A railway track segmentation method, system, computer device and storage medium
By introducing polarization imaging technology and multimodal feature fusion into railway track segmentation, and using the Stokes polarization model to calculate polarization degree and angle images, combined with the U-Net network, the robustness and accuracy problems of railway track segmentation in complex environments are solved, achieving high-precision and high-robust railway track segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)
- Filing Date
- 2026-04-07
- Publication Date
- 2026-06-26
AI Technical Summary
Existing railway track segmentation methods lack accuracy in complex environments, especially under conditions of backlighting, metamerism, and low contrast. This results in poor robustness of the segmentation model, an inability to effectively distinguish between metallic and dielectric materials, and specular reflection and color deepening leading to highlight overflow, causing discontinuities in the segmentation mask.
Using four light intensity images with different polarization directions, the linear polarization degree and polarization angle images are calculated using the Stokes polarization model. A three-dimensional fusion feature tensor is constructed and trained using a U-Net network to enhance the material contrast and geometric features between the track and the background, thus enabling railway track segmentation.
It significantly improves the robustness and accuracy of railway track segmentation, enabling accurate differentiation between track and background in complex environments, reducing false alarm rate, ensuring the integrity and connectivity of segmentation results, and enhancing the model's segmentation performance in environments with strong reflection and low illumination.
Smart Images
Figure CN122289689A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision and railway safety monitoring technology, specifically relating to a railway track segmentation method, system, computer equipment, and storage medium. Background Technology
[0002] Currently, automatic track segmentation and extraction technology is a key foundational component for automatic train operation, intrusion detection, and driver assistance systems. Existing track segmentation methods mostly employ industrial cameras to acquire two-dimensional visible light intensity images. By training neural network models (CNN, U-Net) to learn the color, texture, and geometric features of the track in the visible light intensity images, track pixels are separated from ballast, sleepers, and the surrounding environment.
[0003] While visible light intensity image segmentation methods based on neural network models achieve good results under ideal conditions, in the complex real-world railway operating environment, the smooth metallic surface of the rails is highly susceptible to specular reflection (highlight overflow) under backlight, direct sunlight, or nighttime headlight illumination. Additionally, after rain or in damp environments, puddles or wet ballast surfaces near the rails also produce reflections and darken in color. When specular reflection or color darkening occurs, the pixel values of the track area in the visible light intensity image reach saturation, causing the texture features of the track itself to disappear and making it difficult to distinguish from equally bright environmental noise. This prevents the segmentation algorithm from extracting complete track edges, resulting in broken or incomplete track masks, severely impacting subsequent navigation or obstacle detection.
[0004] In addition, the railway environment is extremely complex. The acquired visible light intensity images may contain ballast (gravel), sleepers, weeds, and even mud and water. When the color of rusty rails is dark brown, it is very similar to the color and grayscale of ballast, sleepers, or dry vegetation (i.e., "weak texture" or "same color" phenomenon). This low contrast makes it difficult for traditional segmentation algorithms to determine accurate segmentation edges. Consequently, the segmentation model may easily confuse the track edge with the ballast edge, resulting in inaccurate track width or severe jagged edges.
[0005] Because the visible light intensity image data sources relied upon by existing technologies only contain information on light intensity and color, they cannot reflect the material properties and geometric reflection characteristics of object surfaces. This results in segmentation models trained on this basis lacking the ability to distinguish between metallic and dielectric materials, the physical understanding of specular reflection, and the enhanced perception of geometric edges under low illumination. Consequently, segmentation accuracy decreases significantly in complex environments such as strong reflections, color interference, and low light, and robustness is insufficient. Summary of the Invention
[0006] To address the aforementioned problems, this invention provides a railway track segmentation method, system, computer equipment, and storage medium.
[0007] To achieve the above objectives, the present invention provides a railway track segmentation method, comprising: The first, second, third, and fourth light intensity images of the target railway track were simultaneously acquired from four different polarization directions.
[0008] Based on the Stokes polarization model, four light intensity images are calculated to obtain a first Stokes parameter image, a second Stokes parameter image, and a third Stokes parameter image. The pixel value of the first Stokes parameter image is half the sum of the corresponding pixel values of the four light intensity images. The second Stokes parameter image is the difference between the corresponding pixel values of the first and third light intensity images. The third Stokes parameter image is the difference between the corresponding pixel values of the second and fourth light intensity images. Using the first, second, and third Stokes parameter images, a linear polarization degree image and a polarization angle image are determined. The first Stokes parameter image, the linear polarization degree image, and the polarization angle image are then fused by channel fusion to obtain a three-dimensional fused feature tensor.
[0009] Region labeling is performed based on the first Stokes parameter image to obtain a binary mask image of the target track region; the U-Net network is trained with the three-dimensional fusion feature tensor as input and the binary mask image as the ground truth label to obtain a model for segmenting the target railway track.
[0010] Preferably, the polarization direction is a linear polarization direction, the first light intensity image is a light intensity image with a 0° linear polarization direction, the second light intensity image is a light intensity image with a 45° linear polarization direction, the third light intensity image is a light intensity image with a 90° linear polarization direction, and the fourth light intensity image is a light intensity image with a 135° linear polarization direction.
[0011] Preferably, the first Stokes parameter image represents the total light intensity of the four light intensity images; the second Stokes parameter image represents the intensity difference between the horizontal polarization direction and the vertical polarization direction; and the third Stokes parameter image represents the intensity difference between two oblique polarization components whose polarization directions differ by 90°.
[0012] Preferably, before performing channel fusion of the first Stokes parameter image, the linear polarization degree image, and the polarization angle image to obtain a three-dimensional fused feature tensor, the method further includes: normalizing the first Stokes parameter image and mapping its pixel values to the [0,1] interval; normalizing the polarization angle image and mapping its pixel values to the [0,1] interval; and fusing the normalized first Stokes parameter image, the polarization angle image, and the linear polarization degree image.
[0013] Preferably, the trained U-Net network outputs a pixel-level probability map of the polarization image of the railway track to be segmented, the pixel-level probability map is thresholded to obtain a binary segmentation mask for the railway track, and the binary segmentation mask is morphologically optimized to obtain the final track segmentation result.
[0014] Preferably, based on the Stokes polarization theory derivation formula, the linear polarization degree image and the polarization angle image are determined using the first Stokes parameter image, the second Stokes parameter image, and the third Stokes parameter image; the linear polarization degree image is used to enhance the material contrast between the track and the background, and the polarization angle image is used to enhance the geometric features of the track edge.
[0015] Preferably, the binary segmentation mask is morphologically optimized by sequentially performing a closing operation to fill the voids inside the mask and an opening operation to smooth the jagged edges of the mask.
[0016] The present invention also provides a railway track segmentation system, comprising: The data acquisition module is used to simultaneously acquire the first, second, third, and fourth light intensity images of the target railway track from four different polarization directions.
[0017] The calculation module is used to calculate four light intensity images based on the Stokes polarization model to obtain a first Stokes parameter image, a second Stokes parameter image, and a third Stokes parameter image. The pixel value of the first Stokes parameter image is half of the sum of the corresponding pixel values of the four light intensity images. The second Stokes parameter image is the difference between the corresponding pixel values of the first and third light intensity images. The third Stokes parameter image is the difference between the corresponding pixel values of the second and fourth light intensity images. Using the first, second, and third Stokes parameter images, the linear polarization degree image and the polarization angle image are determined. The first Stokes parameter image, the linear polarization degree image, and the polarization angle image are fused by channel fusion to obtain a three-dimensional fused feature tensor.
[0018] The training module is used to perform region labeling based on the first Stokes parameter image to obtain a binary mask image of the target track region; the U-Net network is trained with the three-dimensional fusion feature tensor as input and the binary mask image as ground truth label to obtain a model for segmenting the target railway track.
[0019] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the steps in the railway track segmentation method.
[0020] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute any of the steps in the railway track segmentation method.
[0021] The railway track segmentation method provided by this invention has the following beneficial effects: This invention creatively constructs a multi-physical-dimensional supervised signal dataset by labeling track regions on a full-intensity image (i.e., a first Stokes parameter image) and strictly aligning it with a three-dimensional fusion tensor containing polarization physical features (linear polarization degree and polarization angle). This invention upgrades the traditional image-label mapping, which only contains intensity and color information, to a physical feature-label mapping. This allows deep learning models to directly learn the essential relationship between track regions and polarization physical properties during training, such as the metallic material represented by high linear polarization degree and the geometric edges represented by abrupt changes in polarization angle, rather than merely learning unstable apparent color and texture patterns. Therefore, this dataset fundamentally provides the model with the physical criteria needed to solve the three major challenges of metamerism, strong reflection, and blurred edges in weak light, significantly improving the model's generalization ability and segmentation robustness in complex real-world environments. Attached Figure Description
[0022] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0023] Figure 1 This is a flowchart of a railway track segmentation method according to an embodiment of the present invention; Figure 2 This is a first light intensity image according to an embodiment of the present invention; Figure 3 This is the second light intensity image according to an embodiment of the present invention; Figure 4 This is the third light intensity image according to an embodiment of the present invention; Figure 5 This is the fourth light intensity image according to an embodiment of the present invention; Figure 6 This is the first Stokes parameter image of an embodiment of the present invention; Figure 7 This is the second Stokes parameter image according to an embodiment of the present invention; Figure 8 This is the third Stokes parameter image of an embodiment of the present invention; Figure 9 This is a linear polarization degree image according to an embodiment of the present invention; Figure 10This is a polarization angle image according to an embodiment of the present invention; Figure 11 This refers to the three-dimensional fusion feature tensor in this embodiment of the invention. Detailed Implementation
[0024] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.
[0025] This invention provides a method for dividing railway tracks, specifically as follows: Figure 1 As shown, it includes: S1. Simultaneously acquire the first, second, third, and fourth light intensity images of the target railway track from four different polarization directions.
[0026] Acquire multi-channel raw polarization images of a railway scene. Use a polarization image acquisition module (such as a focal plane polarization camera) to capture images of the scene in front of the railway track. In single exposures or continuous acquisitions, simultaneously acquire light intensity response images of the same scene in different polarization directions. Specifically, acquire linearly polarized light intensity images at four specific angles, namely the polarization component images at 0°, 45°, 90°, and 135°, denoted as I0, I... 45 I 90 and I 135 These images form the basic data source for subsequent processing.
[0027] S2. Based on the Stokes polarization model, four light intensity images are calculated to obtain a first Stokes parameter image, a second Stokes parameter image, and a third Stokes parameter image. The pixel value of the first Stokes parameter image is half the sum of the corresponding pixel values of the four light intensity images. The second Stokes parameter image is the difference between the corresponding pixel values of the first and third light intensity images. The third Stokes parameter image is the difference between the corresponding pixel values of the second and fourth light intensity images. Using the first, second, and third Stokes parameter images, the linear polarization degree image and the polarization angle image are determined. The first Stokes parameter image, the linear polarization degree image, and the polarization angle image are fused by channel fusion to obtain a three-dimensional fused feature tensor.
[0028] The Stokes vector and polarization physical characteristic parameters are calculated. The acquired raw image data is processed, and the complete polarization state of the light wave is calculated based on the Stokes vector method. First, the Stokes vectors S0, S1, and S2 are calculated using the light intensity components in each direction. S0 is the first Stokes parameter image, S1 is the second Stokes parameter image, and S2 is the third Stokes parameter image. Here, S0 represents the total light intensity (i.e., half the sum of the pixel values of the four light intensity images in a conventional visual image), S1 represents the intensity difference between linearly polarized light in the horizontal and vertical directions, and S2 represents the intensity difference between linearly polarized light in the 45° and 135° directions. The calculation formula is as follows:
[0029] ; ; ; Subsequently, based on the Stokes vector, two key parameter images reflecting the physical properties of the object are further calculated: the degree of linear polarization image (DoLP) and the angle of polarization image (AoP). The degree of linear polarization characterizes the proportion of polarized components in the reflected light and can effectively distinguish between metallic tracks and scattered ballast. The calculation formula is as follows:
[0030] The polarization angle characterizes the vibration direction of the polarization vector of reflected light. This direction has a strict geometric correspondence with the normal direction of the object's surface, and can keenly detect geometric abrupt changes at the edge of the orbit. The calculation formula is:
[0031] A multimodal polarization feature fusion tensor is constructed. To combine the texture advantages of intensity images with the physical advantages of polarization images, the total intensity image (S0), linear polarization degree image (DoLP), and polarization angle image (AoP) are registered and normalized. Subsequently, these three sets of data are concatenated along the channel dimension to construct a multi-channel input tensor T (a three-dimensional fusion feature tensor) containing intensity, material, and geometric information.
[0032] ; In this tensor, the S0 channel provides the basic visual texture, the DoLP channel provides material contrast in overexposed areas of strong light and areas of the same color, and the AoP channel provides contour geometry constraints in shadow and low-light areas.
[0033] S3. Based on the first Stokes parameter image, perform region labeling to obtain a binary mask image of the target track region; use the three-dimensional fusion feature tensor as input and the binary mask image as ground truth label to train the U-Net network to obtain a model for segmenting the target railway track.
[0034] The first Stokes parameter image is labeled to obtain a binary mask image of the target track region as the ground truth label. The semantic segmentation network U-Net is trained using a multi-channel input tensor T and the labels. Finally, the constructed multi-channel input tensor T is input into the pre-trained U-Net model. During the forward propagation of the network, the encoder uses multi-layer convolutional operations to extract multi-scale fusion features. Polarization features help the network remove specular noise caused by mirror reflections and enhance the boundary response between the track and the background. The decoder upsamples the deep semantic features stepwise and fuses them with shallow detail features to restore spatial resolution. The network's output layer uses a Softmax or Sigmoid activation function to output a pixel-level probability map with the same resolution as the input image. This probability map represents the confidence level of each pixel belonging to the railway track category.
[0035] The output probability map undergoes post-processing. First, a confidence threshold is set for binarization. Pixels with probabilities greater than the threshold are marked as track regions, while those with probabilities less than the threshold are marked as background. Subsequently, morphological operations are used to process the binary image to fill in any minor holes that may exist within the track regions and to smooth jagged edges, ultimately obtaining a continuous and complete railway track segmentation mask, thus completing the track segmentation task.
[0036] from Figures 2 to 5 The original polarization intensity information, to Figures 6 to 8 Stokes polarization parameters, and then Figures 9 to 10 The core polarization characteristics, ultimately to Figure 11 The method utilizes the three-dimensional fusion feature tensor to achieve the progressive purification, transformation, and upgrading of data information through the above steps. It successfully extracts polarization physical characteristics that accurately reflect the material and geometric properties of railway tracks from simple light intensity data. This invention leverages... Figures 2 to 11 The visualization of the processing results at each stage intuitively verifies the effectiveness of polarization features in solving the three major technical challenges of traditional light intensity image track segmentation: the degree of linear polarization image (DoLP) effectively solves the problem of distinguishing between tracks and backgrounds with similar colors and materials; the angle of polarization image (AoP) effectively solves the problem of extracting blurred track edges in low-light and low-contrast scenes; and the overall polarization processing flow based on the Stokes polarization model effectively solves the problem of texture loss caused by metallic track reflections and highlight clipping. A separate first Stokes parameter image ( Figure 6The traditional light intensity image still has inherent limitations, and the linear polarization degree image alone ( Figure 9 ) or polarization angle image ( Figure 10 ) lacks the basic visual texture information of the orbital scene, while Figure 11 The 3D fusion feature tensor integrates the core advantages of three types of features: total light intensity, linear polarization degree, and polarization angle. This allows the fusion features to retain the basic visual texture information of the track scene while also possessing physical material and geometric discrimination capabilities. This provides more robust input features for the training of the U-Net semantic segmentation network, ultimately achieving high-precision and highly robust segmentation of railway tracks under complex lighting conditions.
[0037] Figures 2 to 11 The complete set of accompanying drawings fully demonstrates the core design concept of this invention, which introduces polarization physical information to construct multimodal features. It intuitively shows the process of extracting polarization physical information from the original polarization light intensity data, as well as the process of fusing polarization physical features and light intensity features to form effective training features. This provides intuitive and effective visual experimental support for the technical solution of the railway track segmentation method of this invention.
[0038] This invention provides a railway track segmentation method based on polarization imaging. This method utilizes the material properties and geometric normal features contained in the polarization vector information of light, overcoming the limitations of solely relying on light intensity information in scenarios with strong reflections and metamerism. Addressing the shortcomings of existing railway track segmentation techniques based on visible light intensity images, such as the inability to distinguish metameritic materials, severe interference from strong metallic reflections, and insufficient contour extraction accuracy in low-contrast environments, this invention aims to provide a railway track segmentation method based on polarization imaging.
[0039] To address the technical challenge of existing technologies that rely solely on color and light intensity characteristics to effectively separate rusted rails of similar colors from ballast and sleepers, this invention aims to utilize the strong sensitivity of light wave polarization characteristics to the material properties (metals and dielectrics) of objects, introducing physical dimension discrimination features. This allows for accurate differentiation of tracks from the background even in scenes with extremely similar colors and textures, eliminating false detections and improving the clarity of segmentation boundaries.
[0040] To address the technical challenge of high-brightness overflow in metal tracks under backlight or direct light, which leads to loss of intensity image texture and discontinuity in segmentation masks, this invention aims to leverage the advantage of polarization imaging in capturing light wave vector information. By analyzing the polarization state differences of reflected light, it effectively suppresses high-brightness interference or extracts the inherent polarization characteristics of the reflective region, ensuring the integrity and connectivity of track segmentation results under strong lighting conditions.
[0041] To address the technical challenge of existing technologies that rely excessively on brightness contrast and struggle to capture fine track outlines in low-light areas such as tunnels and shadows, this invention aims to leverage the inherent high correlation between polarization angle information and the normal direction of an object's surface. By enhancing edge information through physical geometric features, it enables precise positioning of track edges even under conditions of weak light intensity contrast, thereby improving the geometric accuracy and robustness of track segmentation.
[0042] This invention significantly improves the robustness of track segmentation under strong light and reflective conditions by introducing polarization imaging technology and a multimodal feature fusion mechanism. Utilizing the physical principle that specular reflection of light from a metal surface has specific polarization characteristics, and by calculating parameters such as the degree of linear polarization (DoLP), the deep learning network can still extract effective features from the polarization channel even when texture information is lost due to highlight clipping in the light intensity image. This mechanism effectively solves the problem of track segmentation mask breakage and loss caused by local overexposure under backlight or direct light conditions in existing technologies, ensuring the integrity of track region extraction under all-weather lighting conditions.
[0043] Furthermore, this invention overcomes the bottleneck in segmentation accuracy caused by the "metachromatic" phenomenon, significantly reducing the false alarm rate of the system. Addressing the technical challenge of distinguishing between rusted rails, ballast, and sleepers due to their extreme similarity in color and texture, this invention utilizes the strong ability of light wave polarization characteristics to distinguish between metallic and dielectric materials, introducing a physical dimension for discrimination. By enhancing the contrast between the foreground track and background debris, this invention can accurately eliminate false detections caused by similarly colored ballast protrusions or non-metallic foreign objects, resulting in smoother and more consistent track contours. Simultaneously, by utilizing the sensitivity of the polarization angle (AoP) to the surface normal direction, this invention can still clearly capture track edges based on geometric normal features even in low-contrast environments such as shadows or low light, compensating for the shortcomings of traditional light intensity imaging that relies on brightness gradients.
[0044] Based on the same inventive concept, the present invention also provides a railway track segmentation system, comprising: The data acquisition module is used to simultaneously acquire the first, second, third, and fourth light intensity images of the target railway track from four different polarization directions.
[0045] The calculation module is used to calculate four light intensity images based on the Stokes polarization model to obtain a first Stokes parameter image, a second Stokes parameter image, and a third Stokes parameter image. The pixel value of the first Stokes parameter image is half of the sum of the corresponding pixel values of the four light intensity images. The second Stokes parameter image is the difference between the corresponding pixel values of the first and third light intensity images. The third Stokes parameter image is the difference between the corresponding pixel values of the second and fourth light intensity images. Using the first, second, and third Stokes parameter images, the linear polarization degree image and the polarization angle image are determined. The first Stokes parameter image, the linear polarization degree image, and the polarization angle image are fused by channel fusion to obtain a three-dimensional fused feature tensor.
[0046] The training module is used to perform region labeling based on the first Stokes parameter image to obtain a binary mask image of the target track region; the U-Net network is trained with the three-dimensional fusion feature tensor as input and the binary mask image as ground truth label to obtain a model for segmenting the target railway track.
[0047] This invention also provides a computer device. At the hardware level, this computer device includes a processor, an internal bus, a network interface, memory, and non-volatile storage, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile storage into the memory and then runs it to implement the railway track segmentation method described above.
[0048] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the railway track segmentation method provided above.
[0049] Specific limitations regarding the calculation system for the railway track segmentation method can be found in the limitations of the railway track segmentation method described above, and will not be repeated here. Each module in the aforementioned railway track segmentation system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.
[0050] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.
Claims
1. A method for dividing railway tracks, characterized in that, include: The first, second, third, and fourth light intensity images of the target railway track were simultaneously acquired from four different polarization directions. Based on the Stokes polarization model, four light intensity images are calculated to obtain a first Stokes parameter image, a second Stokes parameter image, and a third Stokes parameter image. The pixel value of the first Stokes parameter image is half the sum of the corresponding pixel values of the four light intensity images. The second Stokes parameter image is the difference between the corresponding pixel values of the first and third light intensity images. The third Stokes parameter image is the difference between the corresponding pixel values of the second and fourth light intensity images. Using the first, second, and third Stokes parameter images, the linear polarization degree image and the polarization angle image are determined. The first Stokes parameter image, the linear polarization degree image, and the polarization angle image are fused by channel fusion to obtain a three-dimensional fused feature tensor. Region labeling is performed based on the first Stokes parameter image to obtain a binary mask image of the target track region; the U-Net network is trained with the three-dimensional fusion feature tensor as input and the binary mask image as the ground truth label to obtain a model for segmenting the target railway track.
2. The railway track segmentation method according to claim 1, characterized in that, The polarization direction is linear polarization. The first light intensity image is a light intensity image with a 0° linear polarization direction, the second light intensity image is a light intensity image with a 45° linear polarization direction, the third light intensity image is a light intensity image with a 90° linear polarization direction, and the fourth light intensity image is a light intensity image with a 135° linear polarization direction.
3. The railway track segmentation method according to claim 1, characterized in that, The first Stokes parameter image represents the total light intensity of the four light intensity images; the second Stokes parameter image represents the intensity difference between the horizontal polarization direction and the vertical polarization direction; and the third Stokes parameter image represents the intensity difference between two oblique polarization components whose polarization directions differ by 90°.
4. The railway track segmentation method according to claim 1, characterized in that, Before fusing the first Stokes parameter image, the linear polarization degree image, and the polarization angle image to obtain a three-dimensional fused feature tensor, the method further includes: normalizing the first Stokes parameter image and mapping its pixel values to the [0,1] interval; normalizing the polarization angle image and mapping its pixel values to the [0,1] interval; and fusing the normalized first Stokes parameter image, the polarization angle image, and the linear polarization degree image.
5. A railway track segmentation method according to claim 1, characterized in that, The trained U-Net network outputs a pixel-level probability map of the polarization image of the railway track to be segmented. The pixel-level probability map is thresholded to obtain a binary segmentation mask for the railway track. The binary segmentation mask is then morphologically optimized to obtain the final track segmentation result.
6. A railway track segmentation method according to claim 1, characterized in that, Based on the Stokes polarization theory, the linear polarization degree image and polarization angle image are determined using the first Stokes parameter image, the second Stokes parameter image, and the third Stokes parameter image. The linear polarization degree image is used to enhance the material contrast between the track and the background, and the polarization angle image is used to enhance the geometric features of the track edge.
7. A railway track segmentation method according to claim 5, characterized in that, Morphological optimization of the binary segmentation mask includes sequentially performing a closing operation to fill the voids inside the mask and an opening operation to smooth the jagged edges of the mask.
8. A railway track segmentation system, characterized in that, include: The data acquisition module is used to simultaneously acquire the first, second, third, and fourth light intensity images of the target railway track from four different polarization directions. The calculation module is used to calculate four light intensity images based on the Stokes polarization model to obtain a first Stokes parameter image, a second Stokes parameter image, and a third Stokes parameter image. The pixel value of the first Stokes parameter image is half the sum of the corresponding pixel values of the four light intensity images. The second Stokes parameter image is the difference between the corresponding pixel values of the first and third light intensity images. The third Stokes parameter image is the difference between the corresponding pixel values of the second and fourth light intensity images. Using the first, second, and third Stokes parameter images, the linear polarization degree image and the polarization angle image are determined. The first Stokes parameter image, the linear polarization degree image, and the polarization angle image are fused by channel fusion to obtain a three-dimensional fused feature tensor. The training module is used to perform region labeling based on the first Stokes parameter image to obtain a binary mask image of the target track region; the U-Net network is trained with the three-dimensional fusion feature tensor as input and the binary mask image as ground truth label to obtain a model for segmenting the target railway track.
9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 7.