A Deep Learning-Based Method for Damage Detection of Railway Height Restriction Barriers

Abstract

A deep learning-based method for detecting damage anomalies in railway height restriction poles belongs to the technical field of specific computational models for rail transit information monitoring. This invention acquires images of railway height restriction poles using a monitoring device. First, a semantic segmentation model is used to segment the height restriction pole region in the image. Second, the segmented mask undergoes adaptive angle correction. Then, the region corresponding to the horizontal bar portion of the height restriction pole is extracted based on the pixel distribution of the mask. Finally, the mask contour of the horizontal bar is obtained and matched with a preset contour template. Whether the matching degree of the corresponding contour exceeds a threshold determines whether the height restriction pole has suffered damage anomalies. This detection method maximizes the advantages of deep learning methods, improving the efficiency of height restriction pole damage detection.

Description

Technical Field

[0001] This invention discloses a method for detecting damage to railway height restriction poles based on deep learning, belonging to the technical field of specific computational models for rail transit information monitoring. Background Technology

[0002] Railway height restriction barriers are mechanical devices that limit the height of vehicles. They are typically installed on both sides of bridges and underpasses to restrict the passage of vehicles exceeding the prescribed height, preventing them from colliding with and damaging railway bridges and underpasses. However, when a vehicle collides with a height restriction barrier or the barrier is damaged for other reasons, it ceases to restrict the passage of oversized vehicles. In such cases, the railway bridge or underpass faces a greater risk of being struck by subsequent vehicles. Therefore, it is necessary to promptly receive alarm information through monitoring: to immediately report any damage to the height restriction barrier so that maintenance personnel can carry out maintenance work, achieving intelligent operation.

[0003] Traditionally, this hazard is addressed by installing cameras at the entrances of railway bridges or culverts and manually reviewing the surveillance video to obtain information about the height restriction barriers. This method is time-consuming, labor-intensive, and costly. In addition, other methods have emerged in the field for real-time monitoring of the railway height restriction barriers. Chinese patent document CN212026004U discloses a device using a horizontal sensor to obtain information about the height restriction barrier status; however, the sensor solution is susceptible to interference and has a high cost.

[0004] Therefore, with the development of deep learning technology, computer vision technology based on deep learning methods has begun to empower various industries. Chinese patent document CN111104893A discloses a target detection method, device, computer equipment, and storage medium. By determining the target region and road region in a vehicle environment image and performing object segmentation on the target region, target semantic information including the object category and initial contour of the target object is obtained. This achieves a comprehensive description of the target object from multiple perspectives. Then, based on the disparity image and target semantic information, the contour of the target object is further precisely located, greatly improving the accuracy of the contour. Finally, based on the road region and the precisely located contour, the spatial position of the target object is accurately determined. Because it can describe the target object based on the target semantic information and locate the spatial position of the target object based on the high-precision contour, the accuracy of target detection is improved. Chinese patent document CN111666921A discloses a vehicle control method, apparatus, computer equipment, and computer-readable storage medium, comprising: acquiring a road scene image of a target vehicle; performing image semantic segmentation on the road scene image to obtain a semantic segmentation image, the semantic segmentation image including a known semantic region and an unknown semantic region; determining the location information of abnormal obstacles in the unknown semantic region; and controlling the target vehicle to avoid the abnormal obstacles according to the location information. Chinese patent document CN114332020A discloses a method and system for photovoltaic panel localization and defect detection based on visible light images. The method includes: S1: constructing a dataset of visible light images of photovoltaic panels and dividing it into a training set and a test set according to a preset ratio; S2: training a SEFPN semantic segmentation network using the training set to obtain an edge mask image of the photovoltaic panel; S3: inputting the edge mask image of the photovoltaic panel into a contour filtering module to obtain candidate bounding boxes of the photovoltaic panel; S4: inputting the candidate bounding boxes of the photovoltaic panel into a YOLOv5 detection network to obtain the defect category of the photovoltaic panel; S5: the test set sequentially passes through the SEFPN semantic segmentation network, the contour filtering module, and the YOLOv5 detection network to obtain the location and defect category of the photovoltaic panel. This method can be applied to the batch detection of visible light images of photovoltaic panels obtained by UAV aerial photography, avoiding false detections during human identification, improving the overall efficiency of the photovoltaic industry, and greatly contributing to the intelligentization of the photovoltaic industry and reducing the labor costs of photovoltaic power plants.Chinese patent document CN113569819A discloses a method, system, storage medium, and computer equipment for identifying the opening and closing state of a disconnector switch. The method includes: acquiring a target image to be processed, the target image containing a disconnector blade and an insulator, and inputting the target image into a final semantic segmentation model; acquiring the disconnector blade mask and insulator mask output by the final semantic segmentation model after detecting and recognizing the target image; determining whether the disconnector switch is complete based on the insulator mask; if so, extracting the contour of the disconnector blade mask to obtain a set of multiple connected regions; obtaining the area ratio of the first largest connected region to the second largest connected region, and determining whether the area ratio of the first largest connected region to the second largest connected region is greater than a first preset ratio threshold; if so, it indicates that the disconnector switch is in a closed state. The disconnector switch opening and closing state identification method proposed in this invention can greatly reduce background interference, thereby accurately identifying the disconnector switch opening and closing state, and has the advantages of simple operation and low detection cost. As can be seen from the above, computer vision technology based on deep learning methods is beginning to empower various industries; however, its application in intelligent track monitoring, especially in the detection of damage to height restriction poles, is relatively limited.

[0005] In summary, how to provide a reliable and efficient method for identifying damage to railway height restriction barriers and avoid railway track accidents caused by such damage is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a deep learning-based method for detecting abnormal damage to railway height restriction poles.

[0007] The detailed technical solution of this invention is as follows:

[0008] A deep learning-based method for detecting abnormal damage to railway height restriction poles involves acquiring images of the poles using a monitoring device. The method is characterized by: firstly, using a semantic segmentation model to segment the height restriction pole region in the image; secondly, performing adaptive angle correction on the segmented mask; then, to avoid the influence of passing vehicles obstructing the vertical bars of the height restriction pole, extracting the region corresponding to the horizontal bars of the height restriction pole based on the pixel distribution of the mask; finally, obtaining the mask contour of the horizontal bars and matching it with a preset contour template, determining whether the height restriction pole has suffered abnormal damage by checking if the matching degree of the corresponding contour exceeds a threshold.

[0009] According to a preferred embodiment of the present invention, the detection method specifically includes:

[0010] S1: Obtain images of railway height restriction poles through a railway height restriction pole monitoring device, label the railway height restriction pole areas, construct a railway height restriction pole image segmentation dataset, and train a semantic segmentation model to obtain a railway height restriction pole segmentation model; preferably, the semantic segmentation model is the deeplabv3plus semantic segmentation model;

[0011] S2: Input the image of the railway height restriction pole acquired by the real-time monitoring device into the trained semantic segmentation model to obtain the mask image of the railway height restriction pole;

[0012] S3: Rotate the mask image of the railway height restriction bar so that the horizontal bar portion of the height restriction bar in the mask image is parallel to the wide axis of the image; for example... Figure 5 As shown, the purpose of this step is to correct the slight tilt of the image caused by the monitoring equipment.

[0013] A preferred embodiment of the present invention, a method for rotating the mask image of the railway height restriction bar, includes:

[0014] S31: Preset non-zero integers in the range [-α, α] as the rotation angle α of the mask image. n ;

[0015] S32: Rotate the mask image by iterating through the rotation angles in step S31;

[0016] S33: Rotate α from step S32 n The summation of the resulting image along the high axis of the mask image is denoted as:

[0017] M i Let i = (0, 1, ..., height), and record the current rotation angle α and the N largest M values. i and M n =sum(M max1 +...+M max50 ); wherein, the M i It is a set of numbers representing the sum of the number of pixels in the i-th row of the mask image, which is equivalent to projecting the mask onto the high axis of the image; N is the maximum number; M is the maximum number of pixels. n Calculate M i The sum of the N largest numbers in the corresponding group is shown in the appendix. Figure 7 The sum is the sum of the values ​​of the top 50 points;

[0018] Further explanation:

[0019] ① Assume that in step s31 at this point, α = -2;

[0020] ② Rotate the mask image by -2 degrees;

[0021] ③ Project the rotated mask image onto the high axis to obtain the attached image. Figure 7 This diagram visualizes this set of numbers, representing a set of numbers denoted as M. i Assume N = 50 in step S33; calculate the sum of the 50 largest numbers in this set, and denote this sum as M. n④ Within the preset rotation range, change the rotation angle α and repeat steps ①-③ above to calculate the different M values ​​corresponding to different rotation angles α. n ;

[0022] S34: Calculate max(M) n The corresponding rotation angle α n This is the desired adaptive correction rotation angle;

[0023] S4: Obtain the area where the railway height restriction bar is located, including:

[0024] By leveraging the inherent shape properties of the height restriction pole, a region extraction algorithm is used to obtain the area range where the horizontal bar of the height restriction pole is located, in order to avoid false alarms caused by the vertical pole being blocked by vehicles.

[0025] S41: Summate the mask image along the high axis of the image, denoted as M. i , i = (0, 1, ..., height);

[0026] S42: Divide the sum obtained in step S41 by the width of the mask image to get the proportion of the mask to the image width, denoted as:

[0027] R i , i = (0, 1, ..., height);

[0028] S43: Record: In step S42, the proportion of the mask along the high axis to the image width is greater than the preset threshold R. threshold At that time, the corresponding range of the mask image height is [a,b];

[0029] S44: Cut the mask image along [an, b+n] to obtain the image region containing the railway height restriction bar, where n is the number of pixels, referring to the height cut, and n is an empirical threshold; in this step, intuitively, assume that... Figure 7 The horizontal range where the vertical axis is greater than the threshold of 0.15 is [180, 210]. Assuming n is 5, then cutting [175, 215] yields the attached [180, 210]. Figure 8 ;

[0030] S5: Extract and filter the outline of the railway height restriction bar within the area where the horizontal bar is located, while preserving the outline of the railway height restriction bar area;

[0031] S51: Extract the contour within the area of ​​the railway height restriction bar: extract the contour of all masks within the area;

[0032] S52: Calculate the area S of the contours of all masks. The calculation of the area of ​​the mask contours can be done by directly calling existing functions. The specific calling and calculation process is not the content to be protected by this invention.

[0033] S53: Filter out contour areas smaller than the empirical threshold S threshold The mask contour; the empirical threshold S threshold 500 is acceptable;

[0034] S6: Match the extracted contour of the railway height restriction bar area with the contour template. The contour is a previously stored standard contour; subsequent real-time input contours are compared with the standard contour to calculate the matching degree. If no standard contour exists previously, the first input contour is used as the standard contour. The contour template is the contour obtained under normal conditions of the height restriction bar. The matching process uses a conventional algorithm, which is not the entity protected by this invention. Whether the matching is successful determines whether the height restriction bar is damaged.

[0035] S61: The similarity between the horizontal bar region contour and the template contour is calculated using Hu moments. The specific calculation is implemented using a function in OpenCV. Hu moments are constructed using normalized central moments to create seven invariant eigenmotes.

[0036]

[0037] In formula (1), u oo Let μ be the center of the moment. pq The moment is of order (p+q);

[0038] S62: If the contour similarity is greater than the contour matching threshold Match threshold If the match is successful, the railway height restriction barrier is considered normal; otherwise, it is considered damaged or abnormal. The contour matching threshold is set to Match. threshold The threshold is an empirical threshold, and the process of determining it is not the subject of this invention.

[0039] The present invention has the following beneficial effects:

[0040] 1) The present invention provides a method for detecting damage to railway height restriction poles based on deep learning, which includes a semantic segmentation model based on deep learning and related post-processing algorithms. This method can maximize the advantages of deep learning methods and improve the efficiency of height restriction pole damage detection.

[0041] 2) This invention also proposes an adaptive rotation algorithm for the height restriction pole mask image. In the prior art, if the height restriction pole image is tilted, it will affect the extraction of the horizontal bar area and its contour in the next step when determining whether the height restriction pole is damaged. Therefore, this invention adaptively rotates the height restriction pole mask along the height axis of the image to correct the slight tilt of the image caused by the monitoring equipment, so that the horizontal bar part of the height restriction pole is parallel to the width axis of the image.

[0042] 3) In addition, this invention also addresses the situation where the vertical bars of height restriction poles are easily obscured by passing vehicles, leading to misjudgment. It proposes a method for obtaining the horizontal bar region of a height restriction pole mask image. This method determines the rectangular region where the horizontal bar of the height restriction pole is located based on the shape characteristics of the height restriction pole and the distribution of the height restriction pole mask along the image height axis.

[0043] 4) The present invention also proposes a method for judging the damage of height restriction poles: the contour of the height restriction pole crossbar mask and the contour template are matched by Hu moment, and the damage of the height restriction pole is judged by whether the two are successfully matched. Attached Figure Description

[0044] Figure 1 This is a flowchart of the identification method described in this invention;

[0045] Figure 2 This is an image of the labeled region visualization of the semantic segmentation dataset of the height limit bar in step S1 of the embodiment of the present invention, wherein the overall display of the height limit bar is the color of the region "height_limit_rod" shown in the lower right corner;

[0046] Figure 3 The image is the input semantic segmentation model in step S2 of this embodiment of the invention, wherein the height restriction bar is intact;

[0047] Figure 4 The image is the input semantic segmentation model in step S2 of this embodiment of the invention, wherein the height restriction bar is damaged;

[0048] Figure 5 yes Figure 3 The height-limiting pole mask image output after semantic segmentation in step S3 of this embodiment of the invention;

[0049] Figure 6 yes Figure 4 The height-limiting pole mask image output after semantic segmentation in step S3 of this embodiment of the invention;

[0050] Figure 7 yes Figure 3 In step S42 of this embodiment of the invention, a schematic diagram is shown showing the proportion of the height-limiting rod mask to the image width; wherein, the Height of the horizontal axis is the height axis coordinate of the mask image, indicating the nth pixel value, and has no unit; the rat of the vertical axis is the proportion of the height-limiting rod mask to the image width.

[0051] Figure 8 yes Figure 5 An image of the range of the height restriction bar horizontal bar extracted in step S44 of an embodiment of the present invention;

[0052] Figure 9 yes Figure 5 A schematic diagram of the outline of the horizontal bar area of ​​the height limit bar in step S5 of an embodiment of the present invention;

[0053] Figure 10 yes Figure 6 A schematic diagram of the outline of the height limit bar area in step S5 of this embodiment of the invention. Detailed Implementation

[0054] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the specific implementation methods, structures, features, and effects of the present invention are described in detail below with reference to embodiments:

[0055] Example 1

[0056] A deep learning-based method for detecting abnormal damage to railway height restriction poles involves acquiring images of the poles using a monitoring device. The method is characterized by: firstly, using a semantic segmentation model to segment the height restriction pole region in the image; secondly, performing adaptive angle correction on the segmented mask; then, to avoid the influence of passing vehicles obstructing the vertical bars of the height restriction pole, extracting the region corresponding to the horizontal bars of the height restriction pole based on the pixel distribution of the mask; finally, obtaining the mask contour of the horizontal bars and matching it with a preset contour template, determining whether the height restriction pole has suffered abnormal damage by checking if the matching degree of the corresponding contour exceeds a threshold.

[0057] According to a preferred embodiment of the present invention, the detection method specifically includes:

[0058] S1: Obtain images of railway height restriction poles using a monitoring device, label the areas of the railway height restriction poles, construct a railway height restriction pole image segmentation dataset, and train a semantic segmentation model to obtain a railway height restriction pole segmentation model; preferably, the semantic segmentation model is the deeplabv3plus semantic segmentation model; the dataset contains two categories: height restriction poles and background, and the images of the labeled areas in the dataset are shown in the attached figure. Figure 2 As shown;

[0059] S2: Input the image of the railway height restriction pole acquired by the real-time monitoring device into the trained semantic segmentation model to obtain the mask image of the railway height restriction pole; wherein, the input image includes Figure 3 and Figure 4 (Simulated image of a damaged height restriction bar); the mask image obtained after semantic segmentation is as follows. Figure 5 and Figure 6 As shown;

[0060] S3: Rotate the mask image of the railway height restriction bar so that the horizontal bar portion of the height restriction bar in the mask image is parallel to the wide axis of the image; for example... Figure 5 As shown, the purpose of this step is to correct the slight tilt of the image caused by the monitoring equipment.

[0061] A preferred embodiment of the present invention, a method for rotating the mask image of the railway height restriction bar, includes:

[0062] S31: Preset non-zero integers in the range [-α, α] as the rotation angle α of the mask image. n In this embodiment, a non-zero integer within the preset range of [-5, 5] is used as the image rotation angle α. n n = (1, 2, ..., 10);

[0063] S32: Rotate the mask image by iterating through the rotation angles in step S31;

[0064] S33: Rotate α from step S32 n The summation of the resulting image along the high axis of the mask image is denoted as:

[0065] M i Let i = (0, 1, ..., height), and record the current rotation angle α and the N largest M values. i and M n =sum(M max1 +...+M max50 ); wherein, the M i It is a set of numbers representing the sum of the number of pixels in the i-th row of the mask image, which is equivalent to projecting the mask onto the high axis of the image; N is the maximum number; M is the maximum number of pixels. n Calculate M i The sum of the N largest numbers in the corresponding group is shown in the appendix. Figure 7 The sum is the sum of the values ​​of the top 50 points;

[0066] Further explanation:

[0067] ① Assume that in step s31 at this point, α = -2;

[0068] ② Rotate the mask image by -2 degrees;

[0069] ③ Project the rotated mask image onto the high axis to obtain the attached image. Figure 7 This diagram visualizes this set of numbers, representing a set of numbers denoted as M. i Assume N = 50 in step S33; calculate the sum of the 50 largest numbers in this set, and denote this sum as M. n ④ Within the preset rotation range, change the rotation angle α and repeat steps ①-③ above to calculate the different M values ​​corresponding to different rotation angles α. n ;

[0070] S34: Calculate max(M) n The corresponding rotation angle α n This is the desired adaptive correction rotation angle;

[0071] S4: Obtain the area where the railway height restriction bar is located, including:

[0072] By leveraging the inherent shape properties of the height restriction pole, a region extraction algorithm is used to obtain the area range where the horizontal bar of the height restriction pole is located, in order to avoid false alarms caused by the vertical pole being blocked by vehicles.

[0073] S41: Sum the mask image along the high axis of the image, denoted as M. i , i = (0, 1, ..., height);

[0074] S42: Divide the sum obtained in step S41 by the width of the mask image to get the proportion of the mask to the image width, denoted as:

[0075] R i , i = (0, 1, ..., height);

[0076] by Figure 5 As shown, the desired proportion is visualized as follows: Figure 7 As shown, assuming Figure 7 A point (200, 0.55) on the mid-curve represents the mask image. Figure 5 In the 200th row of pixels from top to bottom, white pixels account for 55% of the total pixels in the row; if the threshold is set to 15%, then the position of this row of pixels is determined to be the horizontal bar of the height limit bar.

[0077] S43: Record: In step S42, the proportion of the mask along the high axis to the image width is greater than the preset threshold R. threshold At that time, the corresponding range of the mask image height is [a,b];

[0078] S44: Cut the mask image along [an, b+n] to obtain the image region containing the railway height restriction bar, where n is the number of pixels, referring to the height cut, and n is an empirical threshold; in this step, intuitively, assume that... Figure 7 The horizontal axis range where the vertical axis is greater than the threshold of 0.15 is [180, 210]. Assuming n is 5, then cutting [175, 215] yields the attached [180, 210]. Figure 8 ;

[0079] S5: Extract and filter the outline of the railway height restriction bar within the area where the horizontal bar is located, while preserving the outline of the railway height restriction bar area;

[0080] S51: Extract the contour within the area of ​​the railway height restriction bar: extract the contour of all masks within the area;

[0081] S52: Calculate the area S of the contours of all masks. The calculation of the area of ​​the mask contours can be done by directly calling existing functions. The specific calling and calculation process is not the content to be protected by this invention.

[0082] S53: Filter out contour areas smaller than the empirical threshold S threshold The mask contour; the empirical threshold S threshold 500 can be taken. Figure 5 and Figure 6 As shown, the outlines of the retained height restriction bar areas are as follows: Figure 9 and Figure 10 As shown;

[0083] S6: Match the extracted contour of the railway height restriction bar area with the contour template. The contour is a previously stored standard contour; subsequent real-time input contours are compared with the standard contour to calculate the matching degree. If no standard contour exists previously, the first input contour is used as the standard contour. The contour template is the contour obtained under normal conditions of the height restriction bar. The matching process uses a conventional algorithm, which is not the entity protected by this invention. Whether the matching is successful determines whether the height restriction bar is damaged.

[0084] Will Figure 9 The outline shown is used as a outline template. Figure 10 The outline shown is matched with the outline template;

[0085] S61: Calculate the similarity between the horizontal bar region contour and the template contour using Hu moments. Figure 10 The outline of the horizontal bar region shown is... Figure 9 The similarity of the template contours shown is calculated using a function in OpenCV; the Hu moments are constructed using normalized central moments to create seven invariant eigenmotes:

[0086]

[0087] In formula (1), u oo Let μ be the center of the moment. pq The moment is of order (p+q);

[0088] S62: If the contour similarity is greater than the contour matching threshold Match threshold If the match is successful, the railway height restriction barrier is considered normal; otherwise, it is considered damaged or abnormal. The contour matching threshold is set to Match. threshold The threshold is an empirical threshold, and the process of determining it is not the subject of this invention.

Claims

1. A method for detecting abnormal damage to railway height restriction poles based on deep learning, which acquires images of railway height restriction poles using a railway height restriction pole monitoring device, characterized in that: First, a semantic segmentation model is used to segment the height restriction pole region in the image. Second, the segmented mask undergoes adaptive angle correction. Then, the region corresponding to the horizontal bar portion of the height restriction pole is extracted based on the pixel distribution of the mask. Finally, the mask contour of the horizontal bar is obtained and matched with a preset contour template. Whether the matching degree of the corresponding contour exceeds a threshold determines whether the height restriction pole is damaged or abnormal. Specifically, this includes: S1: Obtain images of railway height restriction poles through railway height restriction pole monitoring devices, label the railway height restriction pole areas, construct a railway height restriction pole image segmentation dataset, and train a semantic segmentation model to obtain a railway height restriction pole segmentation model; S2: Input the image of the railway height restriction pole acquired by the real-time monitoring device into the trained semantic segmentation model to obtain the mask image of the railway height restriction pole; S3: Rotate the mask image of the railway height restriction pole so that the horizontal bar portion of the height restriction pole in the mask image is parallel to the wide axis of the image; The method for rotating the mask image of the railway height restriction pole in step S3 includes: S31: Preset non-zero integers in the range [-α, α] as the rotation angle of the mask image. ; S32: Rotate the mask image by iterating through the rotation angles in step S31; S33: Rotate step S32 The summation of the resulting image along the high axis of the mask image is denoted as: Record the current rotation angle The largest N and ; wherein, the M i It is a set of numbers representing the mask image number. i The sum of the number of pixels in each row; N is the maximum number; the M n calculate M i The sum of the N largest numbers in the corresponding group; S34: Calculation Corresponding rotation angle This is the desired adaptive correction rotation angle; S4: Obtain the area where the railway height restriction bar is located. Specific methods include: S41: Summing the mask image along the high axis of the image, denoted as... ; S42: Divide the sum obtained in step S41 by the width of the mask image to get the proportion of the mask to the image width, denoted as: ； S43: Record: In step S42, the proportion of the mask along the high axis to the image width is greater than a preset threshold. At that time, the corresponding range of the mask image height is [a,b]; S44: Cut the mask image along [am, b+m] to obtain the image region containing the railway height restriction bar, where m is the number of pixels; S5: Extract and filter the outline of the railway height restriction bar within the area where the horizontal bar is located, while preserving the outline of the railway height restriction bar area; S6: Match the extracted outline of the railway height restriction bar area with the outline template, and determine whether the height restriction bar is damaged by checking whether the match is successful.

2. The method for detecting abnormal damage to railway height restriction barriers based on deep learning according to claim 1, characterized in that, Step S6 specifically refers to: S61: The similarity between the outline of the horizontal bar region and the template outline is calculated using Hu moments; Hu moments are constructed using normalized central moments to create 7 invariant eigenmotes. （1） In formula (1), , Center of the square, The moment is of order (p+q); S62: If the contour similarity is less than the contour matching threshold If the match is successful, the railway height restriction barrier is considered to be normal; otherwise, the railway height restriction barrier is considered to be damaged or abnormal.

3. The method for detecting abnormal damage to railway height restriction barriers based on deep learning according to claim 1, characterized in that, The specific steps of step 5 are as follows: S51: Extract the contour within the area of ​​the railway height restriction bar: extract the contour of all masks within the area; S52: Calculate the area S of the contours of all masks; S53: Filter out profiles with an area smaller than the empirical threshold. The mask outline.

Images