A track train part abnormality detection method based on point-to-face distance

Abstract

The application relates to a track train part abnormality detection method based on point-surface distance, and relates to the technical field of track part abnormality detection.The application comprises the following steps: a surface identification step, in which a measured special surface in a target image is identified through a surface identification strategy, and corresponding special identification features are generated; a model indexing step, in which a target part model is called from a part type database according to the special identification features and is imported into a space coordinate system; a vector fitting step, in which a presentation vector of the target part model in the space coordinate system is determined through a model presentation strategy, and a coordinate mapping relationship between the space coordinate system and the target image is constructed; a point confirmation step, in which the graphic plane coordinates of a reference special point in the target image are determined according to the mapping relationship; and a deviation calculation step, in which a difference value of the coordinates of the reference special point and the measured special point is calculated to obtain a special point deviation, and a reference distance deviation is solved through a preset distance solving algorithm.The application has the effect of facilitating detection of abnormal vector changes of track train parts.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of track component anomaly detection, in particular to a track train component anomaly detection method based on point-surface distance. BACKGROUND

[0002] Track inspection is a crucial link in the rail transit system, aiming to ensure the safety and reliability of the track. The traditional inspection method is mainly manual inspection, which is time-consuming and labor-intensive, and is easily affected by human factors.

[0003] In related technologies, in order to facilitate the inspection of track trains, a track component detection system based on YOLOv8 is introduced. The entire system includes five parts: data acquisition, data processing, model training, model optimization and evaluation, real-time detection and system deployment. Through the installation of high-definition cameras on the inspection vehicle, track image data is periodically collected, which can identify foreign objects in areas such as the track bed and switch rails, detect and locate component fractures and missing parts, and further automatically warn and alarm in the presence of abnormal conditions. This can significantly improve inspection efficiency and accuracy, and timely detect and handle track component failures and wear.

[0004] For the above related technologies, mainly using image detection technology to construct component models and identify abnormal image features of components to analyze whether abnormalities occur, but it is difficult to detect spatial vector changes of components, and further improvement is needed. SUMMARY

[0005] In order to facilitate the abnormal change of components on the track train in the vector, the application provides a track train component anomaly detection method based on point-surface distance.

[0006] The application provides a track train component anomaly detection method based on point-surface distance, which adopts the following technical scheme:

[0007] A track train component anomaly detection method based on point-surface distance, comprising:

[0008] A face recognition step, which recognizes the measured special face in the target image through a preset face recognition strategy and generates corresponding special recognition features;

[0009] A model indexing step, which retrieves a target component model from a preset component type database according to the special recognition features, and imports the target component model into a spatial coordinate system;

[0010] The vector fitting step includes a model presentation strategy, which is fitted by the measured special surface and the target part model in the spatial coordinate system to determine a presentation vector of the target part model in the spatial coordinate system, and constructs a coordinate mapping relationship between the spatial coordinate system and the target image according to the presentation vector.

[0011] The point confirmation step determines the planar coordinates of the reference special point in the target image in the target part model according to the mapping relationship.

[0012] The deviation calculation step calculates the difference between the image planar coordinates of the reference special point and the measured special point to obtain a special point deviation, and calculates a corresponding reference distance deviation according to the special point deviation through a preset distance solving algorithm.

[0013] By adopting the above technical solutions, the face recognition step accurately locates the measured special surface, the model indexing step quickly retrieves the target part model, the vector fitting step constructs the spatial coordinate mapping relationship, the point confirmation step determines the planar coordinates of the reference point, and the deviation calculation step quantifies the distance deviation, thereby effectively detecting vector abnormalities of track parts, avoiding subjective errors of manual inspection, improving detection efficiency and accuracy, and providing data support for track safety operation. It helps to reduce the risk of train operation failure.

[0014] Optionally, the face recognition strategy includes:

[0015] According to the pixel value range determined by the target image, a plurality of original inflation points at different image planar coordinates are generated, each inflation point corresponds to a different pixel value range, and a dynamic inflation algorithm is used to expand the image area from the original inflation point in the target image until the preset inflation constraint condition is met.

[0016] The image area with an area greater than a preset reference proportion area is selected as a reference face extraction area, and a plurality of face feature points are determined from the reference face extraction area according to a coordinate positioning algorithm to form the measured special surface.

[0017] By adopting the above technical solutions, the original inflation point is generated based on the pixel value range, the image area is expanded adaptively using the dynamic inflation algorithm, and the reference face is screened based on the reference proportion area, and finally the face feature points are determined, which helps to reduce the interference of uneven light and oil pollution, accurately extract the special surface of the track under the pad and the track bed surface, and significantly improve the recognition robustness of complex texture surfaces compared with the traditional threshold segmentation method. It provides a reliable two-dimensional feature basis for subsequent three-dimensional modeling, and helps to reduce the abnormality missed detection caused by feature extraction error.

[0018] Optionally, the dynamic inflation algorithm is used to calculate a similar inflation value of the adjacent graph region, and if the similar inflation value is greater than the dynamic inflation reference, the adjacent graph region is merged into the current graph region to complete the graph region expansion.

[0019] When none of the adjacent graph regions of the current graph region meets the graph region expansion condition, the screening area of the adjacent graph region is reduced to re-determine the adjacent graph region of the current graph region.

[0020] When the screening area is less than a preset reference trigger area, it is considered that the inflation constraint condition is met.

[0021] The dynamic inflation algorithm uses the following formula for calculation:

[0022]

[0023] The similar inflation value of the adjacent graph regions i and j is used to measure the pixel value distribution similarity, The pixel value vector of the kth pixel in the graph region i is represented, The pixel value vector of the corresponding position kth pixel in the graph region j is represented, and n represents the number of pixels in the feature points participating in the calculation. The Euclidean distance norm between the graph regions is used to calculate the pixel value difference.

[0024] By using the above technical scheme, the dynamic inflation algorithm determines the graph merging condition based on the similar inflation value. When the pixel value similarity of the adjacent graph region is higher than the dynamic reference, the adjacent graph region is merged into the current region. When the condition is not met, the screening area is iteratively reduced until the reference area threshold is triggered. This scheme can adaptively process rust stains, crack edge blur and other situations on the surface of the steel rail, avoid excessive merging or segmentation of regions caused by traditional fixed threshold inflation, make the extracted rail head, clamp and other component contours more consistent with the actual morphology, and help improve the geometric accuracy of subsequent point-to-surface distance calculation.

[0025] Optionally, it further includes a dynamic reference sub-step for determining the dynamic inflation reference, and the dynamic reference sub-step includes:

[0026] The feature recognition algorithm is used to determine the surface influence features in the target image, and the feature type weight of each influence feature is determined according to the type of the surface influence feature. If there is a surface influence feature of the corresponding type in the adjacent graph region, the corresponding feature type weight value is calculated according to the area of the surface influence feature. The feature type weight value is weighted to obtain the attenuation inflation factor by weighting all feature type weight values in the adjacent graph region, and the dynamic inflation reference is generated according to the attenuation inflation factor.

[0027] ​By adopting the technical scheme, the dynamic reference sub-step determines the feature weight according to the surface influence feature type such as a crack and a foreign object, calculates the weight value in combination with the feature area, generates a decay expansion factor by weighting to dynamically adjust the expansion reference, helps to adaptively adjust the detection sensitivity for different types of defects, dynamically suppresses the interference of non-key features in a complex scene such as a ballast water and a steel rail oil stain, and improves the recognition priority of typical faults such as a bolt loss and a gauge block fragmentation, and helps to reduce the false detection rate.

[0028] Optionally, the model presentation strategy is configured with a fitting reliability algorithm for calculating a comprehensive fitting value of the measured special surface corresponding to the target part model under different fitting vectors and determining a fitting vector with the highest comprehensive fitting value as the presentation vector.

[0029] By adopting the technical scheme, the fitting reliability algorithm calculates the comprehensive fitting value under different fitting vectors, selects the presentation vector corresponding to the optimal value, realizes the spatial mapping of the measured special surface and the target part model, helps to improve the matching degree of the model and the measured data, avoids the coordinate deviation caused by traditional manual calibration, especially in the detection of complex components in the turnout area, can accurately align the three-dimensional models of the switch rail and the basic rail, and provides a reliable spatial transformation matrix for subsequent point-surface distance calculation, and helps to improve the consistency of geometric parameter measurement.

[0030] Optionally, the fitting reliability algorithm comprises:

[0031] The comprehensive fitting value is calculated according to a shadow compensation abnormal value, the shadow compensation abnormal value is obtained through the deviation of a shadow vector and a presentation vector, and the shadow vector is generated by extracting a shadow area in a target image;

[0032] The fitting reliability algorithm calculates according to the following model formula:

[0033] ;

[0034] Wherein, represents a comprehensive fitting value corresponding to a fitting vector v, v represents a fitting vector to be evaluated, represents the i-th feature point coordinate of the measured special surface, represents the i-th corresponding point coordinate of the target part model under the fitting vector v, and m represents the total number of feature points, represents a preset shadow compensation coefficient for balancing the influence of the shadow, represents the shadow intensity of the region where the i-th point is located.

[0035] By adopting the technical scheme, the fitting reliable algorithm introduces the shadow compensation value shadow compensation outlier, corrects the comprehensive fitting value based on the shadow intensity, and eliminates the interference of uneven illumination on the positioning of the feature points. In the illumination mutation scene such as a tunnel or a bridge, the mis-matching of the feature points in the shadow area of the steel rail can be reduced. For example, in the detection of the slide bed platform disassembly, the fitting error of the shadow area is compensated, the mapping of the plane coordinates of the three-dimensional model and the measured image is more accurate, and the precision of the disassembly displacement detection is improved.

[0036] Optionally, the vector fitting step further comprises a macro correction sub-step:

[0037] The macro correction sub-step comprises: retrieving a historical presentation vector in a historical database according to the type of the target part model, screening the historical presentation vectors that meet preset similar conditions according to current environmental information, and dividing the historical presentation vectors into a historical presentation vector group according to a preset vector classification strategy, wherein the environmental information comprises shooting position data, shooting device data, track information data, illumination data, and visibility data.

[0038] The presentation vector is corrected according to a mean vector of the historical presentation vector group to generate a new presentation vector.

[0039] By adopting the technical scheme, the macro correction sub-step retrieves the historical presentation vector according to the type of the part, classifies the generation vector group in combination with the shooting position, illumination, and other environmental information, corrects the current presentation vector by using the mean vector, and simultaneously can reuse the optimal mapping relationship in the historical detection data. In repeated detection of the same track section, the coordinate system is quickly calibrated, which is especially suitable for periodic inspection scenes. Compared with the vector fitting from the beginning, the calculation time is reduced, and the real-time detection efficiency is improved.

[0040] Optionally, the vector classification strategy comprises:

[0041] The presentation reliable value of each historical presentation vector is calculated according to the similar frequency value, the interference outlier value, and the measured deviation value, and the historical presentation vector group is generated by taking the historical presentation vectors whose presentation reliable values are ranked within a preset range.

[0042] By adopting the technical scheme, the vector classification strategy calculates the presentation reliable value based on the similar frequency value, the interference outlier value, and the measured deviation value, screens the high-reliable historical vector to generate the group, and helps to eliminate the error mapping relationship caused by abnormal factors such as device jitter and sudden strong light. For example, in the detection in rainy weather, the noise data in the low visibility is automatically filtered, the effectiveness of the historical vector group is ensured, the corrected presentation vector is more in line with the actual track geometry, and the detection stability in the harsh environment is improved.

[0043] Optionally, the distance solving algorithm generates a corresponding deviation mapping function according to the presentation vector, and substitutes the special point deviation into the deviation mapping function to output a corresponding reference distance deviation.

[0044] The distance solving algorithm calculates by using the following model formula:

[0045]

[0046]

[0047] wherein, represents the reference distance deviation, represents a scale conversion matrix, represents an optimal presentation vector a corresponding coordinate conversion matrix, represents a special point deviation vector.

[0048] By using the above technical solution, the distance solving algorithm generates a deviation mapping function based on the optimal presentation vector, and converts the point deviation of the image plane into an actual physical distance deviation. The model integrates the camera internal participation space conversion matrix, realizes accurate conversion of pixel coordinates to millimeter units, and in the steel rail tooth gap detection, can directly output the height difference of up and down tooth gaps and the horizontal distance of left and right tooth gaps. Compared with traditional manual measurement, it meets the high-precision detection needs of the railway industry.

[0049] Optionally, it further includes a fast recognition step, and the fast recognition step includes identifying a trigger condition, and the trigger condition is that when the similarity of the target image and the target image in the historical data exceeds a preset similarity threshold, the coordinate mapping relationship of the target image in the historical data is taken as the coordinate mapping relationship of the current target image to directly enter the point confirmation step.

[0050] By using the above technical solution, the fast recognition step directly enters the point confirmation step by reusing the historical coordinate mapping relationship when the similarity of the target image and the historical image exceeds the threshold. This mechanism can skip time-consuming processes such as face recognition and vector fitting, shorten the processing time of a single image in repeated detection of the same track section, facilitate the adaptation to the moving inspection scene of a fast train, and help to improve the engineering practicability of the track inspection system.

[0051] In summary, the present application includes at least one of the following beneficial technical effects:

[0052] ​​1. The surface recognition step accurately locates measured special surfaces, the model indexing step quickly retrieves the target part model, the vector fitting step establishes spatial coordinate mapping relationships, the point confirmation step determines the plane coordinates of the reference point, and the deviation calculation step quantifies distance deviations. This enables effective automatic detection of vector anomalies in track components, avoiding subjective errors in manual inspections, improving detection efficiency and accuracy, and locating geometric anomalies such as loose bolts and pad displacement in real time. This provides data support for safe track operation and maintenance, helping to reduce the risk of train failures.

[0053] 2. Utilizing a dynamic expansion algorithm to adaptively expand the image area, the reference surface is screened based on the reference proportional area, ultimately determining surface feature points. This helps reduce interference from uneven lighting and oil coverage, accurately extracting special surfaces such as the rail pad and roadbed surface, and reducing missed detections due to feature extraction errors.

[0054] 3. The dynamic benchmark sub-step determines the feature weight through the surface-affecting feature type, such as cracks and foreign matter, calculates the weight value based on the feature area, and generates a weighted attenuation expansion factor to dynamically adjust the expansion benchmark. This helps to adaptively adjust the detection sensitivity for different types of defects. In complex scenarios such as water accumulation on the roadbed and oil pollution on the rails, it dynamically suppresses the interference of non-critical features, improves the recognition priority of typical faults such as missing bolts and broken gauge blocks, and helps to reduce the false detection rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a method flow chart of steps S100 to S500 in this application.

[0056] Figure 2 It is a method flow chart of steps S101 to S102 in this application.

[0057] Figure 3 It is a method flow chart of steps S1011 to S1013 in this application.

[0058] Figure 4 It is a method flow chart of steps S1014 to S1015 in this application.

[0059] Figure 5 This is a method step diagram of steps S302 to S3022 in this application. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solutions and advantages of this application more clear, the following Figures 1-5 It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application.

[0061] The embodiments of the present invention are described in further detail below with reference to the accompanying drawings.

[0062] The embodiment of the application discloses a track train part abnormality detection method based on point-surface distance, accurately positions the measured special surface through the surface recognition step, quickly calls the target part model through the model index step, constructs the spatial coordinate mapping relationship through the vector fitting step, determines the reference point plane coordinate through the point confirmation step, quantifies the distance deviation through the deviation calculation step, so that the vector abnormality of the track part can be effectively detected automatically, the subjective error of artificial inspection is avoided, the detection efficiency and accuracy are improved, the geometric abnormality such as bolt loosening and pad displacement can be positioned in real time, data support is provided for track safety operation, and train operation fault risk control is realized.

[0063] With reference to Figure 1 The method flow of the track train part abnormality detection method based on point-surface distance comprises the following steps:

[0064] The surface recognition step S100 comprises the following steps:

[0065] In track image detection, the surface recognition strategy is the basis for accurately extracting part features. In this step, the original inflation points of different image plane coordinates are generated by presetting the pixel value range, each inflation point corresponds to an independent pixel value interval, and the dynamic inflation algorithm is relied on to expand the image area from the original point as the starting point until the inflation constraint condition is met. Specifically, whether to merge the areas is judged by calculating the similar inflation value of adjacent image areas.

[0066] When the similar inflation value is greater than the dynamic inflation reference, the image areas are merged, otherwise the area is iteratively reduced and screened until the expansion is stopped when the area is less than the reference trigger area. For example, in the steel rail surface detection, this strategy can effectively exclude interference such as oil stains and rust marks, and accurately extract the measured special surfaces such as rail heads and pads. The specific calculation method of the similar inflation value will be further described in the subsequent steps.

[0067] The model index step S200 comprises the following steps:

[0068] Based on the special identification features generated by face recognition, this step retrieves the target part model from the preset part type database and imports it into the spatial coordinate system. The database uses a feature vector indexing mechanism to establish a mapping relationship between the three-dimensional models of components such as bolts and pads and geometric features (such as surface curvature and edge gradient). In specific operations, the similarity between the measured features and the database model is calculated through a feature matching algorithm, and the model with a matching degree greater than 0.8 is selected as the target part model. For example, when the polygonal contour feature of the gauge block is detected, the system automatically retrieves the three-dimensional model of the gauge block of the corresponding model and establishes a spatial coordinate system with the model's center of mass as the origin, providing a benchmark for subsequent three-dimensional fitting.

[0069] Vector fitting step S300: includes a model rendering strategy, which determines the rendering vector of the target part model in the spatial coordinate system by fitting the measured special surface and the target part model in the spatial coordinate system, and constructs a coordinate mapping relationship between the spatial coordinate system and the target image based on the rendering vector;

[0070] This step uses a model rendering strategy to achieve spatial fitting between the measured special surface and the target part model. A reliable fitting algorithm is used to calculate the comprehensive fitting value for different fitting vectors. The vector with the best comprehensive fitting value is selected as the rendering vector, and a coordinate mapping relationship between the spatial coordinate system and the image is constructed.

[0071] The algorithm incorporates a shadow compensation mechanism, dynamically adjusting fitting parameters based on shadow areas within the image to mitigate the effects of uneven lighting. Furthermore, the macro-correction substep uses a historical database of vectors representing similar parts and uses the mean vector to correct the current vector, further improving fitting accuracy. The specific calculation model for the comprehensive fitting value will be detailed in subsequent steps.

[0072] Point confirmation step S400: determining the graphic plane coordinates of the reference special point in the target image in the target part model according to the mapping relationship;

[0073] Based on the coordinate mapping relationship established in the vector fitting step, this step accurately locates the graphic plane coordinates of the datum special points in the target image in the target part model. In specific implementation, a bilinear interpolation algorithm is used to perform coordinate transformation in the mapping relationship grid, and the coordinates of the datum special points in the spatial coordinate system are mapped to the image plane. For complex curved surface parts, a curvature compensation factor is introduced to correct the interpolation results to eliminate positioning errors caused by surface projection deformation. For example, when inspecting high-speed rail fasteners, the system accurately projects the datum special points such as the center points of the bolt holes on the fastener model to the image plane through a mapping relationship, providing an accurate reference for subsequent deviation calculations. Among them, the calculation method of the curvature compensation factor will be further explained in the deviation calculation step.

[0074] Step S500: Calculate the difference between the image plane coordinates of the reference special point and the measured special point to obtain the special point deviation, and use the preset distance solving algorithm to obtain the corresponding reference distance deviation according to the special point deviation.

[0075] In this step, the reference distance deviation is accurately obtained by calculating the difference between the image plane coordinates of the reference special point and the measured special point, and combining the preset distance solving algorithm. First, the weighted Euclidean distance formula is used to calculate the special point deviation, and the weight coefficient is dynamically adjusted according to the importance of the point in the structure of the part. For example, the weight coefficient of the center point of the bolt hole in the rail fastener is higher than that of the edge point, so as to highlight the detection accuracy of the key part. Then, the special point deviation in the image plane is converted into the reference distance deviation in the actual space through a spatial similarity transformation model, which considers the camera distortion parameters and the perspective projection transformation relationship. In order to improve the calculation robustness, the RANSAC algorithm is introduced to remove abnormal points, ensuring the reliability of the deviation calculation result. Specifically, when detecting the installation deviation of the gauge block, the system first calculates the coordinate difference between the reference point and the measured point in the image plane, and then converts it into the actual installation deviation value through the distance solving algorithm, providing a quantitative basis for track state evaluation. The specific parameters of the weighted Euclidean distance formula and the spatial similarity transformation model will be described in detail in the system implementation part.

[0076] With reference to Figure 2 , the face recognition strategy includes:

[0077] Step S101: Determine the pixel value range of the target image to generate a plurality of original inflation points at different image plane coordinates, each inflation point corresponding to a different pixel value range, and perform region expansion on the target image based on the original inflation point as the initial point until the preset inflation constraint condition is met through the preset dynamic inflation algorithm;

[0078] In this step, based on the pixel value distribution characteristics of the target image, original inflation points at different image plane coordinates are generated, each inflation point corresponding to an independent pixel value range (such as a specific interval of the RGB channel). The target image is expanded by region through the preset dynamic inflation algorithm, taking the original inflation point as the starting point. Specifically, the similar inflation value of the adjacent region is calculated, and when the value is greater than the dynamic inflation reference, the adjacent region is merged into the current region. If all adjacent regions do not meet the condition, the screening area is iteratively reduced until the screening area is less than the preset reference trigger area, which is considered to meet the inflation constraint condition. For example, when detecting the surface of the rail head, this algorithm can adaptively exclude the abnormal pixel value area formed by oil stains and accurately expand to the actual contour boundary of the rail head. The specific calculation method of the similar inflation value and the dynamic inflation reference will be further described in the subsequent steps.

[0079] Step S102: Screen the graph area with area greater than the preset reference proportion area as the reference surface extraction graph area, and determine several surface feature points from the reference surface extraction graph area according to the coordinate positioning algorithm to form the measured special surface.

[0080] After the graph area expansion is completed, the graph area with area greater than the preset reference proportion area is screened as the reference surface extraction graph area in this step to filter the noise area (such as the small rust spots on the surface of the steel rail). The reference surface extraction graph area is processed by the coordinate positioning algorithm, and several surface feature points are determined based on edge detection, corner recognition and other technologies, and these points form the measured special surface (such as the upper surface profile of the rail pad) according to the spatial distribution rule. For example, for the ballast water area, the system will preferentially extract the graph area with area exceeding 100 cm², and determine the feature points of the water edge through the sub-pixel level positioning algorithm to ensure the geometric accuracy of the measured special surface. The specific implementation method of the coordinate positioning algorithm will be expanded in the subsequent steps.

[0081] Reference Figure 3 , the dynamic expansion algorithm in step S101 also includes;

[0082] Step S1011: If the similar expansion value is greater than the dynamic expansion reference, the adjacent graph area is merged into the current graph area to complete the graph area expansion;

[0083] In this step, whether to perform the graph area merging operation is determined by comparing the similar expansion value of the adjacent graph area with the dynamic expansion reference. If the similar expansion value is greater than the dynamic expansion reference, it means that the pixel value distribution similarity of the adjacent graph area is high, which belongs to the same component feature area, and it is merged into the current graph area to complete the expansion; if the similar expansion value is not greater than the dynamic expansion reference (i.e. less than or equal to), it is determined that the adjacent graph area and the current area have large feature difference, and no merging is performed, and the other adjacent graph area is continued to be checked. For example, in the steel rail crack detection, when the similar expansion value of the crack edge graph area and the rail head main body is less than the reference, the system will not mistakenly merge the crack area into the rail head main body, ensuring the accuracy of the crack boundary.

[0084] Step S1012: When the adjacent graph area of the current graph area does not meet the graph area expansion condition, the screening area of the adjacent graph area is reduced to re-determine the adjacent graph area of the current graph area;

[0085] When all the adjacent graph regions of the current graph region do not meet the merging condition, the system starts the screening area decay mechanism, i.e., reducing the screening area threshold of the adjacent graph region by a preset proportion (e.g., 20% each time), and re-determining the range of the adjacent graph region meeting the condition. The adjustment mechanism can avoid missing detection of the feature region due to the too large initial screening area, for example, in the scenario of oil stain covering the surface of the steel rail, by gradually reducing the screening area, the rail head profile edge blocked by the oil stain can be accurately positioned. If there is an adjacent graph region meeting the condition after adjustment, return to step S1011 to continue merging; if there is still no region meeting the condition, proceed to the next step.

[0086] Step S1013: When the screening area is less than the preset reference trigger area, it is considered that the inflation constraint condition is met.

[0087] The system monitors the screening area in real time, and when the screening area is less than the preset reference trigger area (e.g., 10 pixels), it is considered that the image details have been sufficiently traversed, the inflation constraint condition is met, and the graph region expansion is stopped; if the screening area is still greater than or equal to the reference trigger area, the iterative adjustment of step S1012 is continued until the termination condition is triggered. This mechanism can balance the detection accuracy and computational efficiency, for example, in the detection of foreign objects on the track bed, when the screening area is reduced to less than the minimum detection threshold of the foreign object (e.g., 10 cm2 corresponding to the pixel area), the system automatically terminates the expansion to avoid invalid calculation. The specific calculation model of the similarity inflation value and the dynamic inflation reference will be described in detail in the subsequent steps.

[0088] Further, in the embodiment, the dynamic inflation algorithm is calculated by the following formula:

[0089] ;

[0090] represents the similarity inflation value of the adjacent graph regions i and j, which is used to measure the similarity of pixel value distribution, represents the pixel value vector of the kth pixel in the graph region i, represents the pixel value vector of the corresponding kth pixel in the graph region j, and n represents the number of pixels in the feature points participating in the calculation, represents the Euclidean distance norm between the graph regions, which is used to calculate the pixel value difference.

[0091] Referring to Figure 4 It also includes a dynamic reference sub-step for determining a dynamic inflation reference, and the dynamic reference sub-step includes:

[0092] Step S1014: The feature recognition algorithm is used to determine the surface impact features in the target image, and the feature type weight of each impact feature is determined according to the type of the surface impact feature. If there is a surface impact feature of the corresponding type in the adjacent image area, the corresponding feature type weight value is calculated according to the area of the surface impact feature, the feature type weight values of all features in the adjacent image area are weighted by the feature type weight value to obtain the attenuation inflation factor, and the dynamic inflation benchmark is generated according to the attenuation inflation factor.

[0093] In this step, the feature recognition algorithm (such as HOG+SVM) is used to scan the target image, locate the surface impact features (such as cracks, foreign matter, and oil stains), and assign an initial feature type weight according to the feature type (crack weight> foreign matter weight> oil stain weight). If there is a surface impact feature of the corresponding type in the adjacent image area, the specific weight value is calculated based on the feature area, and the larger the area, the higher the weight value; if there is no corresponding feature, the initial weight value is used. The weight values of all feature types in the adjacent image area are summed by weighting to obtain the attenuation inflation factor (for example, when the crack feature area accounts for 60% and the foreign matter accounts for 40%, the factor = 0.6 x crack weight + 0.4 x foreign matter weight). The dynamic inflation benchmark is dynamically adjusted according to the attenuation inflation factor: the larger the factor, the smaller the benchmark value, to enhance the detection sensitivity of key features; on the contrary, the smaller the factor, the larger the benchmark value, to suppress noise interference. This mechanism can adaptively balance the detection accuracy in different scenarios, for example, in ballast crack detection, the weight proportion of crack features is automatically increased to ensure that small cracks are effectively identified.

[0094] In addition, step S1015: the model presentation strategy is configured with a fitting reliability algorithm, which is used to calculate the comprehensive fitting value of the measured special surface under different fitting vectors corresponding to the target part model, and determine the fitting vector with the highest comprehensive fitting value as the presentation vector.

[0095] The fitting reliability algorithm calculates the comprehensive fitting value under different fitting vectors, selects the presentation vector corresponding to the optimal value, realizes the spatial mapping of the measured special surface and the target part model, helps to improve the matching degree of the model and the measured data, avoids the coordinate deviation caused by traditional manual calibration, especially in the detection of complex components in the turnout area, accurately aligns the three-dimensional model of the switch rail and the basic rail, provides a reliable spatial transformation matrix for subsequent point-surface distance calculation, and helps to improve the consistency of geometric parameter measurement. The specific fitting reliability algorithm is described in the subsequent steps.

[0096] Reference Figure 5 The fitting reliability algorithm includes:

[0097] Step S301: obtaining the comprehensive fitting value according to the shadow compensation outlier value, wherein the shadow compensation outlier value is obtained by the deviation between the shadow vector and the presentation vector, and the shadow vector is generated by extracting the shadow area in the target image;

[0098] This step extracts shadow regions from the target image to generate shadow vectors. A comprehensive fitting value is then calculated based on the shadow compensation value and shadow compensation outliers to eliminate the interference of uneven illumination on model fitting. Specifically, a brightness threshold-based image segmentation algorithm (such as the Otsu threshold method) is first used to identify shadow regions. Pixels with brightness values ​​below 60% of the global mean brightness are considered shadow pixels, while pixels with brightness values ​​below 60% are considered non-shadow pixels. After smoothing the shadow boundaries using a Gaussian filter, a shadow vector is generated containing the shadow position and intensity (e.g., a mask matrix and intensity gradient matrix for the shadow region).

[0099] When calculating the comprehensive fit value, the algorithm applies shadow compensation to the matching error between the measured feature points of the special surface and the corresponding points on the target part model. For feature points located in the shadow area (shadow intensity > 0.5), the matching error is weighted according to the shadow intensity ratio (for example, the error is magnified by 1.2 times when the intensity is 0.8). For feature points in the non-shadow area (shadow intensity ≤ 0.5), the standard error calculation method is used. For example, when inspecting rails at a tunnel entrance, direct sunlight causes half of the rail head to be in shadow. This compensation mechanism can reduce the fitting error of feature points in the shadow area. The specific quantitative model for shadow compensation will be explained in detail in the following steps.

[0100] The fitting reliability algorithm is calculated using the following model formula:

[0101] ;

[0102] in, Represents the comprehensive fitting value corresponding to the fitting vector v, v represents the fitting vector to be evaluated, Represents the coordinates of the i-th feature point of the measured special surface, represents the coordinates of the i-th corresponding point of the target part model under the fitting vector v, m represents the total number of surface feature points, Indicates the preset shadow compensation coefficient, which is used to balance the impact of shadows. Indicates the shadow intensity of the area where the i-th point is located.

[0103] The vector fitting step also includes a macro correction sub-step S302:

[0104] Step S3021: According to the type of the target part model, the historical rendering vector in the historical database is called, and the historical rendering vector is divided into a historical rendering vector group according to a preset vector classification strategy. The environmental information includes shooting position data, shooting device data, track information data, illuminance data and visibility data.

[0105] In this step, the historical rendering vector of the same type of part is called from the historical database according to the type of the target part model (such as a bolt, a pad, a turnout frog, etc.). The corresponding environmental information is obtained synchronously when calling, including shooting position data (such as milepost number), shooting device data (such as camera model, focal length), track information data (such as curve radius, track gauge), illuminance data (such as lux value) and visibility data (such as haze level). According to the preset vector classification strategy, the historical rendering vector is divided into different historical rendering vector groups according to the environmental information similarity.

[0106] For example, the historical vector with illuminance of 500-1000 lux and visibility >100 m is classified into the "good light group". If the environmental information of a certain historical vector is different from the current detection scene by more than a preset threshold (such as illuminance deviation >30%), it is excluded from the classification, ensuring the environmental matching degree of the vector group.

[0107] Step S3022: The rendering vector is corrected according to the mean vector of the historical rendering vector group to generate a new rendering vector.

[0108] For each historical rendering vector group, the mean vector (the arithmetic mean of each vector parameter) is calculated, and the current rendering vector is corrected based on the mean vector. The specific correction logic is: if the parameter deviation between the current vector and the mean vector is >15% (such as the rotation angle difference is >5°), adjust the current vector parameter by 50% of the deviation; if the deviation is ≤15%, directly generate a new rendering vector by weighted average of the mean vector and the current vector (weight ratio is 3:1). For example, when detecting the rail at the entrance of the tunnel, the historical vector group with the same light condition (illuminance 200-300 lux) is called, and the mean vector can correct the fitting deviation caused by the sudden change of light, so that the coordinate mapping error of the new rendering vector is reduced.

[0109] The vector classification strategy includes:

[0110] Step S30211: According to the similar frequency value, the interference abnormal value and the measured deviation value, the rendering reliability value of each historical rendering vector is calculated, and the historical rendering vector group is generated by taking the historical rendering vector with the rendering reliability value within the preset range.

[0111] In this step, the rendering reliability value of the historical rendering vector is calculated by multi-dimensional index to screen the effective vector to generate the historical group. Specifically:

[0112] Similar frequency value: Count the number of times a historical vector is used under the same part type and environmental conditions. Vectors with a frequency of ≥10 times will be given a basic reliability score of +30 points, otherwise the score will be calculated based on the frequency ratio;

[0113] Interference anomaly value: Checks whether there are any abnormal records such as device jitter and strong light interference when the historical vector is generated. If there is no abnormality, 0 points will be scored, and 10 points will be deducted for each type of abnormality.

[0114] Measured deviation value: Compare the deviation between the actual test result corresponding to the historical vector and the manual re-test result. If the deviation is ≤0.5mm, +40 points will be given; if it is 0.5-1mm, +20 points will be given; and if it is >1mm, 0 points will be given.

[0115] The three scores are summed to obtain the presentation reliability value. The historical vectors ranked in the top 30% are selected to form the historical presentation vector group. The remaining vectors are marked as "low reliability" and excluded. For example, a historical vector used 15 times in similar scenarios with no anomalies and a measured deviation of 0.3mm has a reliability value of 30 + 0 + 40 = 70 points. If the preset threshold for the top 30% is 65 points, the vector will be selected for inclusion in the vector group, ensuring that the correction process is based on highly reliable historical data. The specific quantitative model for presentation reliability will be detailed in subsequent steps.

[0116] The deviation calculation step S500 further includes:

[0117] Step S501: The distance solving algorithm generates a corresponding deviation mapping function according to the presentation vector, and substitutes the special point deviation into the deviation mapping function to output the corresponding reference distance deviation;

[0118] This step generates a deviation mapping function for converting pixel deviations to physical distances based on the optimal rendering vector obtained through vector fitting. Specifically, the algorithm first integrates the spatial transformation parameters (rotation matrix, translation vector) in the rendering vector and combines them with the camera's intrinsic parameter matrix (focal length, principal point coordinates) and the physical pixel ratio (e.g., 0.5 mm / pixel) to construct a coordinate transformation model from the image plane to three-dimensional space. This model uses an interpolation algorithm to generate a continuous deviation mapping function, which directly maps the deviation of a specific point (pixel coordinate difference in the image plane) to the actual physical distance deviation.

[0119] When inputting a specific point deviation vector (such as the coordinate offset of a bolt vertex in an image), the deviation mapping function first performs camera distortion correction on the pixel deviation. Then, using a 3D spatial transformation matrix, the corrected deviation is converted to millimeter-level distance deviation. For example, when inspecting rail misalignment, if the vertical pixel deviation between the reference point in the image and the measured point is 4 pixels, the deviation mapping function can output an actual misalignment height of 2mm (based on a physical scale of 0.5mm / pixel). The specific mathematical model and parameter calibration method of this function will be explained in detail in the following steps.

[0120] The distance solving algorithm uses the following model formula for calculation:

[0121]

[0122]

[0123] wherein, d represents the reference distance deviation, represents the scale conversion matrix, represents the optimal presentation vector corresponding coordinate conversion matrix, represents the special point deviation vector.

[0124] In addition, in addition to step S500, a fast identification step S600 is further included, and the following step scheme is adopted:

[0125] The trigger condition is identified, and when the similarity of the target image and the target image in the historical data exceeds the preset similarity threshold, the coordinate mapping relationship of the target image in the historical data is taken as the coordinate mapping relationship of the current target image to directly enter the point confirmation step.

[0126] When the trigger condition is met, the system retrieves the coordinate mapping relationship (including the presentation vector, the scale conversion matrix and the like) corresponding to the historical image with the highest similarity to the current image from the historical database, directly as the coordinate mapping relationship of the current target image, skips the face recognition and vector fitting steps, and directly enters the point confirmation step. This mechanism can greatly shorten the detection time, for example, the conventional process of a single image processing needs 200 ms, and after reuse of the historical mapping, only 50 ms is needed, which is suitable for high-speed train inspection scenes with a speed of 200 km / h or more. If there are multiple coordinate mapping relationships of high-similarity images in the historical database, the parameter group with the most matching times and the smallest measured deviation is selected as the reuse reference to ensure the mapping accuracy.

[0127] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above functional modules is taken as an example for illustration, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be repeated here.

[0128] The embodiment of the application provides a computer readable storage medium, which stores a computer program capable of being loaded and executed by a processor to perform a track train part abnormality detection method based on point-plane distance.

[0129] ​​​The computer storage medium includes, for example, a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media capable of storing program codes.

[0130] Based on the same inventive concept, the embodiment of the present application provides a kind of intelligent terminal, including memory and processor, memory is stored with the computer program of the track train component abnormality detection method based on point surface distance capable of being loaded and executed by processor.

[0131] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of each functional module is exemplified, and in actual application, the above-mentioned functions can be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. The specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0132] The above are preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. Any feature disclosed in the specification (including the abstract and drawings) can be replaced by other equivalent or similar features unless specifically described. That is, each feature is only an example of a series of equivalent or similar features unless specifically described.

Claims

1. A point-to-face distance-based rail train component anomaly detection method, characterized in that, The method comprises the following steps: a face recognition step, which recognizes a measured special face in a target image and generates a corresponding special recognition feature through a preset face recognition strategy; a model indexing step, which retrieves a target part model from a preset part type database according to the special recognition feature, and imports the target part model into a spatial coordinate system; a vector fitting step, which comprises a model presentation strategy, the model presentation strategy fitting the measured special face and the target part model in the spatial coordinate system to determine a presentation vector of the target part model in the spatial coordinate system, and constructing a coordinate mapping relationship between the spatial coordinate system and the target image according to the presentation vector; the model presentation strategy is configured with a fitting reliability algorithm, which is used to calculate a comprehensive fitting value of the measured special face corresponding to the target part model under different fitting vectors, and determine the fitting vector with the highest comprehensive fitting value as the presentation vector; the fitting reliability algorithm comprises: calculating the comprehensive fitting value according to a shadow compensation abnormal value, the shadow compensation abnormal value is obtained by the deviation of a shadow vector and the presentation vector, and the shadow vector is generated by extracting a shadow area in the target image; the fitting reliability algorithm uses the following model formula for calculation: ； wherein, represents the comprehensive fitting value corresponding to the fitting vector v, v represents the fitting vector to be evaluated, represents the i-th surface feature point coordinate of the measured special surface, represents the i-th corresponding point coordinate of the target part model under the fitting vector v, m represents the total number of surface feature points, represents a preset shadow compensation coefficient, used to balance the influence of the shadow, represents the shadow intensity of the region where the i-th point is located; a point confirmation step, which determines the planar coordinates of a reference special point in the target image in the target part model according to the mapping relationship; a deviation calculation step, which calculates the difference between the planar coordinates of the reference special point and the measured special point to obtain a special point deviation, and calculates a corresponding reference distance deviation according to the special point deviation through a preset distance solving algorithm.

2. The track train component abnormality detection method based on point-to-face distance according to claim 1, characterized in that, The face recognition strategy comprises: determining a pixel value range in the target image to generate a plurality of original inflation points at different image plane coordinates, each inflation point corresponding to a different pixel value range, and performing region expansion in the target image with the original inflation point as the initial point through a preset dynamic inflation algorithm until the preset inflation constraint condition is met; screening a region with an area greater than a preset reference proportion area as a reference face extraction region, and determining a plurality of face feature points from the reference face extraction region to form the measured special face according to a coordinate positioning algorithm.

3. The track train component anomaly detection method based on point-to-face distance according to claim 2, characterized in that, The dynamic inflation algorithm is used to calculate a similar inflation value of adjacent regions, if the similar inflation value is greater than a dynamic inflation reference, the adjacent region is merged into the current region to complete the region expansion; when none of the adjacent regions of the current region meets the region expansion condition, the screening area of the adjacent regions is reduced to re-determine the adjacent regions of the current region; when the screening area is less than a preset reference trigger area, it is considered that the inflation constraint condition is met; The dynamic inflation algorithm uses the following formula for calculation: ； Sij represents the similarity expansion value between adjacent map regions i and j, which is used to measure the similarity of pixel value distribution, Pi,k represents the pixel value vector of the kth pixel in map region i, Pj,k represents the pixel value vector of the kth pixel in map region j, n represents the number of pixels in the feature points participating in the calculation, Dij represents the Euclidean distance norm between map regions, which is used to calculate the pixel value difference.

4. The track train component anomaly detection method based on point-to-face distance according to claim 3, characterized in that, It also comprises a dynamic reference sub-step for determining the dynamic inflation reference, the dynamic reference sub-step comprises: The feature recognition algorithm is used to determine the face influence features in the target image, and the feature type weight of each influence feature is determined according to the type of the face influence feature. If there is a face influence feature of the corresponding type in the adjacent graph area, the corresponding feature type weight value is calculated according to the area of the face influence feature. The feature type weight value is weighted to obtain the attenuation expansion factor. The dynamic expansion reference is generated according to the attenuation expansion factor.

5. The track train component abnormality detection method based on point-to-face distance according to claim 1, characterized in that, The vector fitting step further comprises a macro correction sub-step: The macro correction sub-step comprises retrieving a historical presentation vector in a historical database according to the type of the target part model, screening the historical presentation vector that meets the preset similarity condition according to the current environmental information, and dividing the historical presentation vector into a historical presentation vector group according to a preset vector classification strategy. The environmental information includes shooting position data, shooting device data, track information data, illuminance data and visibility data. The presentation vector is corrected according to the mean vector of the historical presentation vector group to generate a new presentation vector.

6. The track train component anomaly detection method based on point-to-face distance according to claim 5, wherein, The vector classification strategy comprises: The presentation reliability value of each historical presentation vector is calculated according to the similarity frequency value, the interference abnormal value and the measured deviation value, and the historical presentation vector group is generated by taking the historical presentation vector whose presentation reliability value is within the preset range.

7. The track train component anomaly detection method based on point-to-face distance according to claim 1, wherein, The distance solving algorithm generates a corresponding deviation mapping function according to the presentation vector, and substitutes the special point deviation into the deviation mapping function to output the corresponding reference distance deviation. The distance solving algorithm uses the following model formula for calculation: ； ； wherein, denotes a reference distance deviation, denotes a scale conversion matrix, denotes an optimal presentation vector a corresponding coordinate conversion matrix, denotes a special point deviation vector.

8. The track train component anomaly detection method based on point-to-face distance according to claim 1, wherein, It also includes a fast recognition step, which comprises identifying a trigger condition. The identification trigger condition is that when the similarity of the target image and the target image in the historical data exceeds the preset similarity threshold, the coordinate mapping relationship of the target image in the historical data is taken as the coordinate mapping relationship of the current target image, so as to directly enter the point confirmation step.

Images