Steel rail corrugation detection method and device based on computer vision

Abstract

The invention discloses a steel rail corrugation detection method and device based on computer vision, and relates to the technical field of rail transit, and the method comprises the steps: 1, periodically and automatically collecting a high-resolution steel rail surface image, and carrying out the preprocessing; 2, carrying out steel rail region detection on the preprocessed steel rail surface image, and extracting a steel rail sub-image; step 3, carrying out corrugation position detection on the steel rail sub-graph to obtain a corrugation area coordinate; and 4, mapping the coordinates of the corrugation area to a coordinate system of the steel rail surface image to obtain full-image corrugation positioning. According to the method, advanced image processing and deep learning algorithms are adopted, the steel rail corrugation characteristics can be effectively extracted from a complex background, high-precision disease recognition and positioning are achieved, compared with a traditional method, the method has higher detection sensitivity and lower false alarm rate, and the precision ratio and recall ratio of steel rail corrugation detection can be remarkably increased.

Description

Technical Field

[0001] This invention relates to the field of rail transit technology, and more specifically to a method and apparatus for detecting rail corrugation based on computer vision. Background Technology

[0002] Rail transit, as a crucial infrastructure of modern cities, plays a vital role in economic development and social progress. However, during long-term operation, the surface of the rails develops periodic undulations and deformations, known as rail corrugation. This condition is primarily caused by the dynamic interaction between the wheels and rails, and gradually worsens with the increase in train operation frequency. Rail corrugation not only leads to increased vibration and noise during train operation but also accelerates the wear of rails and wheels, seriously affecting driving safety and comfort. If not addressed promptly, corrugation will further deteriorate, potentially leading to fatigue cracks in the rails, increasing the risk of derailment, and threatening passenger safety. Therefore, timely and accurate detection and treatment of rail corrugation are essential for ensuring the safe operation of rail transit systems.

[0003] With the rapid expansion of rail transit networks and the continuous increase in operating mileage, traditional manual inspection methods can no longer meet the growing inspection demands. Manual inspection is not only inefficient but also susceptible to subjective factors, making it difficult to guarantee the consistency and accuracy of inspection results. Furthermore, frequent manual inspections consume valuable operating time, affecting normal train operations. Therefore, developing automated rail corrugation inspection technology has become an urgent need in the current rail transit maintenance field. Automated inspection can not only improve inspection efficiency and accuracy but also achieve all-weather, large-scale monitoring, providing timely and reliable data support for track maintenance decisions.

[0004] Among numerous automated inspection technologies, computer vision-based rail corrugation detection methods offer significant advantages. They are non-contact and non-destructive, causing no additional wear to the rails and making them suitable for long-term, continuous monitoring applications. Compared to other sensors, vision sensors can provide high-resolution images of the rail surface, containing rich texture and geometric information, which is beneficial for capturing minute surface changes. However, the accuracy and sensitivity of current computer vision-based rail corrugation detection technologies still need improvement.

[0005] Therefore, improving the accuracy and sensitivity of computer vision-based rail corrugation detection is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method and apparatus for detecting rail corrugation based on computer vision. It adopts advanced image processing and deep learning algorithms, which can effectively extract rail corrugation features from complex backgrounds, achieve high-precision defect identification and location, and has higher detection sensitivity and lower false alarm rate compared with traditional methods. It can significantly improve the precision and recall of rail corrugation detection.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A computer vision-based method for detecting rail corrugation includes the following steps:

[0009] Step 1: Periodically and automatically acquire high-resolution images of the rail surface and perform preprocessing;

[0010] Step 2: Perform rail region detection on the preprocessed rail surface image and extract rail sub-images;

[0011] Step 3: Detect the corrugation location on the rail sub-map and obtain the coordinates of the corrugated area;

[0012] Step 4: Map the coordinates of the corrugated area to the coordinate system of the rail surface image to obtain the full-image corrugation location.

[0013] Preferably, the preprocessing of the rail surface image includes normalization and size adjustment to enhance the ability to identify the rail area.

[0014] Preferably, the specific process of step 2 is as follows:

[0015] Step 22: Use the target detection model to infer the preprocessed rail surface image, output the rail region bounding box, and obtain the rail region coordinates;

[0016] Step 23: Cropping the preprocessed rail surface image according to the rail region coordinates to obtain a rail sub-image, and performing sub-image preprocessing; sub-image preprocessing includes size normalization, brightness and contrast enhancement, etc. This preprocessing is adapted for subsequent corrugation detection.

[0017] Preferably, the specific implementation process of step 3 is as follows:

[0018] Step 31: Perform feature enhancement processing on the vertical features of the rail sub-image to obtain an enhanced image; the feature enhancement processing includes using methods such as the vertical Sobel operator and Gabor filtering to enhance the vertical features and suppress lateral noise;

[0019] Step 32: Use the ripple detection model to infer the enhanced image, determine the mask and bounding box of the ripple region, and obtain the coordinates of the ripple region.

[0020] Preferably, the object detection model uses the YOLOv5 network as its basic network structure, accelerates the training process through transfer learning, and adopts pre-trained weights on a large-scale object detection dataset for the initial model weights, followed by fine-tuning on a rail dataset. This method significantly reduces training time while improving the model's generalization ability.

[0021] Preferably, the ripple detection model includes a feature extraction and asymmetric downsampling module, and a deep feature fusion and detection head module;

[0022] The feature extraction and asymmetric downsampling module consists of an initial convolutional layer, an asymmetric downsampling layer, and multiple lightweight convolutional blocks connected in sequence. The initial convolutional layer performs standard convolution on the enhanced image to extract basic features. The asymmetric downsampling layer uses convolution or pooling operations with a stride of (2,1) to downsample the basic features and obtain the initial downsampled features. The multiple lightweight convolutional blocks use lightweight structures such as C2f and Bottleneck from the Ultralytics YOLO model to further extract multi-scale features from the initial downsampled features and obtain downsampled multi-scale features.

[0023] The deep feature fusion and detection head module consists of an asymmetric downsampling layer, a feature fusion layer, and a detection head connected in sequence. The asymmetric downsampling layer uses convolution or pooling operations with a stride of (2,1) to downsample the downsampled multi-scale features, reducing the height dimension by a factor of 2 while keeping the width dimension unchanged. Since rail corrugation mainly manifests as periodic changes in the lateral direction, it is necessary to retain lateral detail information. Through asymmetric downsampling, the height is gradually compressed to be close to the width, and the final aspect ratio is close to 1:1, obtaining standard multi-scale features. The feature fusion layer uses an FPN structure to achieve skip connections between features of different scales, fusing standard multi-scale features and downsampled multi-scale features to obtain fused features, enhancing the model's ability to detect corrugated regions of different sizes. The detection head uses the detection head structure of the Ultralytics YOLO model to output corrugation confidence, bounding box coordinates, and other information at each location based on the fused features. It obtains the mask of the corrugated region based on the corrugation confidence at each location and obtains the bounding box based on the bounding box coordinates.

[0024] The corrugation detection model introduces an asymmetric downsampling strategy to preserve lateral feature information. Specifically, the asymmetric downsampling layer only downsamples in the height dimension while keeping the width dimension unchanged. This design effectively preserves the lateral features of rail corrugation and improves detection accuracy. The feature fusion layer's FPN structure uses transposed convolution for upsampling and introduces skip connections to fuse features at different scales.

[0025] Preferably, in step 1, multiple images of the rail surface are acquired from one perspective or multiple images of the rail surface from multiple perspectives. The specific implementation process of step 4 is as follows:

[0026] Step 41: Map the coordinates of the corrugated area of ​​each rail surface image to the coordinate system of the corresponding rail surface image;

[0027] Step 42: Fuse all rail surface images, remove duplicate and false detections, and obtain full-image corrugation localization; specifically:

[0028] Map all mapped rippled area coordinates to the same global coordinate system;

[0029] In the global coordinate system, rail surface images with a confidence level lower than the set confidence threshold are deleted from the corrugated region, and spatially overlapping or adjacent corrugated regions in the remaining different rail surface images are identified to form a candidate corrugated region set.

[0030] For each set of candidate corrugated regions, the coordinates of the corrugated regions are weighted and averaged according to their confidence level, or the coordinates of the corrugated region with the highest confidence level are selected.

[0031] The coordinates of the corrugated regions selected from each candidate corrugated region set are summarized to obtain the full-map corrugation location.

[0032] Preferably, the method also includes step 5, which involves sorting all wave-eroded areas in the full-map wave-erosion localization from largest to smallest based on their area, intensity, and other indicators, and identifying the top n wave-eroded areas as high-risk samples for priority submission for manual review. Here, n is a preset threshold parameter.

[0033] Preferably, both the target detection model and the wave polishing detection model are trained and optimized using the Adam optimizer, the learning rate is optimized using the cosine annealing strategy, and L2 regularization and Dropout techniques are introduced to prevent overfitting. During training, cross-validation is used to evaluate model performance, and hyperparameters are dynamically adjusted based on the performance on the validation set. The training termination condition is set to no significant improvement in the mean accuracy on the validation set for 5 consecutive epochs.

[0034] A computer vision-based rail corrugation detection device includes a high-definition vision sensor, a digital signal acquisition unit, an A / D analog-to-digital conversion circuit, an industrial control computer, sensor matching signal lines, and an equipment mounting beam.

[0035] The high-definition vision sensor is fixed above the rail via the equipment mounting beam to collect images of the rail surface in real time and generate image signals.

[0036] The image signal acquired by the high-definition vision sensor is transmitted to the digital signal acquisition unit via the sensor's matching signal line; if the image signal is an analog signal, it is converted into a digital signal by an A / D analog-to-digital converter circuit and then transmitted to the digital signal acquisition unit; if the image signal is a digital signal, it is directly transmitted to the data signal acquisition unit via the sensor's matching signal line.

[0037] The digital signal acquisition unit performs preliminary processing and buffering of digital signals to generate image data, and transmits the image data to the industrial control computer via a high-speed data cable;

[0038] The industrial control computer, as the core processing unit of the system, is responsible for receiving, storing, and analyzing the acquired image data, running the rail corrugation detection method based on computer vision, and generating a full-image corrugation location.

[0039] Preferably, the industrial control computer uploads the full-map corrugation positioning and raw image data to a remote server or cloud platform via a wireless network connection component, enabling remote monitoring and data sharing. The remote server or cloud platform can be configured with a visual interface to intuitively display the inspection results.

[0040] Preferably, it also includes power supply equipment and accessories to provide a stable power supply for all hardware such as high-definition vision sensors, data signal acquisition devices, A / D analog-to-digital conversion circuits, and industrial control computers, ensuring the continuous and reliable operation of the system.

[0041] Preferably, the high-definition vision sensor adopts a high-definition track camera. Five high-definition track cameras are installed on the equipment mounting beam under the vehicle of the comprehensive track defect detection vehicle and are evenly distributed along the transverse direction of the vehicle to ensure complete coverage of the entire rail surface. The equipment mounting beam adopts a vibration reduction design to reduce the impact of vehicle vibration on image quality during operation.

[0042] Preferably, the industrial PC can also be equipped with an alarm module, which automatically sends information to relevant personnel when severe corrugation is detected. The industrial PC also has a data storage and analysis module, supporting historical data queries and trend analysis. As an edge computing device, the industrial PC can utilize a low-power, high-performance embedded processor, possessing sufficient computing power to run trained deep learning models. To adapt to complex field environments, edge computing devices can also adopt shockproof, dustproof, and wide-temperature designs to ensure stable operation under various harsh conditions.

[0043] As can be seen from the above technical solution, compared with the prior art, this invention discloses a computer vision-based method and device for detecting rail corrugation, which can solve the important problem of automated detection of rail corrugation. It combines high-definition image acquisition, deep learning models, and edge computing technology to achieve real-time, efficient, and accurate detection of rail corrugation, providing strong support for the safe operation and preventative maintenance of rail transit systems. Specifically, the use of a high-resolution line scan camera and advanced image processing technology enables the acquisition of clear rail surface images at high speeds. High-quality data acquisition provides a solid foundation for subsequent corrugation detection, improving the overall system's detection accuracy. The two-stage detection strategy first detects the rail area, and then detects the corrugation location within the identified rail area. This method effectively alleviates the problem of insufficient corrugation defect samples and reduces the false detection rate. In particular, the asymmetric downsampling network structure used in the second-stage detection better preserves the lateral feature information of the rail corrugation, further improving detection accuracy. The device structure fully considers the needs of the actual application environment, providing data support for long-term track maintenance decisions. Its modular design gives it good scalability, allowing for the flexible addition of other types of rail defect detection functions as needed. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the rail corrugation detection method based on computer vision provided by the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] This invention discloses a computer vision-based method for detecting rail corrugation, such as... Figure 1 As shown, it includes the following steps:

[0048] S1: Periodically and automatically acquire high-resolution images of the rail surface and perform preprocessing;

[0049] S2: Perform rail region detection on the preprocessed rail surface image and extract rail sub-images;

[0050] S3: Detect the corrugation location on the rail sub-map and obtain the coordinates of the corrugated area;

[0051] S4: Map the coordinates of the corrugated area to the coordinate system of the rail surface image to obtain the full-image corrugation location.

[0052] Further preprocessing of the rail surface image includes normalization and size adjustment to enhance the ability to identify rail areas.

[0053] Furthermore, the specific process of S2 is as follows:

[0054] S21: Use an object detection model to infer the preprocessed rail surface image, output the rail region bounding box, and obtain the rail region coordinates;

[0055] S22: The preprocessed rail surface image is cropped according to the rail region coordinates to obtain a rail sub-image, and then preprocessed. The sub-image preprocessing includes size normalization, brightness and contrast enhancement, etc. The preprocessing is adapted to subsequent corrugation detection.

[0056] Furthermore, the object detection model uses the YOLOv5 network as its basic network structure and accelerates the training process through transfer learning. The initial weights of the model are pre-trained on a large-scale object detection dataset, and then fine-tuned on a rail dataset. This method significantly reduces training time while improving the model's generalization ability.

[0057] Furthermore, the specific implementation process of S3 is as follows:

[0058] S31: Perform feature enhancement processing on the vertical features of the rail sub-image to obtain an enhanced image; the feature enhancement processing includes using methods such as the vertical Sobel operator and Gabor filtering to enhance the vertical features and suppress lateral noise;

[0059] S32: Use the ripple detection model to infer the enhanced image, determine the mask and bounding box of the ripple region, and obtain the coordinates of the ripple region.

[0060] Furthermore, the ripple detection model includes a feature extraction and asymmetric downsampling module, and a deep feature fusion and detection head module;

[0061] The feature extraction and asymmetric downsampling module consists of an initial convolutional layer, an asymmetric downsampling layer, and multiple lightweight convolutional blocks connected in sequence. The initial convolutional layer performs standard convolution on the enhanced image to extract basic features. The asymmetric downsampling layer uses different downsampling ratios in the height and width dimensions, and uses convolutional operations or pooling operations with a stride of (2,1) to downsample the basic features to obtain the initial downsampled features. The multiple lightweight convolutional blocks use lightweight structures such as C2f and Bottleneck from the Ultralytics YOLO model to further extract multi-scale features from the initial downsampled features to obtain downsampled multi-scale features.

[0062] The deep feature fusion and detection head module consists of an asymmetric downsampling layer, a feature fusion layer, and a detection head connected in sequence. The asymmetric downsampling layer uses convolution or pooling operations with a stride of (2,1) to downsample the downsampled multi-scale features, gradually compressing the height to be close to the width, and finally achieving an aspect ratio close to 1:1 to obtain standard multi-scale features. The feature fusion layer uses FPN (Feature Pyramid Network) or PAN (Path Aggregation Network) structures to achieve skip connections between features of different scales, fusing standard multi-scale features and downsampled multi-scale features to obtain fused features, enhancing the model's ability to detect erosion regions of different sizes. The detection head uses the detection head structure of the Ultralytics YOLO model to output information such as erosion confidence and bounding box coordinates at each position based on the fused features. It obtains the mask of the erosion region based on the erosion confidence at each position and obtains the bounding box based on the bounding box coordinates.

[0063] The corrugation detection model introduces an asymmetric downsampling strategy to preserve lateral feature information. Specifically, the asymmetric downsampling module only performs downsampling in the height dimension while keeping the width dimension unchanged. This design effectively preserves the lateral features of rail corrugation and improves detection accuracy. The feature fusion layer uses transposed convolution for upsampling and introduces skip connections to fuse features at different scales.

[0064] The ripple detection model employs two cascaded n-models. The first n-model focuses on preliminary feature extraction and asymmetric downsampling, and its output feature map is directly used as the input of the second n-model. The second n-model further mines features and completes the detection task, thereby achieving end-to-end data flow and completing the ripple region detection. Considering the inconsistent saliency of the lateral and vertical features of rail corrugations, an innovative downsampling strategy was adopted in the model design. Typically, downsampling operations in convolutional neural networks (such as max pooling or convolution with a stride of 2) simultaneously reduce the width and height of the feature map to reduce computational complexity and increase the receptive field. This symmetrical downsampling method performs well when processing ordinary images, but may lead to the loss of lateral feature information of special targets such as rail corrugations. Therefore, this invention proposes an asymmetric downsampling strategy. In some network layers, only the height dimension is reduced while the width dimension is retained to maximize the preservation of lateral feature information. This asymmetric downsampling is mainly applied to the detection stage of the second n-model. At this time, the input image is a rectangular rail region, which usually has a small aspect ratio. After downsampling, the aspect ratio of the image is close to 1:1, which is beneficial for feature extraction in subsequent convolutional layers.

[0065] The wave polishing detection model employs a lightweight network structure, which not only ensures real-time detection but also enables the model to be deployed on small, lightweight, and low-power edge computing devices. This feature greatly improves the system's flexibility and applicability, allowing the method of this invention to achieve efficient wave polishing detection in various complex field environments. By combining asymmetric downsampling and a lightweight network structure, the method of this invention significantly improves computational efficiency while maintaining high detection accuracy, thus ensuring real-time performance in practical applications.

[0066] Furthermore, S3 also includes determining whether there is erosion at the corresponding location through erosion confidence level, and judging the severity of erosion through manual review.

[0067] Furthermore, S1 acquires multiple rail surface images from one perspective or multiple rail surface images from multiple perspectives. The specific implementation process of S4 is as follows:

[0068] S41: Map the coordinates of the corrugated area of ​​each rail surface image to the coordinate system of the corresponding rail surface image;

[0069] S42: Fuse all rail surface images, remove duplicate and false detections, and obtain full-image corrugation localization; specifically:

[0070] Map all mapped rippled area coordinates to the same global coordinate system;

[0071] In the global coordinate system, rail surface images with a confidence level lower than the set confidence threshold are deleted from the corrugated region, and spatially overlapping or adjacent corrugated regions in the remaining different rail surface images are identified to form a candidate corrugated region set.

[0072] For each set of candidate corrugated regions, the coordinates of the corrugated regions are weighted and averaged according to their confidence level, or the coordinates of the corrugated region with the highest confidence level are selected.

[0073] The coordinates of the corrugated regions selected from each candidate corrugated region set are summarized to obtain the full-map corrugation location.

[0074] Furthermore, S5 is also included, which sorts all wave-eroded areas in the full-map wave-erosion localization from largest to smallest based on indicators such as area and intensity, and identifies the top n wave-eroded areas as high-risk samples, which are then submitted for manual review first. n is a preset threshold parameter.

[0075] Furthermore, both the target detection model and the wave polishing detection model were trained and optimized using the Adam optimizer, with the learning rate optimized using a cosine annealing strategy. L2 regularization and Dropout techniques were introduced to prevent overfitting. During training, cross-validation was used to evaluate model performance, and hyperparameters were dynamically adjusted based on the performance on the validation set. The training termination condition was set to no significant improvement in mAP (mean accuracy) on the validation set for 5 consecutive epochs.

[0076] On the other hand, in one specific embodiment, a rail corrugation detection device based on computer vision includes a high-definition vision sensor, a digital signal acquisition unit, an A / D analog-to-digital conversion circuit, an industrial control computer, sensor matching signal lines, and an equipment mounting beam.

[0077] The high-definition vision sensor is fixed above the rail via the equipment mounting beam to collect images of the rail surface in real time and generate image signals.

[0078] The image signal acquired by the high-definition vision sensor is transmitted to the digital signal acquisition unit via the sensor's matching signal line; if the image signal is an analog signal, it is converted into a digital signal by an A / D analog-to-digital converter circuit and then transmitted to the digital signal acquisition unit; if the image signal is a digital signal, it is directly transmitted to the data signal acquisition unit via the sensor's matching signal line.

[0079] The digital signal acquisition unit performs preliminary processing and buffering of digital signals to generate image data, and transmits the image data to the industrial control computer via a high-speed data cable;

[0080] The industrial control computer, as the core processing unit of the system, is responsible for receiving, storing, and analyzing the acquired image data, running the rail corrugation detection method based on computer vision, and generating a full-image corrugation location.

[0081] Furthermore, high-speed data cables can use USB, Ethernet, etc.

[0082] Furthermore, the industrial control computer uploads the full-map corrugation positioning and raw image data to a remote server or cloud platform via a wireless network connection component, enabling remote monitoring and data sharing. The remote server or cloud platform can be configured with a visual interface to intuitively display the inspection results. The entire device automates and enables real-time processing from data acquisition to result output during the inspection process, improving the efficiency and accuracy of rail corrugation inspection and providing comprehensive technical support for the safe operation of rail transit systems.

[0083] Furthermore, the wireless network connection component can be a Wi-Fi, 4G / 5G module, etc.

[0084] Furthermore, it also includes power supply equipment and accessories, providing a stable power supply for all hardware such as the high-definition vision sensor, data signal acquisition unit, A / D analog-to-digital conversion circuit, and industrial control computer, ensuring continuous and reliable system operation. This structural design constitutes a complete rail corrugation detection device, ensuring efficient operation throughout the entire process from data acquisition to result output.

[0085] Furthermore, the high-definition vision sensor employs high-definition track cameras. Five high-definition track cameras are installed on the equipment mounting beam under the vehicle of the comprehensive track defect detection vehicle, and are evenly distributed along the transverse direction of the vehicle to ensure complete coverage of the entire rail surface. The equipment mounting beam adopts a vibration reduction design to reduce the impact of vehicle vibration on image quality during operation.

[0086] Furthermore, the industrial PC can be equipped with an alarm module that automatically sends information to relevant personnel when severe corrugation is detected. It also features a data storage and analysis module, supporting historical data queries and trend analysis. As an edge computing device, the industrial PC can utilize a low-power, high-performance embedded processor, possessing sufficient computing power to run trained deep learning models. To adapt to complex field environments, edge computing devices can also adopt shockproof, dustproof, and wide-temperature designs, ensuring stable operation under various harsh conditions.

[0087] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0088] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A computer vision-based method for detecting rail corrugation, characterized in that, Includes the following steps: Step 1: Acquire images of the rail surface and perform preprocessing; Step 2: Perform rail region detection on the preprocessed rail surface image and extract rail sub-images; Step 3: Detect the corrugation location on the rail sub-map and obtain the coordinates of the corrugated area; Step 4: Map the coordinates of the corrugated area to the coordinate system of the rail surface image to obtain the full-image corrugation location.

2. The rail corrugation detection method based on computer vision according to claim 1, characterized in that, The preprocessing of the rail surface image includes normalization and size adjustment.

3. The rail corrugation detection method based on computer vision according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 21: Use the target detection model to infer the preprocessed rail surface image, output the rail region bounding box, and obtain the rail region coordinates; Step 22: Crop the preprocessed rail surface image according to the rail area coordinates to obtain a rail sub-image.

4. The rail corrugation detection method based on computer vision according to claim 1, characterized in that, The specific implementation process of step 3 is as follows: Step 31: Perform feature enhancement processing on the vertical features of the rail sub-image to obtain an enhanced image; Step 32: Use the ripple detection model to infer the enhanced image, determine the mask and bounding box of the ripple region, and obtain the coordinates of the ripple region.

5. The rail corrugation detection method based on computer vision according to claim 3, characterized in that, The object detection model uses the YOLOv5 network, and the training process is accelerated through transfer learning.

6. The rail corrugation detection method based on computer vision according to claim 4, characterized in that, The ripple detection model includes a feature extraction and asymmetric downsampling module, and a deep feature fusion and detection head module; The feature extraction and asymmetric downsampling module includes an initial convolutional layer, an asymmetric downsampling layer, and multiple lightweight convolutional blocks connected in sequence. The initial convolutional layer performs standard convolution on the enhanced image to extract basic features. The asymmetric downsampling layer uses convolution or pooling operations with different downsampling ratios in the height and width dimensions to downsample the basic features and obtain the initial downsampled features. The multiple lightweight convolutional blocks further extract multi-scale features from the initial downsampled features to obtain downsampled multi-scale features. The deep feature fusion and detection head module includes an asymmetric downsampling layer, a feature fusion layer, and a detection head connected in sequence. The asymmetric downsampling layer uses convolution or pooling operations with different downsampling ratios in the height and width dimensions to downsample the downsampled multi-scale features and obtain standard multi-scale features. feature The fusion layer fuses standard multi-scale features and downsampled multi-scale features to obtain fused features; The detection head outputs the erosion confidence and bounding box coordinates for each location based on the fusion features. It obtains the mask of the erosion region based on the erosion confidence and generates the bounding box based on the bounding box coordinates.

7. The rail corrugation detection method based on computer vision according to claim 1, characterized in that, Step 1 involves acquiring multiple images of the rail surface from one perspective or multiple images of the rail surface from multiple perspectives. The specific implementation process of step 4 is as follows: Step 41: Map the coordinates of the corrugated area of ​​each rail surface image to the coordinate system of the corresponding rail surface image; Step 42: Fuse all rail surface images, remove duplicate and false detection images, and obtain full-image corrugation localization.

8. The rail corrugation detection method based on computer vision according to claim 1, characterized in that, The process also includes step 5, which involves sorting all wave-eroded areas in the full-map wave-erosion localization from largest to smallest based on their area and intensity indices, and identifying the top n wave-eroded areas as high-risk samples for priority submission for manual review.

9. A computer vision-based rail corrugation detection device, characterized in that, The rail corrugation detection method based on computer vision according to any one of claims 1-8 includes a high-definition vision sensor, a digital signal acquisition unit, an A / D analog-to-digital conversion circuit, an industrial control computer, sensor matching signal lines, and an equipment mounting beam; The high-definition vision sensor is fixed above the rail via the equipment mounting beam to collect images of the rail surface in real time and generate image signals. The image signal acquired by the high-definition vision sensor is transmitted to the digital signal acquisition unit via the sensor's matching signal line; if the image signal is an analog signal, it is converted into a digital signal by an A / D analog-to-digital converter circuit and then transmitted to the digital signal acquisition unit; if the image signal is a digital signal, it is directly transmitted to the data signal acquisition unit via the sensor's matching signal line. The digital signal acquisition unit performs preliminary processing and buffering of digital signals to generate image data, and transmits the image data to the industrial control computer via a high-speed data cable; The industrial control computer receives, stores, and analyzes the collected image data, runs a computer vision-based rail corrugation detection method, and generates a full-image corrugation location.

10. A rail corrugation detection device based on computer vision according to claim 9, characterized in that, The industrial control computer uploads the full-map corrugation positioning and raw image data to a remote server or cloud platform via a wireless network connection component.

Images