

[0138] Step 127 continues training to generate a new model.

[0139] Step 128 repeats step 124 to test the new model effect. If satisfactory, keep the model. If not satisfactory, repeat the above steps 115 to 127.

[0140] The above is an introduction to a method embodiment. The following further illustrates the solution of the present invention through electronic device and storage medium embodiments.

[0141] The electronic device of the present invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0142] Multiple components in the device are connected to the I / O interface, including: input units, such as keyboards, mice, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as disks, optical disks, etc.; and communication units, such as network cards, modems, wireless communication transceivers, etc. The communication unit allows the device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunication networks.

[0143] The processing unit executes the various methods and processes described above, such as methods S1 to S 16For example, in some embodiments, methods S1 to S 16 It can be implemented as a computer software program, which is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed on the device via the ROM and / or the communication unit. When the computer program is loaded into the RAM and executed by the CPU, the methods S1 to S1 described above can be executed. 16 Alternatively, in other embodiments, the CPU may be configured to execute methods S1 to S1 by any other appropriate means (eg, by means of firmware). 16 .

[0144] The functions described above herein may be performed at least in part by one or more hardware logic components. For example, without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip systems (SOCs), load programmable logic devices (CPLDs), and the like.

[0145] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the processor or controller, enables the functions / operations specified in the flow chart and / or block diagram to be implemented. The program code can be executed entirely on the machine, partially on the machine, partially on the machine as a stand-alone software package and partially on a remote machine, or entirely on a remote machine or server.

[0146] In the context of the present invention, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0147] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for identifying track areas based on a large target semantic segmentation model in a limited scene, characterized in that: The method includes: Step S1, determining a track area recognition algorithm, and using an encoder-decoder structure model to perform semantic segmentation on the image; Step S2, screening the feature backbone network model, and selecting the ResNet residual network as the backbone network prototype architecture; Step S3, reconstruct the backbone network architecture, add the first layer of convolutional network and three groups of Block networks; Step S4: the number of layers of the new backbone network model is 40; Step S5, add an activation layer ReLU after each Block; Step S6, using a dilated convolution method to design the encoder structure; Step S7: Under each branch of the coding region, a Batch Normalization is designed to normalize the outputs of different dilation rate layers; Step S8, after normalizing the data of the three different branches, three activation layers ReLU are added respectively; Step S9, upsampling after adding an activation layer to the first pooling layer branch of the encoder; Step S 10 , connect the outputs of the three branches of the encoder, and then add a convolutional layer as the output of the encoder; Step S 11 , decoder structure design, introducing shallow features in the backbone feature network for splicing; Step S 12 , the shallow features introduced are used as the last set of outputs of the first Block of the backbone feature network, and the output is convolved and parameter normalized; Step S 13 , add activation layer ReLU to the output layer of the shallow network; Step S 14 , the encoder output layer needs to be upsampled after adding the activation layer ReLU; Step S 15 , step S 13 and S 14 The output layer of is connected and an upsampling layer is added, the size of which is the same as the input image, thus restoring the full image size; Step S 16 , step S 13 A convolutional layer is added after the output layer to build a model, which is used to determine whether a pixel is a track area.

2. According to claim 1, a method for identifying track areas based on a large target semantic segmentation model in a limited scene is characterized in that: The first convolutional network layer of step S3 is named conv; The first Block contains three groups; The second Block contains four groups; The third Block contains six groups.

3. The method for identifying track areas based on a large target semantic segmentation model in a limited scene according to claim 2, characterized in that: Each group of the first Block includes the following three layers: The convolution kernel kernel_size is 1*1, the number of convolution kernel filters is 64, the stride is 1*1, the bias is 64, and the convolution layer is named conv2_block1_X_1; The convolution kernel kernel_size is 3*3, the number of convolution kernel filters is 256, the stride is 1*1, the bias is 64, and the convolution layer is named conv2_block1_X_2; The convolution kernel kernel_size is 1*1, the number of convolution kernel filters is 256, the stride is 1*1, the bias is 256, and the convolution layer is named conv2_block1_X_3.

4. According to claim 2, a method for identifying track areas based on a large target semantic segmentation model in a limited scene is characterized in that: Each group of the second Block includes the following three layers: The convolution kernel kernel_size is 1*1, the number of convolution kernel filters is 128, the stride is 2*2, the bias is 128, and the convolution layer is named conv2_block2_X_1; The convolution kernel kernel_size is 3*3, the number of convolution kernel filters is 128, the stride is 1*1, the bias is 128, and the convolution layer is named conv2_block2_X_2; The convolution kernel kernel_size is 1*1, the number of convolution kernel filters is 512, the stride is 1*1, the bias is 512, and the convolution layer is named conv2_block2_X_3.

5. The method for identifying track areas based on a large target semantic segmentation model in a limited scene according to claim 2, characterized in that: Each group of the third Block includes the following three layers: The convolution kernel kernel_size is 1*1, the number of convolution kernel filters is 256, the stride is 2*2, the bias is 256, and the convolution layer is named conv2_block3_X_1; The convolution kernel kernel_size is 3*3, the number of convolution kernel filters is 256, the stride is 1*1, the bias is 256, and the convolution layer is named conv2_block3_X_2; The convolution kernel kernel_size is 1*1, the number of convolution kernel filters is 1024, the stride is 1*1, the bias is 1024, and the convolution layer is named conv2_block3_X_3.

6. The method for identifying track areas based on a large target semantic segmentation model in a limited scene according to claim 1, characterized in that: In step S6, the encoder is designed into a three-branch structure according to the usage environment of the semantic segmentation model, the characteristics that the scene in the tunnel is limited and the objects in the track area are large.

7. The method for identifying track areas based on a large target semantic segmentation model in a limited scene according to claim 6, characterized in that: The three branch structures are as follows: 1) The first branch, the only pooling layer of the encoder module, uses an average pooling layer to take the average value of the image; The pooling window Pool_size is set to 27*50, the pooling operation strides is set to 27*50, and the pooling layer is named Encoder_Average_Pooling; The kernel_size is set to 1*1, the number of filters is 36, the strides is 1*1, the bias is False, and the layer is named Pooling_1x1conv2d; 2) The second branch uses a dilated convolutional layer with a dilation rate of 1; The kernel_size is set to 1*1, the number of filters is 36, the strides is 1*1, the bias is False, and the layer is named Encoder_conv2d_d1; 3) The third branch uses a dilated convolutional layer with a dilation rate of 6; The convolution kernel kernel_size is set to 3*3, the number of convolution kernel filters is 36, the stride is 1*1, the bias is False, and the layer is named Encoder_conv2d_d6.

8. The method for identifying track areas based on a large target semantic segmentation model in a limited scene according to claim 1, characterized in that: The size of the up-sampled data in step S9 is: the input image height is divided by 16 and padded to an integer, and the input image width is divided by 16 and padded to an integer.

9. The method for identifying track areas based on a large target semantic segmentation model in a limited scene according to claim 1, characterized in that: The step S 10 The convolutional layer in is designed as follows: The convolution kernel size is 1*1, the number of convolution kernels is 36, the stride is 1*1, the bias is set to False, and the layer is named Encoder_conv2d_Final.

10. The method for identifying track areas based on a large target semantic segmentation model in a limited scene according to claim 1, characterized in that: The step S 12 The convolutional layer in is designed as follows: The convolution kernel_size is 1*1, the number of convolution kernels filters is 48, the strides of the convolution kernel is 1*1, the bias is False, and the convolution layer is named Decoder_Low_Pre.

11. The method for identifying track areas based on a large target semantic segmentation model in a limited scene according to claim 1, characterized in that: The step S 14 The dimensions of the upsampled data in are: the input image height divided by 4 and padded to an integer, and the input image width divided by 4 and padded to an integer.

12. The method for identifying track areas based on a large target semantic segmentation model in a limited scene according to claim 1, characterized in that: The step S 16 The convolutional layer in is designed as follows: The convolution kernel kernel_size is 1*1, the number of convolution kernel filters is 2, the stride is 1*1, the bias is set to 2, and the network layer is named Con_Decoder_Pre.

13. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 12 is implemented.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 12 is implemented.