Visual image-based catenary hot defect geometric parameter false alarm filtering method

Abstract

The application discloses a catenary hot defect geometric parameter false alarm filtering method based on visual images, relates to railway catenary operation state monitoring technology, and discloses the catenary hot defect geometric parameter false alarm filtering method based on visual images. Through analysis of alarm data, extraction of out-of-limit information, sorting processing and a rejudgment mechanism of multi-source visual information, the disclosed method can effectively distinguish between real out-of-limit and false alarms caused by environmental interference, can effectively identify and filter geometric parameter false alarms, can reduce the workload of manual review, can improve the accuracy and reliability of the alarm system, and can thus guarantee the operation safety of the railway catenary.

Description

Technical Field

[0001] This application relates to railway catenary operation status monitoring technology, and in particular to a method for filtering false alarms of geometric parameters of thermal defects in catenary based on visual images. Background Technology

[0002] During real-time monitoring, the vehicle-mounted overhead contact line operation status detection device needs to continuously calculate geometric parameters such as pull-out value and guide height value to assess the system's operating status. The actual operating environment is complex and variable, with factors such as dense station structures, undulating mountainous terrain, dim lighting inside tunnels, and vibration interference in bridge areas, which can easily lead to distorted detection data and a large number of false geometric parameter alarms. These false alarms significantly increase the manual verification burden on maintenance personnel, drastically reducing the efficiency of effective alarm processing. Furthermore, frequent false alarms can cause staff fatigue, potentially leading to the overlooking of real safety hazards and the failure of the safety warning mechanism. Existing technologies have extensively researched overhead contact line geometric parameter detection algorithms, but these mainly focus on parameter calculation under ideal conditions, performing reasonably well against high-definition, clean backgrounds. However, when faced with complex scenarios such as insufficient bow-line contrast, strong interference from soft and hard cross-section structures, degraded image quality, or special areas (such as tunnel entrances and exits), the robustness of these algorithms is severely insufficient. For example, template matching methods suffer from a sharp drop in accuracy when the pantograph is partially obscured or the background is cluttered; contact point detection is prone to errors due to feature blurring or interference; and single-frame image analysis struggles to distinguish between actual over-limit conditions and environmental noise, resulting in a persistently high false alarm rate. Literature review indicates that current technologies lack post-processing mechanisms for alarm data, particularly automated solutions that can combine multi-source visual information for false alarm screening, thus failing to effectively address the false alarm filtering needs under complex operating conditions.

[0003] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this application is to provide a method for filtering false alarms of geometric parameters of thermal defects in contact wires based on visual images, which aims to effectively identify and filter false alarms of geometric parameters, thereby improving the accuracy and reliability of the alarm system.

[0005] To achieve the above objectives, this application proposes a method for filtering false alarms of geometric parameters of thermal defects in overhead contact lines based on visual images. The method includes:

[0006] By analyzing the alarm data of the vehicle-mounted overhead contact line geometric parameters, image frame data and geometric detection result data are obtained; the geometric detection result data includes the detected geometric value data and its corresponding image acquisition time information;

[0007] The geometric detection results data are analyzed and processed to extract the out-of-limit geometric value data and its corresponding image information data.

[0008] The data is sorted according to the degree of exceeding the limit in the geometric values ​​to obtain sorted over-limit frame data; if the geometric detection result data is that the double-branch pull-out exceeds the limit, then the sorting is based on the size of the distance between the two branches.

[0009] The sorted out-of-limit frame data is re-judged to confirm whether it exceeds the limit in order to filter out false alarms, and the judgment result data is output. The re-judgment process includes: using the reference camera image data in the sorted out-of-limit frame data to confirm the contact point data; using the visible light image data in the sorted out-of-limit frame data to obtain depth estimation image data; using the depth estimation image data to detect the pantograph area data; using the depth estimation image data and the pantograph area data to detect the contact wire position data; using the contact point data, the pantograph area data and the contact wire position data to calculate geometric value data; comparing the geometric value data with a preset standard value; and determining whether it is a false alarm based on the comparison result.

[0010] In one embodiment, the sorting process based on the distance between the two branches includes:

[0011] When the geometric detection result data shows that the double-bracket pullout exceeds the limit, the sum of the absolute values ​​of the double-bracket pullout values ​​is calculated to obtain the double-bracket spacing data;

[0012] The data is sorted by priority based on the size of the bi-branch spacing data to obtain the sorted over-limit frame data.

[0013] In one embodiment, the step of confirming contact point data using reference camera image data from the sorted over-limit frame data includes:

[0014] The pixel value comparison process is performed on the reference camera image data in the sorted over-limit frame data; the pixel value comparison process is to compare the pixel value of the current point to be confirmed with the pixel values ​​of its left neighbor and right neighbor within its neighborhood range respectively;

[0015] If the pixel value of the current point to be confirmed is greater than the pixel value of both the left and right neighboring points, then reliable contact point data is output as contact point data; otherwise, a neighborhood search process is performed to reconfirm the contact point data. The neighborhood search process involves searching for points that meet the contrast conditions within the neighborhood, centered on the current point to be confirmed, and growing several points upwards to verify reliability.

[0016] In one embodiment, the step of obtaining depth estimation image data using visible light image data from the sorted over-limit frame data includes:

[0017] The visible light image data in the sorted over-limit frame data is input into a monocular depth estimation model for processing; the monocular depth estimation model generates depth estimation image data by predicting the depth value of each pixel; wherein, pixels closer to the camera have higher brightness values ​​and pixels farther away have lower brightness values.

[0018] In one embodiment, the step of detecting pantograph region data using the depth estimation image data includes:

[0019] The depth estimation image data is input into a trained deep learning object detection model for inference processing; the inference processing includes segmenting the depth estimation image data, detecting object features, performing model classification and bounding box regression processing, and outputting pantograph area data.

[0020] In one embodiment, the step of detecting contact wire position data using the depth estimation image data and pantograph area data includes:

[0021] Input the depth estimation image data and the pantograph area data, and find the location data of the first point with the largest brightness value in the depth estimation image data above the pantograph area data;

[0022] The first line segment point set data is obtained by performing downward growth processing based on the first point position data; the downward growth processing is to determine whether the growth length conforms to the line characteristics, and if it does not conform, the first point position data with the largest brightness value is searched again.

[0023] The first line segment point set data is fitted to output the contact line position data.

[0024] In one embodiment, the step of detecting contact wire position data using the depth estimation image data and pantograph area data further includes:

[0025] When no contact line position data is output, the depth estimation image data and pantograph area data are input, and the second point position data that meets the contrast condition is found; the point that meets the contrast condition is the point whose pixel value differs from the neighboring pixel value by more than a preset threshold.

[0026] The second point location data is used to perform growth processing to generate the second line segment point set data, and the second line segment point set data is then fitted to generate candidate line feature data.

[0027] Compare the mean brightness values ​​of all candidate line feature data, and select the line feature data with the largest mean brightness value as the contact line position data.

[0028] In one embodiment, the step of calculating geometric data using the contact point data, pantograph area data, and contact wire position data includes:

[0029] Input the contact point data, pantograph area data, and contact wire position data;

[0030] The pixel values ​​are converted into physical values ​​based on the calibration model, and geometric value data is output. The calibration model is a preset correspondence between pixel values ​​and physical values.

[0031] In one embodiment, the step of comparing the geometric value data with a preset standard value includes:

[0032] Input the geometric data and compare it with a preset standard value;

[0033] If the geometric value data meets the over-limit condition, the judgment result data is a valid alarm; otherwise, it is a false alarm and is filtered; wherein, the over-limit condition is that the absolute value of the geometric value is greater than the preset pull-out standard value or the guide height value exceeds the preset guide height range.

[0034] In one embodiment, the step of confirming the contact point data using the reference camera image data in the sorted over-limit frame data further includes:

[0035] Based on the image acquisition time information, search for other camera image data at the same time to use them for auxiliary judgment.

[0036] The proposed method for filtering false alarms of geometric parameters of thermal defects in overhead contact lines based on visual images effectively distinguishes between genuine over-limits and false alarms caused by environmental interference through parsing alarm data, extracting over-limit information, sorting and processing, and a re-judgment mechanism based on multi-source visual information. It can effectively identify and filter false alarms of geometric parameters, reduce the workload of manual verification, improve the accuracy and reliability of the alarm system, and thus ensure the safe operation of railway overhead contact lines. Attached Figure Description

[0037] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0038] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating an embodiment of the method for filtering false alarms of geometric parameters of thermal defects in contact wires based on visual images, as provided in this application.

[0040] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0041] The technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of this application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0042] It should be understood that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0043] In existing technologies, research on catenary geometric parameter detection algorithms is relatively extensive, but it mainly focuses on parameter calculation under ideal conditions, performing reasonably well against high-definition, clean backgrounds. However, when faced with complex scenarios such as insufficient pantograph-line contrast, strong interference from soft and hard cross-section structures, degraded image quality, or special areas (such as tunnel entrances and exits), the robustness of these algorithms is severely insufficient. For example, template-matching methods experience a sharp drop in matching accuracy when the pantograph is partially obscured or the background is cluttered; contact point detection is prone to errors due to feature blurring or interference points; and single-frame image analysis struggles to distinguish between actual exceedances and environmental noise, resulting in a persistently high false alarm rate. Literature review indicates that current technologies lack post-processing mechanisms for alarm data, particularly automated solutions that can combine multi-source visual information for false alarm screening, failing to effectively address the false alarm filtering needs under complex operating conditions.

[0044] Based on this, embodiments of this application provide a method for filtering false alarms of geometric parameters of thermal defects in overhead contact lines based on visual images, referring to... Figure 1 The method for filtering false alarms of contact wire thermal defect geometric parameters based on visual images includes steps S100 to S400, wherein:

[0045] Step S100: By parsing the vehicle-mounted overhead contact line geometric parameter alarm data, image frame data and geometric detection result data are obtained; the geometric detection result data includes the detected geometric value data and its corresponding image acquisition time information;

[0046] Step S200: Analyze and process the geometric detection result data to extract the out-of-limit geometric value data and its corresponding image information data;

[0047] Step S300: Sort the data according to the degree of exceeding the limit of the geometric value data to obtain sorted exceeding frame data; wherein, if the geometric detection result data is that the double-branch pull-out exceeds the limit, sorting is performed based on the size of the distance between the two branches.

[0048] Step S400: The sorted out-of-limit frame data is re-judged to confirm whether it exceeds the limit in order to filter out false alarms, and the judgment result data is output. The re-judgment process includes: using the reference camera image data in the sorted out-of-limit frame data to confirm the contact point data; using the visible light image data in the sorted out-of-limit frame data to obtain depth estimation image data; using the depth estimation image data to detect the pantograph area data; using the depth estimation image data and the pantograph area data to detect the contact wire position data; using the contact point data, the pantograph area data and the contact wire position data to calculate geometric value data; comparing the geometric value data with a preset standard value; and determining whether it is a false alarm based on the comparison result.

[0049] In this embodiment, parsing the vehicle-mounted overhead contact line geometric parameter alarm data is the process of extracting information related to geometric parameters from the alarm system. This can be achieved by reading alarm files or receiving real-time data streams, such as using file parsing tools to extract structured data or receiving and parsing data packets through a network interface. The main purpose is to obtain basic data related to geometric parameters. During the acquisition of image frame data and geometric detection result data, the image frame data can originate from the vehicle-mounted camera or other image acquisition devices, while the geometric detection result data is generated by a geometric parameter detection algorithm. The purpose is to ensure time synchronization between the image and the geometric values. Extracting out-of-limit geometric value data and its corresponding image information can be achieved in various ways. For example, threshold filtering can be applied to the geometric detection result data to extract geometric value data exceeding a preset range, or a classification model can be used to label and filter the geometric detection result data. This is mainly to focus on potential problem areas and reduce the processing burden of irrelevant data.

[0050] In this embodiment, sorting based on the distance between the two supports can be understood as prioritizing data according to the distance relationship between the pull-out values ​​of the two supports. This can be achieved by calculating the absolute value of the difference between the pull-out values ​​of the two supports and sorting them according to their numerical values. For example, a quicksort algorithm or a bubble sort algorithm can be used to complete the sorting operation. The purpose is to optimize the processing order to quickly identify high-risk false alarms. In addition, confirming contact point data using reference camera image data can be achieved in various ways. For example, high-contrast regions can be extracted as candidate contact point regions based on the grayscale distribution of the image, or significant feature points in the image can be identified through edge detection algorithms. This is mainly to overcome the problem of misjudgment caused by unclear features or interference points. Obtaining depth estimation image data using visible light image data can be achieved in ways other than monocular depth estimation techniques. For example, the 3D scene structure can be reconstructed from multi-view images, or depth information can be inferred using known scene prior information. The purpose is to provide scene distance information to assist target recognition.

[0051] Detecting pantograph region data using depth estimation image data can be achieved in various ways. For example, specific regions can be extracted as candidate pantograph regions based on image segmentation algorithms, or template matching techniques can be used to locate the pantograph position. The aim is to accurately locate the pantograph region to avoid background interference. Detecting contact wire position data using both depth estimation image data and pantograph region data can also be achieved in various ways. For example, candidate contact wire positions can be extracted based on brightness gradient changes, or straight line features can be detected using Hough transform. The aim is to reliably locate the contact wire to solve identification problems in complex environments.

[0052] This application integrates visual image analysis and deep learning technologies to systematically verify geometric parameter alarms, thereby effectively distinguishing between real defects and false alarms caused by environmental interference. Compared to the false alarm problems caused by complex environmental interference such as stations, mountains, tunnels, and bridges in existing technologies, this application improves the accuracy of geometric parameter detection and reduces the need for manual screening through comprehensive analysis and recalculation of multi-source data.

[0053] This application obtains image frame data and geometric detection result data by parsing the geometric parameter alarm data of the vehicle-mounted overhead contact system. The geometric detection result data includes the detected geometric values ​​and their corresponding image acquisition time information, ensuring precise time synchronization between the image and the geometric values, providing a reliable basis for subsequent analysis. Furthermore, the geometric detection result data is analyzed and processed to extract the out-of-limit geometric value data and its corresponding image information, focusing on potential problem areas, reducing the processing burden of irrelevant data, and improving resource utilization efficiency. The out-of-limit geometric value data is sorted according to the degree of exceeding the limit to obtain sorted out-of-limit frame data. If the geometric detection result data shows an out-of-limit situation with two branches extended, the sorting is based on the distance between the two branches, prioritizing frames with more severe out-of-limit situations or those more susceptible to interference, optimizing the processing order to quickly identify high-risk false alarms. Specifically, when re-processing the sorted out-of-limit frame data, reference camera image data is used to confirm the contact point data, and high-contrast images are used to accurately identify the contact point location, overcoming the influence of indistinct features or interference points; visible light image data is used to obtain depth estimation image data, and a monocular depth estimation model is used to infer the 3D depth structure from the 2D image, providing scene distance information to assist in identifying targets in complex backgrounds; depth estimation image data is used to detect pantograph area data, and combined with depth information, the pantograph is accurately located, avoiding recognition failures caused by background interference; depth estimation image data and pantograph area data are used to detect contact wire position data, and the contact wire is reliably located using depth information and area constraints, solving the problem of insufficient pantograph line contrast or interference between soft and hard cross-sections; geometric value data is calculated using contact point data, pantograph area data, and contact wire position data, and physical geometric parameters are recalculated based on multi-source data to improve measurement accuracy; the geometric value data is compared with preset standard values, and the comparison results are used to determine whether there are false alarms. Invalid alarms are filtered through rigorous verification to ensure that the output results only contain real defects, thereby efficiently retaining valid alarms and reducing manual intervention. Therefore, this method integrates visual image analysis and deep learning technology to systematically verify geometric parameter alarms, effectively distinguish between real defects and false alarms caused by environmental interference, thereby improving alarm accuracy and reducing the need for manual screening.

[0054] In one feasible implementation, sorting based on the size of the two-branch spacing includes: when the geometric detection result data shows that the two-branch pull-out exceeds the limit, calculating the sum of the absolute values ​​of the two-branch pull-out values ​​to obtain the two-branch spacing data; and performing priority sorting based on the size of the two-branch spacing data to obtain the sorted over-limit frame data.

[0055] In this embodiment, the dual-branch spacing data refers to a numerical index generated by quantifying the comprehensive offset of the dual-branch pull-out values. This can be achieved by summing the absolute values ​​of the dual-branch pull-out values ​​to eliminate directional interference and unify the evaluation criteria. Priority sorting can be understood as arranging the out-of-limit frame data according to the magnitude of the dual-branch spacing data. This can be implemented using common sorting algorithms (such as quicksort and mergesort) to ensure that severely out-of-limit frame data is processed first.

[0056] In this embodiment, the proposed solution effectively addresses the issue of ambiguous sorting criteria when two supports extend beyond the limit by clearly defining the calculation logic and sorting mechanism for the dual-support spacing data. In actual operation, firstly, for cases where two supports extend beyond the limit, the extension value of each support is extracted and the sum of their absolute values ​​is calculated, thereby generating dual-support spacing data. This process transforms the previously difficult-to-quantify overall offset of the two supports into a specific numerical indicator, avoiding evaluation bias caused by directional differences. Subsequently, based on the generated dual-support spacing data, the over-limit frame data is prioritized according to its numerical value, ensuring that frames with more significant overall offset are selected first. This design not only improves the objectivity and consistency of the sorting results but also significantly enhances the targeting and efficiency of false alarm filtering. Furthermore, this solution is closely integrated with the aforementioned analysis and processing flow of geometric detection result data. By accurately quantifying and sorting cases of dual-support extension exceeding the limit, it further optimizes the filtering effect of over-limit frame data, thereby reducing the workload of manual review and lowering the risk of missing genuine alarms.

[0057] In one feasible implementation, the step of confirming contact point data using reference camera image data in the sorted over-limit frame data includes: performing pixel value comparison processing on the reference camera image data in the sorted over-limit frame data; the pixel value comparison processing involves comparing the pixel value of the current point to be confirmed with the pixel values ​​of its left and right neighboring points within its neighborhood range; if the pixel value of the current point to be confirmed is greater than both the pixel values ​​of its left and right neighboring points, then reliable contact point data is output as contact point data; otherwise, a neighborhood search process is performed to reconfirm the contact point data; the neighborhood search process involves searching for points that meet the contrast conditions within the neighborhood range centered on the current point to be confirmed, and growing upwards by several points to verify reliability.

[0058] In this embodiment, pixel value comparison processing refers to identifying local peak features by analyzing the brightness difference between the current point to be confirmed and its neighboring points. This can be achieved using methods such as pixel-by-pixel scanning and window sliding. The purpose of introducing this technique is to eliminate global noise interference through local comparison, thereby accurately locating contact points in complex environments. Neighborhood search processing can be understood as a candidate point search mechanism based on contrast filtering. It can be implemented by setting a threshold for the difference between the pixel value and the mean of the neighborhood or a gradient change threshold, aiming to solve the problem that contact points are difficult to directly identify in low-contrast areas. Growing several points upwards to verify reliability is a verification operation simulating the continuity of the contact line. This can be accomplished by gradually expanding along the vertical direction and statistically analyzing the continuity features, with the aim of ensuring that the confirmed contact points conform to actual physical characteristics.

[0059] In this embodiment, the proposed solution achieves accurate identification of contact points in complex environments through a series of ordered technical steps. First, based on pixel value comparison processing, candidate points with local advantages are quickly identified by utilizing the brightness relationship between the current point to be identified and its left and right neighboring points. This local comparison method effectively avoids interference from global noise, and is particularly suitable for scenarios with weak bow line contrast. Second, when the initial comparison fails to meet the conditions, a neighborhood search process is triggered. Potential contact points are filtered out by setting contrast conditions. This mechanism significantly reduces the false positive rate in low-contrast areas. Finally, by growing several points upwards, the geometric continuity of the candidate points is further verified. Only the set of points that conforms to the physical characteristics of the contact line is ultimately confirmed, thus eliminating the influence of isolated noise points or temporary interference points. These steps work together to form a complete process from initial screening to fine verification, ensuring high reliability of contact point identification and providing a solid foundation for subsequent geometric parameter calculations. Furthermore, this solution, combined with sorted over-limit frame data, further improves the efficiency and accuracy of false alarm filtering.

[0060] In one feasible implementation, the step of obtaining depth estimation image data using visible light image data from the sorted over-limit frame data includes: inputting the visible light image data from the sorted over-limit frame data into a monocular depth estimation model for processing; the monocular depth estimation model generates depth estimation image data by predicting the depth value of each pixel; wherein, pixels closer to the camera have higher brightness values ​​and pixels farther away have lower brightness values.

[0061] In this embodiment, the monocular depth estimation model refers to a deep learning-based algorithm structure, which can be implemented using convolutional neural networks, generative adversarial networks, or autoencoders. The purpose of introducing this model is to infer three-dimensional depth information from two-dimensional visible light images, thereby providing a reliable basis for subsequent false alarm filtering. Depth estimation image data can be understood as image data that expresses depth information in the form of brightness features, which can be visualized by assigning different brightness values ​​to each pixel.

[0062] In this embodiment, the problem of inaccurate depth information acquisition in complex scenes can be effectively solved by the above-mentioned technical solution. First, by using the existing visible light images of the vehicle system as the input source, the reliance on additional sensors is avoided, and the deep learning model can adapt to environments with strong interference, such as stations and tunnels. Second, the monocular depth estimation model generates depth estimation image data by predicting the depth value of each pixel, simplifying the computational complexity of the real-time vehicle system and ensuring that continuous depth representation can still be generated in special areas. Furthermore, by adopting an encoding method that prioritizes pixels closer to the camera with higher brightness values ​​and pixels farther away with lower brightness values, depth information is intuitively converted into brightness features, enabling subsequent detection steps to reliably locate targets based on brightness differences. In addition, this solution, combined with the aforementioned steps of acquiring out-of-limit geometric value data and its corresponding image information data, can more accurately identify the contact line position, effectively suppress the influence of background interference and ambient light changes, thereby improving the accuracy of false alarm filtering.

[0063] In one feasible implementation, the step of detecting pantograph region data using the depth estimation image data includes: inputting the depth estimation image data into a trained deep learning target detection model for inference processing; the inference processing includes segmenting the depth estimation image data, detecting target features, performing model classification and bounding box regression processing, to output pantograph region data.

[0064] In this embodiment, depth estimation image data refers to image data containing distance information generated by a monocular depth estimation model, which can be achieved by predicting the depth value of each pixel. In practical applications, deep learning object detection models can be detection models based on convolutional neural networks, such as YOLO and Faster R-CNN, aiming to improve the accuracy of pantograph area detection in complex scenes. Segmenting the depth estimation image data can be understood as dividing the image into multiple sub-regions, specifically implemented through semantic segmentation or instance segmentation techniques, aiming to reduce overall processing complexity and focus on potential target areas. Detecting target features refers to extracting the contour and structural features of the pantograph from the segmented regions, which can be achieved through edge detection algorithms or feature extraction networks, aiming to enhance the recognition ability of key visual attributes. Performing model classification refers to distinguishing the pantograph area from interfering objects based on the extracted features, which can be achieved through support vector machines or deep learning classifiers, aiming to reduce background misclassification. Bounding box regression processing can be understood as the process of dynamically adjusting the position and size of the bounding box, which can be achieved through optimization algorithms, aiming to accurately match the actual shape and position of the pantograph.

[0065] In this embodiment, the proposed solution processes depth estimation image data using a deep learning object detection model. It leverages distance information from the depth estimation image to enhance the model's understanding of the scene's spatial structure, avoiding the limitations of relying solely on visible light images in complex backgrounds. The segmentation operation decomposes the image into manageable regional units, reducing overall processing complexity and focusing on potential target areas. The target feature detection step extracts the pantograph's contour and structural features from the segmented regions, strengthening the recognition of key visual attributes. The model classification step distinguishes the pantograph region from interfering objects based on features, reducing background false positives. The bounding box regression processing step dynamically adjusts the bounding box position and size to ensure accurate matching of the pantograph's actual shape and position. The final output pantograph region data provides reliable input for subsequent geometric parameter calculations, ensuring accurate false alarm filtering. Based on this, combined with the aforementioned acquired depth estimation image data, it effectively addresses occlusion and background interference issues in complex scenes, significantly improving the robustness of pantograph region detection.

[0066] In one feasible implementation, the step of detecting contact wire position data using the depth estimation image data and pantograph area data includes: inputting the depth estimation image data and pantograph area data; finding the first point position data with the largest brightness value in the depth estimation image data above the pantograph area data; performing downward growth processing based on the first point position data to obtain a first line segment point set data; the downward growth processing involves determining whether the growth length conforms to line characteristics, and if not, re-finding the first point position data with the largest brightness value; and performing fitting processing on the first line segment point set data to output the contact wire position data.

[0067] In this embodiment, the first point location data refers to the pixel location information with the maximum brightness value in the depth estimation image data above the pantograph area data. This can be achieved through a traversal search algorithm or a local extremum detection algorithm, with the aim of quickly locating the possible initial point of the contact line. The downward growth process can be understood as extending vertically from the initial point and dynamically judging whether the growth length conforms to line characteristics. This can be achieved by setting a length threshold and angle constraints, ensuring that continuous linear structures are captured rather than isolated interference points. The first line segment point set data refers to a set of points conforming to line characteristics obtained through the downward growth process. This can be stored using a linked storage structure or an array, with the aim of providing a reliable data foundation for subsequent fitting processing.

[0068] In this embodiment, the scheme combines depth estimation image data with pantograph area data, limiting the search range to above the pantograph area. Utilizing the characteristic that pixels closer to the camera in the depth image have higher brightness values, it prioritizes bright areas as initial points, thus avoiding background interference introduced by full-image search. When performing downward growth processing based on the location data of the first point with the highest brightness value, it dynamically judges whether the growth length conforms to line characteristics, effectively distinguishing between real contact lines and isolated interference points. If the growth length does not conform to line characteristics, the point with the highest brightness is searched again. This iterative mechanism ensures that continuous linear structures can be captured even in complex environments. By fitting the first line segment point set data, the influence of local noise is eliminated, outputting a stable geometric position, providing accurate input for subsequent geometric value calculations. The entire process, through the design of brightness localization, growth verification, and fitting output, significantly improves the robustness of contact line detection in complex environments, solves the problem of contact line position recognition failure caused by non-contact line interference points or environmental interference, and thus improves the reliability of false alarm filtering.

[0069] In one feasible implementation, the step of detecting contact wire position data using the depth estimation image data and pantograph area data further includes: when no contact wire position data is output, inputting the depth estimation image data and pantograph area data, and searching for second point position data that meets the contrast condition; the point that meets the contrast condition is a point whose pixel value differs from the neighboring pixel value by a preset threshold; performing growth processing based on the second point position data to generate a second line segment point set data, and performing fitting processing on the second line segment point set data to generate candidate line feature data; comparing the average brightness of all candidate line feature data, and selecting the line feature data with the largest average brightness as the contact wire position data.

[0070] In this embodiment, the second point location data refers to the point in the depth estimation image data where the difference between the pixel value and its neighboring pixel values ​​exceeds a preset threshold. In practical applications, this can be achieved by calculating the grayscale difference between each pixel and its eight neighboring pixels. The purpose is to utilize the significant contrast characteristics of the contact line edge to avoid false detections caused by ambient light interference. The second line segment point set data can be understood as a set of points continuously extended along the contact line direction from high-contrast points. It can be implemented using a region growing algorithm based on seed points, aiming to simulate the geometric shape of the contact line and overcome the defect that isolated points cannot represent complete line features. Candidate line feature data refers to smooth line features generated by mathematically fitting a discrete point set. It can be implemented using the least squares method or the RANSAC algorithm, aiming to eliminate the influence of noise points and improve the reliability of line features.

[0071] In this embodiment, the proposed solution provides a robust backup mechanism to address the initial contact line position detection failure by combining contrast features and depth information. First, without outputting contact line position data, the system analyzes depth estimation image data and pantograph area data to find a second point position that meets the contrast criteria. This step fully utilizes the stability characteristics of the contact line edge in complex scenes. Subsequently, a second line segment point set is generated based on the second point position data through a growth process. This process expands the point set along the continuity direction of the contact line, effectively simulating the true geometry of the contact line. Next, candidate line feature data is generated by fitting the second line segment point set data. This operation transforms the discrete point set into smooth line features, thereby improving the reliability of the candidate lines. Finally, by comparing the average brightness of all candidate line feature data and selecting the largest value as the contact line position data, the system leverages the characteristic that objects closer to the camera in the depth estimation image have higher brightness, ensuring effective differentiation between background interference and the real target.

[0072] Based on this, this scheme forms an organic whole with the aforementioned steps. By introducing contrast conditions and brightness mean screening mechanisms, it not only solves the problem of backup strategies when the initial method fails, but also further enhances the accuracy of contact line position identification under complex interference environments, thereby ensuring the continuity and reliability of false alarm filtering.

[0073] In one feasible implementation, the step of calculating geometric data using the contact point data, pantograph area data, and contact wire position data includes: inputting the contact point data, pantograph area data, and contact wire position data; converting pixel values ​​into physical values ​​based on a calibration model, and outputting geometric data; wherein, the calibration model is a preset correspondence between pixel values ​​and physical values.

[0074] In this embodiment, contact point data refers to data identifying the actual interaction position between the pantograph and the contact wire, which can be obtained by comparing pixel values ​​or performing neighborhood search processing on reference camera image data. Pantograph area data can be understood as data defining the physical boundary range of the pantograph, which can be generated by inference processing of depth estimation image data using a deep learning object detection model. Contact wire position data describes the spatial trajectory characteristics of the contact wire, which can be generated by finding and fitting the point with the maximum brightness value. The purpose of introducing these multi-source visual data is to comprehensively capture the dynamic changes in geometric relationships and avoid the bias caused by occlusion or interference from a single data source in complex scenes. The calibration model refers to the preset mapping relationship between pixel values ​​and physical values, which can be achieved by calibrating systematic errors such as lens distortion and installation angle during the camera calibration process. In particular, in environments prone to perspective distortion, such as tunnels and bridges, the calibration model can accurately map image coordinates to actual physical dimensions, thereby ensuring the physical comparability of geometric parameters such as pull-out value and guide height value.

[0075] In this embodiment, the above-mentioned scheme integrates multi-source visual data and introduces a calibration model to achieve accurate conversion from pixels to physical values, fundamentally improving the accuracy of geometric parameter calculation. Based on this, contact point data, pantograph area data, and contact wire position data serve as the basis for calculation, each derived from the accurate identification of key elements of the overhead contact system in previous steps, and are input together to provide a complete and reliable basis. The calibration model, by calibrating system errors, solves the problem of pixel position distortion caused by image distortion, perspective effects, or background noise, making the geometric parameter calculation results more objective. The final output geometric value data directly reflects the true state of the overhead contact system, avoiding the inherent scale ambiguity of pixel values, thus reliably distinguishing between true exceedances and false alarms in subsequent comparisons, improving the robustness of the filtering process. In this way, the problem of significantly increased geometric parameter calculation errors under complex environmental interference is solved, the reliability of false alarm filtering is improved, the workload of manual screening is reduced, and more efficient technical support is provided for railway operation and maintenance.

[0076] In one feasible implementation, the step of comparing the geometric value data with a preset standard value includes: inputting the geometric value data and comparing it with the preset standard value; if the geometric value data meets the over-limit condition, the result data is judged as a valid alarm; otherwise, it is a false alarm and is filtered; wherein, the over-limit condition is that the absolute value of the geometric value is greater than the preset pull-out standard value or the guide height value exceeds the preset guide height range.

[0077] In this embodiment, geometric data refers to the data output after converting pixel values ​​into physical values ​​through a calibration model. This data can be calculated using contact point data, pantograph area data, and contact wire position data, aiming to ensure the objectivity and accuracy of the geometric parameters. The preset standard value can be understood as a safety threshold set according to actual operational requirements. It can be determined through historical data analysis or experimental testing, aiming to provide a clear basis for judgment. Exceeding the limit condition refers to the absolute value of the geometric value exceeding the preset pull-out standard value or the guide height value exceeding the preset guide height range. This can be achieved using dual judgment criteria, aiming to distinguish between different safety boundaries of the pull-out value and the guide height value, thereby avoiding the omission of actual thermal defects.

[0078] In this embodiment, the scheme ensures that the judgment process is based on objective data rather than subjective experience by inputting geometric value data and performing standardized comparisons, thus avoiding arbitrary judgments caused by unclear thresholds. In specific execution, the geometric value data is first compared with preset standard values. If the out-of-limit conditions are met, it is determined to be a valid alarm. This design directly targets the core characteristics of overhead contact line geometric parameter detection, ensuring that real thermal defects are not missed. Data that does not meet the out-of-limit conditions is automatically filtered as false alarms, reducing interference from invalid data and avoiding the waste of resources for manual review. Furthermore, the combination of the absolute value of the output and the range of the guide height value as dual judgment criteria adapts to the diversity of complex scenarios in overhead contact line operation, improving the targeting and accuracy of false alarm filtering, thereby significantly reducing the workload of maintenance personnel and improving the effectiveness of safety warnings.

[0079] Based on this, the above method solves the problem of ambiguity in false alarm filtering by accurately standardizing the comparison logic of geometric value data. Especially under complex environmental interference, it can efficiently and accurately filter false alarm data, retain valid alarms, reduce the workload of manual screening, and ensure the reliability of safety warnings.

[0080] In one feasible implementation, the step of confirming the contact point data using the reference camera image data in the sorted over-limit frame data further includes: searching for other camera image data at the same time based on the image acquisition time information, so as to use them to assist in the judgment.

[0081] In this embodiment, the image acquisition time information refers to the image acquisition timestamp precisely recorded by the vehicle system. This can be achieved using high-precision clock synchronization technology to ensure that the acquired images from other cameras strictly correspond to the scene state at the same physical moment as the reference image. The image data from other cameras can be understood as complementary perspective images captured by cameras at different installation locations. This can be achieved through a multi-camera collaborative acquisition system to provide uncontaminated visual information, thereby enhancing the reliability of contact point confirmation.

[0082] In this embodiment, the scheme effectively solves the problem of insufficient reliability of contact point identification in complex scenes by introducing a multi-source image collaborative verification mechanism. Based on the feature of searching other camera image data at the same time based on image acquisition time information, the timestamp accurately recorded by the on-board system is used as a synchronization benchmark, avoiding interference from dynamic scene changes caused by train operation. In complex environments such as tunnels or bridge areas, cameras at different installation positions may capture the catenary structure from complementary angles. When the reference camera is obstructed or interfered with by lighting, other camera images can provide uncontaminated visual information, establishing multi-dimensional data support for contact point confirmation. After integrating multi-view image data into the contact point confirmation process, the reliability of feature points can be verified by cross-comparing pixel features and spatial relationships in the pixel value comparison process, combined with the neighborhood contrast information of other images. When the neighborhood search of the reference image fails due to interference, the clear areas of other images can provide alternative judgment criteria, thereby significantly improving the robustness of contact point positioning and fundamentally reducing false alarms caused by misjudgment of single-point data. This not only improves the accuracy of contact point confirmation but also effectively reduces the problem of false alarm filtering failure caused by interference from complex environments, providing reliable technical support for railway operation and maintenance.

[0083] In the embodiments of this application, the method for filtering false alarms of geometric parameters of thermal defects in overhead contact lines based on visual images effectively distinguishes between true over-limits and false alarms caused by environmental interference through a mechanism of parsing alarm data, extracting over-limit information, sorting and processing, and re-judging multi-source visual information. It can effectively identify and filter false alarms of geometric parameters, reduce the workload of manual verification, improve the accuracy and reliability of the alarm system, and thus ensure the safe operation of railway overhead contact lines.

[0084] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. All equivalent structural transformations made under the technical concept of this application using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included within the scope of patent protection of this application.

Claims

1. A method for filtering false alarms of geometric parameters of thermal defects in overhead contact wires based on visual images, characterized in that, The method includes: By analyzing the alarm data of the vehicle-mounted overhead contact line geometric parameters, image frame data and geometric detection result data are obtained; the geometric detection result data includes the detected geometric value data and its corresponding image acquisition time information; The geometric detection results data are analyzed and processed to extract the out-of-limit geometric value data and its corresponding image information data. The data is sorted according to the degree of exceeding the limit in the geometric values ​​to obtain sorted over-limit frame data; if the geometric detection result data is that the double-branch pull-out exceeds the limit, then the sorting is based on the size of the distance between the two branches. The sorted out-of-limit frame data is re-evaluated to confirm whether it exceeds the limit in order to filter out false alarms, and the evaluation result data is output. The re-evaluation process includes: using the reference camera image data in the sorted out-of-limit frame data to confirm the contact point data; using the visible light image data in the sorted out-of-limit frame data to obtain depth estimation image data; using the depth estimation image data to detect the pantograph area data; using the depth estimation image data and the pantograph area data to detect the contact wire position data; using the contact point data, the pantograph area data, and the contact wire position data to calculate geometric value data; comparing the geometric value data with a preset standard value; and determining whether it is a false alarm based on the comparison result. Sorting based on the distance between two branches includes: When the geometric detection result data shows that the double-bracket pullout exceeds the limit, the sum of the absolute values ​​of the double-bracket pullout values ​​is calculated to obtain the double-bracket spacing data; The data is sorted by priority based on the size of the bi-branch spacing data to obtain the sorted over-limit frame data.

2. The method for filtering false alarms of contact wire thermal defect geometric parameters based on visual images as described in claim 1, characterized in that, The steps for verifying contact point data using the reference camera image data in the sorted over-limit frame data include: The pixel value comparison process is performed on the reference camera image data in the sorted over-limit frame data; the pixel value comparison process is to compare the pixel value of the current point to be confirmed with the pixel values ​​of its left neighbor and right neighbor within its neighborhood range respectively; If the pixel value of the current point to be confirmed is greater than the pixel value of both the left and right neighboring points, then reliable contact point data is output as contact point data; otherwise, a neighborhood search process is performed to reconfirm the contact point data. The neighborhood search process involves searching for points that meet the contrast conditions within the neighborhood, centered on the current point to be confirmed, and growing several points upwards to verify reliability.

3. The method for filtering false alarms of contact wire thermal defect geometric parameters based on visual images as described in claim 1, characterized in that, The steps for obtaining depth estimation image data using the visible light image data from the sorted over-limit frame data include: The visible light image data in the sorted over-limit frame data is input into a monocular depth estimation model for processing; the monocular depth estimation model generates depth estimation image data by predicting the depth value of each pixel; wherein, pixels closer to the camera have higher brightness values ​​and pixels farther away have lower brightness values.

4. The method for filtering false alarms of contact wire thermal defect geometric parameters based on visual images as described in claim 1, characterized in that, The steps for detecting pantograph region data using the depth estimation image data include: The depth estimation image data is input into a trained deep learning object detection model for inference processing; the inference processing includes segmenting the depth estimation image data, detecting object features, performing model classification and bounding box regression processing, and outputting pantograph area data.

5. The method for filtering false alarms of contact wire thermal defect geometric parameters based on visual images as described in claim 1, characterized in that, The steps for detecting contact wire position data using the depth estimation image data and pantograph area data include: Input the depth estimation image data and the pantograph area data, and find the location data of the first point with the largest brightness value in the depth estimation image data above the pantograph area data; The first line segment point set data is obtained by performing downward growth processing based on the first point position data; the downward growth processing is to determine whether the growth length conforms to the line characteristics, and if it does not conform, the first point position data with the largest brightness value is searched again. The first line segment point set data is fitted to output the contact line position data.

6. The method for filtering false alarms of contact wire thermal defect geometric parameters based on visual images as described in claim 5, characterized in that, The step of detecting contact wire position data using the depth estimation image data and pantograph area data further includes: When no contact line position data is output, the depth estimation image data and pantograph area data are input, and the second point position data that meets the contrast condition is found; the point that meets the contrast condition is the point whose pixel value differs from the neighboring pixel value by more than a preset threshold. The second point location data is used to perform growth processing to generate the second line segment point set data, and the second line segment point set data is then fitted to generate candidate line feature data. Compare the mean brightness values ​​of all candidate line feature data, and select the line feature data with the largest mean brightness value as the contact line position data.

7. The method for filtering false alarms of contact wire thermal defect geometric parameters based on visual images as described in claim 1, characterized in that, The steps for calculating geometric data using the contact point data, pantograph area data, and contact wire position data include: Input the contact point data, pantograph area data, and contact wire position data; The pixel values ​​are converted into physical values ​​based on the calibration model, and geometric value data is output. The calibration model is a preset correspondence between pixel values ​​and physical values.

8. The method for filtering false alarms of contact wire thermal defect geometric parameters based on visual images as described in claim 1, characterized in that, The steps for comparing the geometric data with preset standard values ​​include: Input the geometric data and compare it with a preset standard value; If the geometric value data meets the over-limit condition, the judgment result data is a valid alarm; otherwise, it is a false alarm and is filtered; wherein, the over-limit condition is that the absolute value of the geometric value is greater than the preset pull-out standard value or the guide height value exceeds the preset guide height range.

9. The method for filtering false alarms of contact wire thermal defect geometric parameters based on visual images as described in claim 1, characterized in that, The step of verifying the contact point data using the reference camera image data in the sorted over-limit frame data further includes: Based on the image acquisition time information, search for other camera image data at the same time to use them for auxiliary judgment.

Images