Track line segmentation method for high-precision detection of tail end of track

Abstract

The invention discloses a track line segmentation method for high-precision detection of the tail end of a track. The method comprises the following steps: constructing to-be-segmented track line image data; inputting the to-be-segmented track line image data into a track line segmentation network RTEW-Net for processing to obtain a train operation track line segmentation result; wherein the rail line segmentation network RTEW-Net is used for extracting rail line tail end preliminary features from the train operation rail line feature graph, performing feature detail maintenance on the preliminary features, and retaining the rail line tail end features in combination with the train operation rail line feature graph; the track line segmentation network RTEW-Net is obtained through training of a training set, and the training set is a track line image and a corresponding label. According to the invention, accurate segmentation of the track line is realized, and the running safety of the train is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of track line segmentation, and in particular to a track line segmentation method for high-precision detection of track ends. Background Art

[0002] Rail transit is an important part of the modern transportation system, and high-precision segmentation of tracks is a key technology to ensure safe and automatic driving of intelligent trains. Trains run at high speeds, and the end distance within the field of view is reached in a short time. In order to ensure driving safety, high-precision and accurate detection of the end of the railway track is particularly important.

[0003] Currently, due to the performance of data acquisition equipment, the features at the end of the track often degrade, resulting in reduced detection accuracy. In complex actual track operation scenarios, such as the entrance and exit areas of tunnels, the difficulty of feature extraction at the end of the track is further increased due to the rapid changes in lighting. This feature loss makes the existing track segmentation method ineffective at detecting the end of the track, which seriously affects the safety of smart trains.

[0004] Traditional track segmentation methods have achieved certain results in the field of track segmentation, but many studies focus on the visual representation of the entire track, which shows great limitations in the detection of the end in complex scenes. Therefore, although the existing track segmentation methods perform well in common scenes, their performance still has a lot of room for improvement in complex scenes at the end of the track. In response to these challenges, it is particularly important to develop segmentation methods specifically for track end detection in order to further improve the safety of train operation. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a track line segmentation method for high-precision detection of the track end, aiming to solve the problem that traditional track line segmentation methods are difficult to achieve accurate segmentation when faced with the track line end where the features disappear or the features are unclear, and to achieve accurate segmentation of the track line, thereby improving the safety of train operation.

[0006] To achieve the above object, the present invention provides a track line segmentation method for high-precision detection of track ends, comprising:

[0007] Constructing the track line image data to be segmented;

[0008] The track line image data to be segmented is input into the track line segmentation network RTEW-Net for processing to obtain the train running track line segmentation result;

[0009] Among them, the track line segmentation network RTEW-Net is used to extract preliminary features of the track line end from the train running track line feature map, maintain feature details of the preliminary features, and retain the track line end features in combination with the train running track line feature map; the track line segmentation network RTEW-Net is obtained through training with a training set, and the training set is a track line image and corresponding annotations.

[0010] Preferably, constructing the training set includes:

[0011] The initial track line images in the real train operation scenario are collected, the initial track line images are manually annotated and screened, and the screened images are combined with the initial track line images to construct the RailMixed2024 dataset, i.e., the training set; wherein the real train operation scenario includes image data in which the end features of the track line disappear or are unclear due to entering and exiting tunnels and extremely severe weather conditions.

[0012] Preferably, the structure of the track segmentation network RTEW-Net is an encoding-decoding structure, the encoder structure includes a full attention calculation module FTM, a global response normalization module GRN and an intelligent weight maintenance module WWM connected in sequence, and the decoder is an Uperhead decoder.

[0013] Preferably, inputting the track line image data to be segmented into the track line segmentation network RTEW-Net for processing comprises:

[0014] The track line image data to be segmented is subjected to feature dimension upgrading through the Patch Embedding method to obtain a feature map, and the size of the feature map is normalized;

[0015] The normalized feature map is subjected to attention calculation by the full attention calculation module FTM, and the feature map is aggregated with global features by the global response normalization module GRN, and recalibrated to output the normalized feature map;

[0016] The feature map processed by the GRN is combined with the normalized feature map through an intelligent weight maintenance module WWM to output the train track segmentation result.

[0017] Preferably, the track line image data to be segmented is subjected to feature dimension upgrading by a Patch Embedding method to obtain a feature map, comprising:

[0018] Based on the track segmentation network RTEW-Net, the track image data to be segmented is downsampled by deep features through PatchEmbedding operation and Full Transformer Block calculation, and downsampled through PatchMerging, and this process is repeated and iterated until the feature map is output;

[0019] Among them, each downsampling operation combined with Block calculation constitutes an operation stage. During the Block calculation process, LN operation, FTM module calculation, GELU activation and MLP classifier calculation are performed in sequence.

[0020] Preferably, the attention calculation is performed by the full attention calculation module FTM, including: Base-patch calculation and pad-patch calculation;

[0021] The method for calculating Base-patch is:

[0022] FTM 1 (x)=ReverseSeparate(attn(Separate(x)))

[0023] The pad-patch calculation method is:

[0024] FTM 2 (x) = RmPad(FTM 1 (Pad(x)))

[0025] Where Pad and RmPad represent the operations of expanding and removing padding according to the patch shape, Separate and ReverseSeparate represent the operations of splitting the feature map into patches and restoring them, attn represents the execution of attention calculation, and FTM 1 (x) is the calculation result of Base-patch, FTM 2 (x) is the pad-patch calculation result, x is the input feature Figure 1 ;

[0026] In a block calculation, the base-patch calculation is performed first, and then the pad-patch calculation is performed.

[0027] Preferably, the method for the global response normalization module GRN to aggregate global features of the feature map and recalibrate is:

[0028] X i+1 =γ*N(X i )+β+X i

[0029] In the formula, γ and β are trainable parameters, X i+1 Output feature Figure 1 ,N(X i ) is the regularization method, X i is the input feature map.

[0030] Preferably, the intelligent weight maintenance module WWM retains the normalized feature map information and the processed feature map information through an adaptively adjusted weight maintenance method, and sets adaptive parameters participating in global optimization to organically combine the normalized feature map information and the processed feature map information, so as to realize the model's adaptive retention and adjustment of the detail features in the feature map.

[0031] Preferably, training the track line segmentation network RTEW-Net comprises:

[0032] The parameters of RTEW-Net are optimized through the cross entropy loss function and the adamW optimizer, and the intelligent residual link operation is performed to obtain the trained track segmentation network RTEW-Net.

[0033] Preferably, the method for performing the intelligent residual link operation is to add trainable weights to the calculated feature map and optimize it during the network training process, specifically:

[0034] x out =ω 1 *x 1 +ω 2 *x 2

[0035] In the formula, x out Output feature Figure 2 ,ω 1 and ω 2 are trainable parameters that participate in global optimization independently, x 1 and x 2 Feature map information after processing by different modules.

[0036] Compared with the prior art, the present invention has the following advantages and technical effects:

[0037] The present invention can effectively extract the missing or unclear features at the end of the track line, maintain the depth, complete the accurate segmentation of the end of the track line, and has good generalization, thereby completing the complete and accurate perception of the train running track and ensuring the safety of train operation.

[0038] The present invention improves the calculation effect of traditional attention on subtle features by designing a full-transformer module (FTM), and realizes the preliminary retention of the feature information of the end of the track line; introduces a global normalization response module (GRN) to aggregate global features and recalibrate them to achieve feature normalization output, improve feature detail maintenance, and cope with complex track scenes and drastic lighting changes; designs an intelligent weight maintenance module (Wise Weigh Maintain (WWM), which can adaptively adjust the intensity of feature learning at each layer, realize the deep retention of the unclear feature details at the end of the track line, and deepen the model's detection ability for the end of the track line. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0040] Figure 1 This is a display diagram of the RM2024 dataset track image and its labels according to an embodiment of the present invention;

[0041] Figure 2 This is a display diagram of the RS19 dataset track image and its labels according to an embodiment of the present invention;

[0042] Figure 3 This is a flow chart of track line segmentation by RTEW-Net according to an embodiment of the present invention;

[0043] Figure 4 This is the overall structure diagram of RTEW-Net in an embodiment of the present invention;

[0044] Figure 5 It is a calculation flow chart of the full attention calculation module of an embodiment of the present invention;

[0045] Figure 6 It is a schematic diagram showing the prediction results of RTEW-Net and SOTA model on the RM2024 dataset according to an embodiment of the present invention;

[0046] Figure 7 This is a visualization diagram of the prediction results of the RTEW-Net and SOTA models of an embodiment of the present invention on the RS19 dataset. DETAILED DESCRIPTION

[0047] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0048] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0049] The present invention proposes a track line segmentation method for high-precision detection of track ends, comprising:

[0050] Constructing the track line image data to be segmented;

[0051] The track line image data to be segmented is input into the track line segmentation network RTEW-Net for processing to obtain the train running track line segmentation result;

[0052] Among them, the track segmentation network RTEW-Net is used to extract preliminary features of the track end from the train running track feature map, maintain the feature details of the preliminary features, and retain the track end features in combination with the train running track feature map; the track segmentation network RTEW-Net is obtained by training with a training set, and the training set is a track line image and corresponding annotations.

[0053] Furthermore, the structure of the track segmentation network RTEW-Net is an encoding-decoding structure. The encoder structure includes a full attention calculation module FTM, a global response normalization module GRN and an intelligent weight maintenance module WWM connected in sequence, and the decoder is an Uperhead decoder.

[0054] Specifically, this embodiment proposes a track line segmentation network (RTEW-Net) for high-precision detection of track ends. The network model integrates the full attention calculation module (FTM), the global response normalization module (GRN) and the intelligent weight maintenance module (WWM) in sequence. The three modules are connected in series through a linear calculation process, which effectively improves the performance of the algorithm, enables the method to adapt to complex track environments, and has a good recognition effect on the track end, realizes accurate segmentation of the track line, and ensures the safety of train operation.

[0055] Furthermore, a training set is constructed, including:

[0056] The initial track line images in the real train operation scenario are collected, the initial track line images are manually annotated and screened, and the screened images are combined with the initial track line images to construct the RailMixed2024 (RM2024) dataset, i.e., the training set; wherein the real train operation scenario includes image data in which the end features of the track line disappear or are unclear due to entering and exiting tunnels and extremely severe weather conditions.

[0057] Specifically, refer to Figure 1 In this embodiment, track line image data under real train operation scenes are collected. The scenes are complex and rich, including image data where track line end features disappear or are unclear due to situations such as entering and exiting the track and extremely bad weather. The collected images are manually annotated. Figure 2 The RS19 dataset is displayed, and some images are filtered out. After processing, they are combined with the collected images to construct the RailMixed2024 dataset. There are 2142 images collected in the field, and 800 images are filtered out from the RS19 dataset, totaling 2942 track images, which are divided into training set and test set in a ratio of 9:1.

[0058] Furthermore, the track line image data to be segmented is input into the track line segmentation network RTEW-Net for processing, including:

[0059] The track line image dataset is subjected to feature dimension upgrading through the Patch Embedding method to obtain feature maps, and the size of the feature maps is normalized.

[0060] The normalized feature map is subjected to attention calculation by the full attention calculation module FTM, and the feature map is aggregated with global features by the global response normalization module GRN, and recalibrated to output the normalized feature map;

[0061] The feature map processed by the GRN is combined with the normalized feature map through an intelligent weight maintenance module WWM to output the train track segmentation result.

[0062] Specifically, Figure 3 , This embodiment proposes RTEW-Net, the overall structure of which is an encoding-decoding structure. In this embodiment, Uperhead is selected as the decoder. Uperhead is a general decoder that can be used in conjunction with a specific encoder to achieve feature fusion and result prediction. The encoder structure is designed through the comprehensive design of FTM, GRN and WWM to achieve efficient feature extraction of the track line. In the encoder part, the image is downsampled to obtain a high-dimensional feature map. Before entering the encoder, the image undergoes a reconstruction operation, and the pixel value is adjusted to between 0 and 1 and conforms to the normal distribution state.

[0063] Furthermore, the track line image dataset is subjected to feature dimension upgrading through the Patch Embedding method to obtain a feature map, including:

[0064] Based on the track segmentation network RTEW-Net, the track image data to be segmented is downsampled by deep features through PatchEmbedding operation and Full Transformer Block (FTBlock) calculation, and downsampled by Patch Merging, and this process is repeated and iterated until the feature map is output;

[0065] Among them, each downsampling operation combined with Block calculation constitutes an operation stage. During the Block calculation process, LN operation, FTM module calculation, GELU activation and MLP classifier calculation are performed in sequence.

[0066] Specifically, refer to Figure 4 , the encoder of the RTEW-Net proposed in this embodiment is a hierarchical structural design as a whole. In the first calculation, the image is downsampled by deep features through Patch Embedding operation and Full Transformer Block calculation, and then downsampled through Patch Merging. This process is iterated four times. Each downsampling operation is combined with Block calculation to form an operation stage. During the Block calculation process, LN operation, FTM module calculation, GELU activation and MLP classifier calculation are performed in sequence. The MLP module is derived from ViT and Swin, and compensates for the local calculation problems inherent in the attention mechanism through fully connected operations. During a single Block calculation process, the operation of the feature map is as follows:

[0067] x stage+1 =FTBlock(DownSampling(x stage ))

[0068] In the formula, x stage+1 Output features after a single Block calculation is completed Figure 3 , DownSampling is a downsampling operation, in the first layer of Block calculation it is a Patch Embedding operation, and in all other cases it is a Patch Merging operation, x stage For input features Figure 2 ;

[0069] For FTBlock, a complete calculation process is as follows:

[0070] x i+1 =ω 1 *GRN(GELU(FTM 1 (LN(x i ))))+ω 2 *x i

[0071] x i+2 =ω 3 *MLP(LN(x i+1 ))+ω 4 *x i+1

[0072] x i+3 =ω 5 *GRN(GELU(FTM 2 (LN(x i+2 ))))+ω 6 *x i+2

[0073] x i+4 =ω 7 *MLP(LN(x i+3 ))+ω 8 *x i+3

[0074] In the formula, x i For input features Figure 3 , x i+1 、x i+2 、x i+3 are the output feature maps of their respective layers and the input feature maps calculated by the next layer, respectively. 1 to 8 are trainable parameters that participate in global optimization independently, FTM 1 like Figure 5 As shown, it is the mathematical implementation of the Base-Patch module, FTM 2 like Figure 5 As shown in the figure, it is the mathematical implementation of the Pad-Patch module, GRN is the global response normalization module, MLP is the fully connected calculation method, GELU is a common activation function of neural networks, and LN is a common regularization operation of neural networks.

[0075] For an input image of size H×W×3, the Patch embedding operation first changes its shape to After each stage operation, the height and width of the feature map are halved and the number of channels is doubled. The final size is 8×C.

[0076] Furthermore, attention calculation is performed through the full attention calculation module FTM, including base-patch calculation and pad-patch calculation; in order to maintain and settle symbol consistency and facilitate understanding of the structure of base-patch and pad-patch.

[0077] The method for calculating Base-patch is:

[0078] FTM 1(x)=ReverseSeparate(attn(Separate(x)))

[0079] The pad-patch calculation method is:

[0080] FTM 2 (x) = RmPad(FTM 1 (Pad(x)))

[0081] Where Pad and RmPad represent the operations of expanding and removing padding according to the patch shape, Separate and ReverseSeparate represent the operations of splitting the feature map into patches and restoring them, attn represents the execution of attention calculation, and FTM 1 (x) is the calculation result of Base-patch, FTM 2 (x) is the pad-patch calculation result;

[0082] In a block calculation, the base-patch calculation is performed first, and then the pad-patch calculation is performed.

[0083] Specifically, refer to Figure 5 ,The FTM module is the core computing module of the feature extraction network, in which the main feature extraction operations are implemented. The design of this module is inspired by ViT, Swin Transformer, MobileViT, LPT and their variants. The basic goal is to better extract the overall features of the target and achieve global attention interaction at a low computational cost.

[0084] for Figure 5 In the Base-patch calculation method, the feature map is initially divided into patches as the basic calculation unit according to preset parameters, and attention calculation is performed. After the divided and calculated patches are deeply connected, each individual patch is regarded as a sequence unit similar to the NLP task and input into the Transformer for calculation. After the calculation is completed, the feature map is restored according to the inverse step of segmentation.

[0085] Next is the pad-patch calculation method. Based on the size of the divided patch, the feature map is expanded by half the patch size in four directions. The expanded image is re-divided according to the patch size and subsequent attention calculations are performed like Base-patch. After completing the Transformer calculation, the feature map is restored to its original shape and the padding elements in the four directions are removed to obtain the restored feature map.

[0086] Furthermore, the global response normalization module GRN aggregates global features of the feature map and recalibrates it as follows:

[0087] X i+1 =γ*N(X i )+β+X i

[0088] In the formula, γ and β are trainable parameters, X i+1 Output feature Figure 1 , X i is the input feature map, N(x) is a regularization method, and is defined as follows:

[0089]

[0090] In the formula, H and W represent the height and width of the feature map respectively, and ∈ is a small real number, x ij Represents the pixel at coordinate (i, j) in the feature map.

[0091] Specifically, the global response normalization (GRN) module aims to enhance channel contrast and selectivity. This embodiment uses GRN for feature map normalization after attention calculation, not before MLP calculation, but after FTM module calculation.

[0092] Furthermore, the intelligent weight maintenance module WWM retains the normalized feature map information and the processed feature map information through an adaptive weight maintenance method, and sets the adaptive parameters participating in the global optimization to organically combine the normalized feature map information and the processed feature map information, so as to realize the model's adaptive retention and adjustment of the detailed features in the feature map.

[0093] Specifically, the intelligent weight maintenance module (WWM) designs a weight maintenance method that can be adaptively adjusted to retain the feature map information before and after the calculation, and sets the adaptive parameters participating in the global optimization to organically combine the two feature maps, thereby allowing the model to adaptively retain and adjust the detailed features in the feature map, deepen the sensitivity to subtle features, and achieve accurate segmentation of the end of the track line.

[0094] Furthermore, the track segmentation network RTEW-Net is trained, including:

[0095] The parameters of RTEW-Net are optimized through the cross entropy loss function and the adamW optimizer, and the intelligent residual link operation is performed to obtain the trained track segmentation network RTEW-Net.

[0096] The method of performing the intelligent residual link operation is to add trainable weights to the calculated feature map and optimize it during the network training process, specifically:

[0097] x out =ω 1 *x 1 +ω 2 *x 2

[0098] In the formula, x out Output feature Figure 2 ,ω 1 and ω 2 are trainable parameters that participate in global optimization independently, x 1 and x 2 Feature map information after processing by different modules.

[0099] Specifically, the intelligent residual link operation aims to improve the model's ability to retain sparse features, thereby enhancing the extraction of features that disappear or are unclear at the end of the track line. The traditional residual link operation inevitably reduces the model's feature learning ability to some extent. This reduction occurs when each newly learned feature is added to the previously learned feature, resulting in a weakening of the strength of the newly learned feature.

[0100] In order to solve this problem and enhance the learning ability of the model, this embodiment adds trainable weights to the feature graph involved in the calculation and optimizes them during the network training process.

[0101] This operation enables the model to adaptively adjust the strength of feature learning at each layer, thereby improving the model's learning ability at the end of the track while maintaining the stability advantage of the original residual connection.

[0102] Reference Figure 6 This example shows a visual comparison of the method of the present invention and other SOTA methods on the RM2024 dataset. The comparison methods include Swin Transfrormer, Vision Transformer, Lraspp, BiFormer, ConvNeXtV2 and RTEW-Net of the present invention.

[0103] In the first four rows of railway images, the railway terminus is disturbed by different degrees of environment: the first row is affected by the highlight at the tunnel exit, the fourth row is affected by the shadow in the middle of the tunnel, and the second and third rows are affected by lens pollution. These factors lead to the degradation of railway terminus features. RTEW-Net effectively reconstructs these features using the FTM module and achieves better railway terminus segmentation. In contrast, the Swin model has difficulties in segmenting the railway terminus on the right side of the first row, has problems with the large-scale objects on the right side of the second row and the railway terminus on the left side, and performs poorly on the railway terminus on the left side of the third row. ViT and Lraspp perform well on some images, but perform poorly in complex scenes. BiFormer and ConvNeXt accurately segment and locate the railway position, but the overall segmentation accuracy is low and the terminus is almost completely failed to segment. The fifth row shows a typical train running scene. Swin still performs poorly on the segmentation of the left railway terminus, while only RTEW-Net accurately segments the terminus, demonstrating the effectiveness of the RTEW-Net module design.

[0104] The specific comparison indicators are shown in Table 1, which shows the main test targets and mIoU data of rs19. These targets are annotated using an improved semi-automatic method, and the training results are better than other categories. As shown in Table 1, RTEW-Net achieved the best results on the railway-related targets of rs19, and the IoU of the most critical target rail track was improved by 3.2 compared with the second best model. The IoU of rail raised and rail embedded increased by 4.5 and 6.6 respectively. The overall mIoU increased by 3.6, showing a significant advantage over other models. This shows that RTEW-Net has superior feature extraction capabilities for railway-related targets and verifies the effectiveness of the proposed algorithm.

[0105] Reference Figure 7 , showing the performance of different models in railway terminal segmentation. Figure 7In Figure 1, the first five rows show the inference results of the test set, and the last four rows show the results of the training set. The first column is the original acquired image, the second column is the ground truth, the third column is the result of RTEW-Net, and the remaining columns are the inference results of the comparison models. Visually, RTEW-Net provides the best segmentation results, almost completely segmenting all railway targets. In the first four rows of railway images, the railway terminus is affected by different degrees of environmental interference: the first row is affected by the highlight at the tunnel exit, the fourth row is affected by the shadow in the middle of the tunnel, and the second and third rows are affected by lens contamination. These factors lead to the degradation of the railway terminus features. RTEW-Net effectively reconstructs these features using the FTM module and achieves better railway terminus segmentation. In contrast, the Swin model has difficulty in segmenting the railway terminus on the right side of the first row, has problems with the large-scale objects on the right side of the second row and the railway terminus on the left side, and performs poorly on the railway terminus on the left side of the third row. ViT and Lraspp perform well on some images, but perform poorly in complex scenes. BiFormer and ConvNeXt accurately segment and locate the railway position, but the overall segmentation accuracy is low and the terminus is almost completely failed to segment. The fifth row shows a typical train running scene. Swin still performs poorly on the segmentation of the left railway terminal, while only RTEW-Net accurately segments the terminal, demonstrating the effectiveness of the RTEW-Net module design.

[0106] The specific comparison indicators for the rs19 dataset are shown in Table 2, which shows the main test targets and mIoU data of rs19. RTEW-Net achieved the best results on the railway-related targets of rs19, and the IoU of the most critical target rail track was improved by 3.2 compared with the second best model. The IoU of rail raised and rail embedded were improved by 4.5 and 6.6 respectively. The overall mIoU was improved by 3.6, showing a significant advantage over other models. This shows that RTEW-Net has superior feature extraction capabilities on railway-related targets and verifies the effectiveness of the proposed algorithm.

[0107] Table 1

[0108]

[0109] Table 2

[0110]

[0111] The comprehensive module performance of the present invention can adapt to complex track operation scenarios and effectively extract new messages at the track end with unclear characteristics in extreme cases, thereby achieving accurate segmentation of the track line and ensuring the safety of train travel, which has high practical value.

[0112] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A track line segmentation method for high-precision detection of track ends, characterized in that: include: Constructing the track line image data to be segmented; The track line image data to be segmented is input into the track line segmentation network RTEW-Net for processing to obtain the train running track line segmentation result; Among them, the track line segmentation network RTEW-Net is used to extract preliminary features of the track line end from the train running track line feature map, maintain feature details of the preliminary features, and retain the track line end features in combination with the train running track line feature map; the track line segmentation network RTEW-Net is obtained through training with a training set, and the training set is a track line image and corresponding annotations.

2. The track line segmentation method for high-precision detection of track end according to claim 1, characterized in that: Constructing the training set includes: The initial track line images in the real train operation scenario are collected, the initial track line images are manually annotated and screened, and the screened images are combined with the initial track line images to construct the RailMixed2024 dataset, i.e., the training set; wherein the real train operation scenario includes image data in which the end features of the track line disappear or are unclear due to entering and exiting tunnels and extremely severe weather conditions.

3. The track line segmentation method for high-precision detection of track end according to claim 1, characterized in that: The structure of the track segmentation network RTEW-Net is an encoding-decoding structure, the encoder structure includes a full attention calculation module FTM, a global response normalization module GRN and an intelligent weight maintenance module WWM connected in sequence, and the decoder is an Uperhead decoder.

4. The track line segmentation method for high-precision detection of track end according to claim 3, characterized in that: Inputting the track line image data to be segmented into the track line segmentation network RTEW-Net for processing includes: The track line image data to be segmented is subjected to feature dimension upgrading through the Patch Embedding method to obtain a feature map, and the size of the feature map is normalized; The normalized feature map is subjected to attention calculation by the full attention calculation module FTM, and the feature map is aggregated with global features by the global response normalization module GRN, and recalibrated to output the normalized feature map; The feature map processed by the GRN is combined with the normalized feature map through an intelligent weight maintenance module WWM to output the train track segmentation result.

5. The track line segmentation method for high-precision detection of track end according to claim 4, characterized in that: The track line image data to be segmented is subjected to feature dimension upgrading by using a Patch Embedding method to obtain a feature map, including: Based on the track line segmentation network RTEW-Net, the track line image data to be segmented is downsampled by deep features through Patch Embedding operation and Full Transformer Block calculation, and downsampled through Patch Merging, and this process is repeated and iterated until the feature map is output; Among them, each downsampling operation combined with Block calculation constitutes an operation stage. During the Block calculation process, LN operation, FTM module calculation, GELU activation and MLP classifier calculation are performed in sequence.

6. The track line segmentation method for high-precision detection of track end according to claim 5, characterized in that: Attention calculation is performed through the full attention calculation module FTM, including: Base-patch calculation and pad-patch calculation; The method for calculating Base-patch is: FTM1(x)=ReverseSeparate(attn(Separate(x))) The pad-patch calculation method is: FTM2(x)=RmPad(FTM1(Pad(x))) Where Pad and RmPad represent the operations of expanding and removing padding according to the patch shape, respectively; Separate and ReverseSeparate represent the operations of splitting the feature map into patches and restoring them, respectively; attn represents the execution of attention calculation; FTM1(x) is the result of Base-patch calculation; FTM2(x) is the result of pad-patch calculation; x is the input feature map 1; In a block calculation, the base-patch calculation is performed first, and then the pad-patch calculation is performed.

7. The track line segmentation method for high-precision detection of track end according to claim 4, characterized in that: The method for the global response normalization module GRN to aggregate global features of the feature map and recalibrate is as follows: X i+1 =γ*N(X i )+β+X i In the formula, γ and β are trainable parameters, X i+1 is the output feature map 1, N(X i ) is the regularization method, X i is the input feature map.

8. The track line segmentation method for high-precision detection of track end according to claim 3, characterized in that: The intelligent weight maintenance module WWM retains the normalized feature map information and the processed feature map information through an adaptive weight maintenance method, and sets adaptive parameters participating in global optimization to organically combine the normalized feature map information and the processed feature map information, thereby realizing the model's adaptive retention and adjustment of detail features in the feature map.

9. The track line segmentation method for high-precision detection of track end according to claim 1, characterized in that: Training the track segmentation network RTEW-Net includes: The parameters of RTEW-Net are optimized through the cross entropy loss function and the adamW optimizer, and the intelligent residual link operation is performed to obtain the trained track segmentation network RTEW-Net.

10. The track line segmentation method for high-precision detection of track end according to claim 9, characterized in that: The method of performing the intelligent residual link operation is to add trainable weights to the calculated feature map and optimize it during the network training process, specifically: x out =ω1*x1+ω2*x2 In the formula, x out To output feature map 2, ω1 and ω2 are trainable parameters that participate in global optimization independently, and x1 and x2 are feature map information processed by different modules.

Images