Method and device for detecting rail surface block drop of turnout rail

Abstract

The application provides a turnout rail piece rail surface block detection method and equipment, the method comprises a turnout rail piece identification step, a two-dimensional image extraction step, a rail surface block identification step, a block effective point cloud generation step, and a block feature size calculation step. Since the profile characteristics of the turnout rail piece are identified in the profile data, the starting point of the turnout rail piece is identified, and the corresponding two-dimensional image is extracted based on the starting point and the predetermined extension length to identify the block, so that the identified rail surface block can correspond to the corresponding type of turnout rail piece. Since the camera image and the track profile data collected by the line laser sensor are combined, not only can the neural network efficiently identify the block from the image, but also the corresponding profile data can be further extracted to generate the effective point cloud, which is used to calculate the depth and other feature sizes of the block, so that more comprehensive detection data of the rail surface block can be obtained, the depth and other sizes of the block do not need to be manually supplemented, and the method can be applied to actual field detection to improve the detection efficiency.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of turnout rail detection, and particularly relates to a turnout rail surface spalling detection method and device. BACKGROUND

[0002] Spalling of the rail surface of the rail is a common damage, and when the rail surface spalling is relatively serious, it will affect the passenger's ride experience and even the safety of train operation, and the rail surface spalling may also indicate further problems, such as problems of the material of the rail itself, small curve radius of the track line, high frequency of train operation, deviation of track geometry, etc., therefore, it is necessary to detect and further analyze the rail surface spalling.

[0003] There are various rails in the turnout, including the center rail, wing rail, guard rail, etc., and the rail surface spalling of these rails needs to be detected. Due to the complex structure of the turnout and the multiple types of rails, in the actual field measurement, the rail surface spalling is mainly observed manually, and then the corresponding measuring ruler is used to measure the specific size of the spalling and record it. Such detection method is low in efficiency, and due to the random distribution, different sizes and discontinuous distribution of the rail surface spalling, the manual detection method is easy to miss. Although there are some automatic detection methods for the rail surface spalling in the prior art, these methods are difficult to be used in the field detection of the turnout: on the one hand, in the detection of the turnout, not only the rail surface spalling needs to be identified, but also the identified rail surface spalling needs to be positioned to the corresponding rail, and the existing method does not provide such a solution; on the other hand, the existing method mainly uses image algorithm or neural network to identify the rail surface spalling from the camera image, and such method cannot extract the depth of the spalling, and the depth of the spalling is a very important index, therefore, in the actual field detection, even if the existing method is used, manual measurement of the size data such as the depth of the spalling is still needed, and the efficiency of the field detection is also limited.

[0004] Therefore, in order to improve the efficiency and coverage of the detection of the rail surface spalling in the detection of the turnout, a new detection method for the rail surface spalling of the turnout rail is needed. SUMMARY

[0005] The present application is carried out to solve the above problems, and aims to provide a detection method and device which can automatically position to the target turnout rail, detect the rail surface spalling of the target turnout rail and obtain the corresponding characteristic size of the spalling, and the present application adopts the following technical scheme:

[0006] The application provides a turnout rail piece rail surface block detection method for detecting the rail surface block of a turnout rail piece, characterized in that the detection is based on two-dimensional images collected by a camera, rail profile data collected by a line laser sensor, and corresponding mileage information. The method comprises: a turnout rail piece identification step of identifying the starting point of the turnout rail piece from the two-dimensional images or multiple frames of profile data based on the profile features of the starting point of the turnout rail piece, and obtaining the corresponding mileage information; a two-dimensional image extraction step of extracting a plurality of frames of two-dimensional images corresponding to the turnout rail piece as detected images based on the mileage information of the starting point of the turnout rail piece and the predetermined extension length of the turnout rail piece; a rail surface block identification step of sequentially inputting the detected images into a rail surface block identification model, which outputs the block position information of the rail surface block, wherein the rail surface block identification model is a trained neural network model; a block effective point cloud generation step of extracting multiple frames corresponding to the rail surface block from multiple frames of profile data based on the block position information and the calibration parameters of the camera and the line laser sensor, and stacking them to generate a block effective point cloud; and a block feature size calculation step of calculating the feature size of the rail surface block based on the block position information and the block effective point cloud.

[0007] The turnout rail piece rail surface block detection method provided by the application can also have the following technical features: the turnout rail piece comprises a center rail and wing rails arranged on both sides of the center rail, the turnout rail piece identification step comprises: based on the mileage information, profile matching identification of the cross section starting point of the center rail is performed on the collected profile data frame by frame to identify the cross section starting point of the center rail and obtain the corresponding mileage information; based on the mileage information of the cross section starting point of the center rail, a predetermined number of frames of profile data connected after the cross section starting point of the center rail are obtained and stacked to generate a point cloud; based on the point cloud, the center rail theoretical starting point is identified by fitting a plane, and the corresponding mileage information is obtained; based on the mileage information of the center rail theoretical starting point, profile data within a predetermined position range before the center rail theoretical starting point is extracted; profile matching identification of the wing rail starting point is performed on the extracted profile data frame by frame to identify the wing rail starting point and obtain the corresponding mileage information; and in the two-dimensional image extraction step, based on the mileage information of the center rail theoretical starting point, the extension length of the center rail, the mileage information of the wing rail starting point, and the extension length of the wing rail, a plurality of frames of two-dimensional images covering the center rail and the wing rails are extracted.

[0008] The turnout rail piece rail surface block detection method provided by the application can also have the following technical features: the heart rail theoretical starting point is identified by fitting a plane, including: fitting a plane between the heart rail section starting point and the heart rail theoretical starting point based on the point cloud by using the RANSAC algorithm; fitting an inclined plane after the heart rail theoretical starting point based on the point cloud by using the RANSAC algorithm; calculating the coordinates of the points on the intersection line of the fitted plane and the fitted inclined plane, and obtaining the corresponding mileage information based on the coordinates, or respectively calculating the center lines of the fitted plane and the fitted inclined plane, and calculating the coordinates of the intersection point of the two center lines, and obtaining the corresponding mileage information based on the coordinates.

[0009] The turnout rail piece rail surface block detection method provided by the application can also have the following technical features: the profile matching of the heart rail section starting point includes: determining whether there is a profile at a predetermined position in the profile data based on a predetermined heart rail section starting point matching template, and when it is determined that there is a profile, the profile is intercepted according to a predetermined profile interception algorithm to obtain the profile of the upper surface of the heart rail, and it is determined whether the intercepted profile is a horizontal straight line segment with a predetermined length, and when it is determined that it is, the heart rail section starting point is identified, and the mileage information corresponding to the profile data of this frame is taken as the mileage information of the heart rail section starting point; the profile matching of the wing rail starting point includes: determining whether there is a profile at a predetermined position in the profile data based on a predetermined wing rail starting point matching template, and when it is determined that there is a profile, the profile is intercepted according to a predetermined profile interception algorithm to obtain the profile of the rail head part of the two side wing rails, the intercepted profile is matched with a predetermined standard wing rail starting point profile, the similarity is calculated, and when the similarity reaches a predetermined similarity threshold, it is determined that the wing rail starting point is identified, and the mileage information corresponding to the profile data of this frame is taken as the mileage information of the wing rail starting point.

[0010] The turnout rail piece rail surface block detection method provided by the application can also have the following technical features: the track surface block area identification model is an improved YOLOv5 neural network model, which includes: a Backbone module for extracting features of the two-dimensional image, an up-sampling layer added after the Backbone module; a PAN+Bi-FPN module for fusing the extracted features to obtain fused features; and a Head module for identifying block areas and rail surface areas in the two-dimensional image based on the fused features, and outputting pixel coordinates of corresponding block anchor frames and their first confidence, and pixel coordinates of rail surface anchor frames and their second confidence.

[0011] The track turnout rail surface block detection method provided by the application can also have the following technical features: the track surface block identification step further comprises: removing the block anchor frame whose first confidence is lower than the preset confidence threshold and the block anchor frame located outside the track surface anchor frame from the identified block anchor frame.

[0012] The track turnout rail surface block detection method provided by the application can also have the following technical features: the block effective point cloud generation step comprises: based on the mileage information corresponding to the to-be-detected image, extracting a plurality of corresponding frames from the profile data and stacking them to generate a rail point cloud; based on the distance between the line laser sensor and the rail surface of the turnout rail, the rail point cloud is cropped to obtain a rail surface point cloud; based on the camera intrinsic parameter of the camera, the pixel coordinates of the block anchor frame are converted into camera coordinates, and based on the calibration relationship between the camera and the line laser sensor, the camera coordinates of the block anchor frame are converted into world coordinates; based on the world coordinates of the block anchor frame, a block effective point cloud corresponding to the track surface block is extracted from the rail surface point cloud.

[0013] The track turnout rail surface block detection method provided by the application can also have the following technical features: the block feature size calculation step comprises: based on the world coordinates of the block anchor frame, the length and width thereof are calculated, and the maximum value thereof is taken as the length of the track surface block; based on the block effective point cloud, a rail surface plane is fitted through the RANSAC algorithm; the inclination angle of the fitted rail surface plane is calculated, and the block effective point cloud is horizontally corrected based on the inclination angle; in the horizontally corrected block effective point cloud, the maximum value and the minimum value of the Z-axis coordinate value of the traversed points are found out, and the difference between the maximum value and the minimum value is calculated as the depth of the track surface block.

[0014] The track turnout rail surface block detection method provided by the application can also have the following technical features: after the block feature size calculation step, a block severity determination step is further included, which determines the severity of the track surface block based on the feature size and a predetermined track surface block severity standard, wherein the track surface block severity standard comprises a plurality of severity levels and corresponding threshold information.

[0015] The application provides a turnout rail piece rail surface block detection device for detecting rail surface block of a turnout rail piece, which is characterized by detecting based on two-dimensional images collected by a camera, track profile data collected by a line laser sensor and corresponding mileage information. The device comprises: a turnout rail piece identification part for identifying the starting point of the turnout rail piece from multiple frames of the two-dimensional images or multiple frames of the profile data based on the profile features of the starting point of the turnout rail piece and obtaining corresponding mileage information; a two-dimensional image extraction part for extracting a plurality of frames of the two-dimensional images corresponding to the turnout rail piece based on the mileage information of the starting point of the turnout rail piece and the predetermined extension length of the turnout rail piece; a rail surface block identification part for sequentially inputting the extracted two-dimensional images into a rail surface block identification model, which outputs block position information of the rail surface block, wherein the rail surface block identification model is a trained neural network model; a block effective point cloud generation part for extracting multiple frames corresponding to the rail surface block from multiple frames of the profile data based on the block position information and the calibration parameters of the camera and the line laser sensor and stacking them to generate a block effective point cloud; and a block feature size calculation part for calculating the feature size of the rail surface block based on the block effective point cloud.

[0016] Inventive action and effect

[0017] According to the turnout rail piece rail surface block detection method and device, the method comprises a turnout rail piece identification step, a two-dimensional image extraction step, a rail surface block identification step, a block effective point cloud generation step and a block feature size calculation step. Since the starting point of the turnout rail piece is identified in the profile data based on the profile features of the turnout rail piece, and the corresponding two-dimensional images are extracted based on the starting point of the turnout rail piece and the predetermined extension length for identifying the rail surface block, the identified rail surface block can correspond to the corresponding type of turnout rail piece. Since the camera image and the track profile data collected by the line laser sensor are combined, not only can the neural network be used to efficiently identify the rail surface block from the camera image, but the corresponding profile data can be further extracted to generate an effective point cloud, and the depth and other feature sizes of the block can be calculated based on the effective point cloud, so that more comprehensive detection data of the rail surface block can be obtained, and the depth and other sizes of the block do not need to be manually measured, which can be applied to actual field detection to improve the detection efficiency. In addition, since only the part of the profile data corresponding to the identified rail surface block is extracted to generate a point cloud for further analysis, and the profile data of the whole track does not need to be analyzed, the amount of calculation is relatively small, which is also very beneficial to the application of field detection. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a flowchart of the turnout rail piece rail surface block detection method in the embodiment of the application;

[0019] Figure 2is a perspective view of the track detection vehicle in an embodiment of the present application;

[0020] Figure 3 is a distribution diagram of sensors in a vehicle body unit in an embodiment of the present application;

[0021] Figure 4 is a perspective view of a load wheel assembly in an embodiment of the present application;

[0022] Figure 5 is a sectional view of the load wheel assembly in an embodiment of the present application;

[0023] Figure 6 is a side view of a pressing assembly in an embodiment of the present application;

[0024] Figure 7 is a perspective view of the pressing assembly in an embodiment of the present application;

[0025] Figure 8 is a flowchart of a switch rail piece identification step in an embodiment of the present application;

[0026] Figure 9 is a schematic diagram of a center rail cross-section starting point profile identification in an embodiment of the present application;

[0027] Figure 10 is a flowchart of a center rail theoretical starting point identification by fitting a plane in an embodiment of the present application;

[0028] Figure 11 is a block diagram of a rail surface block identification model in an embodiment of the present application;

[0029] Figure 12 is a flowchart of a block area identification step in an embodiment of the present application;

[0030] Figure 13 is a flowchart of a block effective point cloud generation step in an embodiment of the present application;

[0031] Figure 14 is a flowchart of a block size acquisition step in an embodiment of the present application;

[0032] Figure 15 is a block diagram of a switch rail piece rail surface block detection device in an embodiment of the present application;

[0033] Figure 16 is a structural schematic diagram of a single turnout in an embodiment of the present application;

[0034] Figure 17 is a sectional view of a center rail and two side wing rails thereof in an embodiment of the present application;

[0035] Figure 18 is a side view of a center rail starting point vicinity center rail in an embodiment of the present application.

[0036] Figure label:

[0037] Track inspection vehicle 100; vehicle body 20; vehicle body unit 21; housing 211; inclined end 2111a; camera 215; refractor 216; first-line laser sensor 217a; second-line laser sensor 217b; third-line laser sensor 217c; fourth-line laser sensor 217d; bearing mechanism 30; bearing wheel assembly 31; bearing wheel bracket 311; bearing wheel shaft 312; bearing wheel 313; encoder receiving slot 3133; bearing 314; brake component 315; odometer encoder 317; clamping assembly 32; clamping bracket 321; mounting block 3211; guide rod 3212; guide component 322; guide rail 3221; slider 3222; elastic element 323; wheel Body support 324; clamping wheel 326; locking component 327; locking mating part 3271; locking component 3272; wrench component 328; pushing mechanism 40; turnout rail component rail surface drop detection equipment 60; model storage unit 601; detection information storage unit 602; turnout drop detection communication unit 603; turnout rail component identification unit 604; two-dimensional image extraction unit 605; rail surface drop identification unit 606; drop effective point cloud generation unit 607; drop feature size calculation unit 608; turnout drop information storage unit 609; turnout drop detection control unit 610; single turnout 90; wing rail 96; frog rail 97; frog rail theoretical starting point 971; plane 972; inclined plane 973; frog rail cross-section starting point 974. Detailed Implementation

[0038] To make the technical means, creative features, objectives and effects of the present invention easy to understand, the following describes in detail the method and equipment for detecting rail surface sag of turnout rail components of the present invention with reference to embodiments and accompanying drawings.

[0039] <Example 1>

[0040] This embodiment provides a method and equipment for detecting spalling on the surface of turnout rail components. It is used to detect spalling on the surface of rail components in turnouts. The rail components at turnouts include various types such as main rails, switch rails, connecting rails, frog rails, wing rails, and guard rails. In this embodiment, the frog rail and the wing rails on both sides are used as examples for specific explanation.

[0041] Figure 16 This is a schematic diagram of the structure of a single turnout in this embodiment.

[0042] Figure 17 This is a cross-sectional view of the center rail and its two side wing rails in this embodiment.

[0043] like Figures 16-17As shown, taking a single turnout 90 as an example, the frog 97 is a variable cross-section rail component. Its starting point is triangular and pointed. From its starting point to the other end, its cross-sectional width gradually increases. The frog 97 is flanked by wing rails 96. The middle part of the wing rails 96 is relatively close to the frog 97. The two ends of the wing rails 96 bend outward (towards the direction away from the frog). The cross-sectional shape of the wing rails 96 is similar to that of the main rail.

[0044] Figure 18 This is a side view of the cardiac track near the starting point of the cardiac track in this embodiment.

[0045] like Figure 18 As shown, between the starting point 974 of the cross-section of the heart track and the theoretical starting point 971 of the heart track, the upper end face of the heart track 97 is a plane 972. After the theoretical starting point 971 of the heart track, the upper end face of the heart track is a smaller inclined plane 973. The intersection of the plane 972 and the inclined plane 973 is the theoretical starting point 971 of the heart track.

[0046] Figure 1 This is a flowchart of the method for detecting rail surface breakage of turnout components in this embodiment.

[0047] like Figure 1 As shown, the method for detecting rail surface spalling on turnout rail components includes the following steps:

[0048] Turnout data acquisition step S1: Acquire two-dimensional images of the turnout using a camera, acquire the turnout profile data using a line laser sensor, and acquire mileage information using a mileage encoder.

[0049] Step S2 for identifying turnout rail components: Based on the contour features of the starting point of the turnout rail component to be detected, the starting point of the turnout rail component is identified from the track profile data and the corresponding mileage information is obtained.

[0050] Step S3 for extracting two-dimensional images: Based on the mileage information of the starting point of the turnout rail component and the preset extension length of the turnout rail component, extract the two-dimensional image corresponding to the turnout component.

[0051] Step S4 for identifying the falling block area: Input the extracted two-dimensional images into the track surface falling block identification model in sequence. The model outputs the falling block location information.

[0052] Step S5 for generating effective point cloud of dropped block: Based on the location information of the dropped block and the calibration relationship of the camera and line laser sensor, extract the multiple frames corresponding to the dropped block from the multi-frame profile data, and stack them to generate effective point cloud of dropped block.

[0053] Step S6 for obtaining the size of the dropped block: Based on the location information of the dropped block and the characteristic size of the dropped block on the cloud computing orbital plane at the valid point of the dropped block.

[0054] Block drop degree determination step S7: based on the length and depth of the rail surface block drop and the preset rail block drop severity standard, the severity of the rail surface block drop is determined.

[0055] The above steps will be described in detail below.

[0056] Turnout data acquisition step S1: a two-dimensional image of the turnout is acquired by a camera, profile data of the turnout is acquired by a line laser sensor, and mileage information is acquired by a mileage encoder.

[0057] As one of the examples, a track detection vehicle provided with a camera, a line laser sensor and a mileage encoder is shown below.

[0058] Figure 2 is a perspective view of the track detection vehicle in this embodiment.

[0059] As shown in Figure 2 , the track detection vehicle 100 includes a vehicle body 20, a bearing mechanism 30 and a pushing mechanism 40, the vehicle body 20 is movably placed on two main line rails 9 through the bearing mechanism 30 at the bottom, and the pushing mechanism 40 is installed on the vehicle body 20 for the detection personnel to push the vehicle body 20.

[0060] The vehicle body 20 is a two-part structure, including two mirror-symmetrical vehicle body units 21. Each vehicle body unit 21 includes a shell 211 and sensors, electronic control components and the like arranged in the shell 211.

[0061] Figure 3 is a schematic view of the distribution of sensors in the vehicle body unit in this embodiment.

[0062] As shown in Figure 3 , the shell 211 of each vehicle body unit 21 is provided with a camera 215, a refracting mirror 216, a first line laser sensor 217a, a second line laser sensor 217b, a third line laser sensor 217c, and an illumination unit (not shown in the figure). The shell 211 has an opening below, and the camera and the line laser sensor can both detect the main line rail 9 below through the opening.

[0063] The camera 215 is installed on the top of the shell 211 through a corresponding bracket, and is installed horizontally with its lens facing the horizontal direction. The refracting mirror 216 is a 45° refracting mirror, also installed on the top of the shell 211 through a corresponding bracket, and installed in front of the lens of the camera 215. Through the refraction of the refracting mirror 216, the camera 215 can take pictures of the rail below. In this way, the installation of the camera 215 is more stable, and the line laser sensor can be avoided from being blocked.

[0064] The lighting unit includes a plurality of light tubes, which are also installed in the top of the shell 211 through corresponding supports, for providing sufficient lighting for the shooting of the camera 215, and the shielding of the shell 211 can reduce the influence of the change of external environmental light conditions on the shooting of the camera 215.

[0065] The first line laser sensor 217a is installed in the shell 211 through a corresponding support and located in the inclined end 2111a of the shell 211. The first line laser sensor 217a is installed horizontally. When the track detection vehicle 100 is placed on the basic rail, the first line laser sensor 217a is located on the outside of the basic rail, slightly higher than the basic rail and faces the basic rail, and the line laser projected by the first line laser sensor 217a covers the non-working edge of the basic rail, the lower jaw position on the side of the non-working edge, and part of the rail surface, and can obtain the corresponding point data.

[0066] The second line laser sensor 217b is installed obliquely on the top of the shell 211 through a corresponding support. When the track detection vehicle 100 is placed on the basic rail, the second line laser sensor 217b is located above the outside of the main rail and faces the basic rail, and the line laser projected by the second line laser sensor 217b covers the non-working edge of the basic rail, the waist on the side of the non-working edge, and the rail surface, and can obtain the corresponding point data.

[0067] The third line laser sensor 217c is installed obliquely on the top of the shell 211 through a corresponding support, and is closer to the middle of the vehicle body 20 than the second line laser sensor 217b. When the track detection vehicle 100 is placed on the basic rail, the third line laser sensor 217c is located above the inside of the basic rail and faces the basic rail, and the line laser projected by the third line laser sensor 217c covers the working edge of the basic rail, the waist on the side of the working edge, and the rail surface, and can obtain the corresponding point data. When there is a turnout rail such as a frog 92 on the inside of the basic rail 91, the line laser projected by the third line laser sensor 217c can also cover the upper end surface of the rail.

[0068] The fourth line laser sensor 217d is also installed obliquely on the top of the shell 211 through a corresponding support, and the inclination angle thereof is substantially the same as that of the third line laser sensor 217c, and the installation position thereof is closer to the middle of the vehicle body than that of the third line laser sensor 217c. When the track detection vehicle 100 is placed on the basic rail, the fourth line laser sensor 217d is located above the inside of the basic rail, close to the position between the two basic rails, and faces the position beside the rail bottom of the basic rail. When there is another turnout rail on the inside of the basic rail, the line laser projected by the fourth line laser sensor 217d covers the upper end surface and the inside surface of the turnout rail. It can be seen that, taking the frog 92 as an example, the line laser of the fourth line laser sensor 217d can cover the frog 92 in the abutting state (abutting with the basic rail) and the frog 92' in the repelling state.

[0069] In this embodiment, the models of the several line laser sensors on the track detection vehicle 100 are the same.

[0070] As shown in Figure 2 , the bearing mechanism 30 includes four bearing wheel assemblies 31 and four pressing assemblies 32. The four bearing wheel assemblies 31 are respectively installed at the lower part of the width direction of the vehicle body 20, and are installed in pairs at the position close to one end of the length direction of the vehicle body 20, corresponding to the two rails 9 respectively. The four pressing assemblies 32 are also installed at the lower part of the width direction of the vehicle body 20, and are respectively located beside the four bearing wheel assemblies 31. The track detection device 100 is movably placed on the two rails 9 through the four bearing wheel assemblies 31, at this time, the four pressing assemblies 32 are respectively located at the inner side of the corresponding side rail 9.

[0071] Figure 4 is a perspective view of the bearing wheel assembly in this embodiment.

[0072] Figure 5 is a sectional view of the bearing wheel assembly in this embodiment.

[0073] As shown in Figures 4-5 , the bearing wheel assembly 31 includes a bearing wheel support 311, a bearing wheel rotating shaft 312, a bearing wheel 313, a bearing 314 and a brake component 315.

[0074] The two ends of the bearing wheel rotating shaft 312 are respectively installed on the bearing wheel support 311 through two bearings 314. The bearing wheel 313 is installed on the bearing wheel rotating shaft 312 and is limited by the two bearings 314. In this embodiment, the bearing wheel 313 is made of insulating material, preferably ceramic or POM plastic material, and the insulation resistance value of the bearing wheel 313 is not less than 1MΩ, so as to avoid the detection device 100 to make the two rails connected to be powered, preventing interference with the circuit system of the track.

[0075] The bearing wheel 313 has an encoder accommodating groove 3133 inside, and a distance detection encoder 317 is arranged in the two encoder accommodating grooves 3133 of the bearing wheel 313, for obtaining the corresponding mileage information when the bearing wheel 313 rolls along the rail 9. The size of the encoder accommodating groove 3133 is slightly larger than the outer size of the distance detection encoder 317, so that the distance detection encoder 317 does not directly contact with the bearing wheel 313, which can avoid affecting the rotation of the bearing wheel 313 due to friction.

[0076] It should be noted that the structure of the vehicle body 20 and the installation position of the line laser sensor enable the plurality of line laser sensors to scan the profile data of the same cross section of the straight rail 9. Under the control of the corresponding controller, the plurality of line laser sensors synchronously scan after a predetermined time or a predetermined distance during movement of the trolley. Each frame of point data (point cloud) collected by the line laser sensor contains several hundred to several thousand coordinate points, and each coordinate point contains mileage information (Y axis), height information (Z axis), width information (X axis), and brightness information.

[0077] Similarly, under the control of the corresponding controller, the camera takes a picture once after a predetermined distance or a predetermined distance during movement of the trolley, and obtains a frame of two-dimensional image. In this embodiment, a two-dimensional image is obtained every 500 mm, and a frame of profile point data is obtained every 2 mm.

[0078] In addition, before starting data acquisition, the camera and all line laser sensors are calibrated to obtain corresponding calibration parameters.

[0079] Figure 6 is a side view of the pressing assembly in this embodiment.

[0080] Figure 7 is a perspective view of the pressing assembly in this embodiment.

[0081] As shown in Figures 6-7 , the pressing assembly 32 includes a pressing support 321, a guide component 322, an elastic member 323, a wheel body support 324, a pressing wheel rotating shaft (not shown), a pressing wheel 326, a locking component 327, and a wrench component 328.

[0082] The pressing support 321 includes a mounting block 3211 and a guide rod 3212. The guide rod 3212 is a cylindrical rod, and its extension direction is consistent with the length direction of the vehicle body 20.

[0083] The guide component 322 includes a guide rail 3221 and a sliding block 3222. The guide rail 3221 is fixedly installed on the housing 211, and its extension direction is the same as that of the guide rod 3212. The sliding block 3222 is slidably installed on the guide rail 3221.

[0084] The elastic member 323 is a spring, which is also sleeved on the guide rod 3212. One end of the spring abuts against the mounting block 3211, and the other end is embedded in a cylindrical groove on the wheel body support 324, and abuts against the groove bottom.

[0085] The pressing wheel rotating shaft is installed in a notch at the end of the wheel body support 324, and the pressing wheel 326 is rotatably installed on the pressing wheel rotating shaft and partially embedded in the notch. The pressing wheel 326 is also made of the above-mentioned insulating material.

[0086] Therefore, under the spring force of the elastic member 323, the wheel body support 324 and the pressing wheel 326 thereon can be pressed towards the inner side of the rail, so that the pressing wheel 326 is tightly attached to the acting side of the rail. In this embodiment, when the track detection vehicle 100 is placed on the rail, the pressing wheel 326 is located 16 cm below the rail surface.

[0087] The locking member 327 comprises a locking fitting member 3271 and a locking member 3272.

[0088] The locking member 3272 is a locking screw, which is installed at the corresponding locking member hole of the mounting block 3211, and the screw end portion thereof is movable along the hole.

[0089] The locking fitting member 3271 is in the shape of a strip plate, and a plurality of circular holes are formed along the length direction thereof. One end of the locking fitting member 3271 is installed on the wheel body support 324, and the other end thereof passes through the locking fitting hole of the mounting block 3211 and is movable along the hole. The above-mentioned locking member hole and the locking fitting hole are in communication. When the screw end portion of the locking member 3272 extends downward, the screw end portion passes through the locking member hole and reaches the locking fitting hole, and penetrates into the circular hole at one end of the locking fitting member 3271, so as to fix (lock) the relative positions of the wheel body support 324 and the mounting block 3211.

[0090] In this embodiment, among the two pressing assemblies 32 corresponding to one rail 9, the locking member 3272 is in the unlocked state, and the elastic member 323 is movable freely; among the two pressing assemblies 32 on the other side, the locking member 3272 is in the locked state, and the positions of the wheel body support 324 and the mounting block 3211 are fixed, that is, the position of the pressing wheel 326 relative to the vehicle body 20 is fixed. That is to say, in this embodiment, the two pressing assemblies 32 on one side are in the form of fixed side wheels, and the two pressing assemblies 32 on the other side are in the form of spring side wheels.

[0091] The wrench member 328 comprises two combined connecting rods, which are used by the detection personnel to adjust the relative position of the pressing wheel 326 relative to the vehicle body 20, so as to more conveniently place the track detection vehicle 100 on the rail.

[0092] The pushing mechanism 40 comprises a trolley rod and a notebook support installed at the end of the trolley rod. The angle of the trolley rod is adjustable, and the angle of the notebook support relative to the trolley rod is also adjustable.

[0093] In step S1, the detection personnel places the above-mentioned track detection vehicle 100 on the turnout track, and starts the camera, line laser sensor, data acquisition device, etc. on the vehicle, and then pushes the track detection vehicle 100 along the turnout track, and the track detection vehicle 100 automatically acquires track data according to the setting during movement.

[0094] Step S2 for identifying turnout rail components: Based on the contour features of the starting point of the turnout rail component to be detected, the starting point of the turnout rail component is identified from the profile data and the corresponding mileage information is obtained.

[0095] As described above, in this embodiment, the turnout rail component to be detected includes the frog rail and the wing rails on both sides. First, the starting point of the frog rail is identified and located, and then the starting points of the wing rails on both sides are further identified and located based on the identified starting point of the frog rail.

[0096] Figure 8 This is a flowchart of the turnout rail component identification steps in this embodiment.

[0097] like Figure 8 As shown, the turnout rail component identification step S2 specifically includes the following steps:

[0098] Step S2-1: Based on the mileage information, perform profile matching and recognition of the starting point of the heart track section frame by frame on the collected profile data, identify the starting point of the heart track section and obtain the corresponding mileage information.

[0099] Figure 9 This is a schematic diagram of the identification of the starting point of the central track section in this embodiment.

[0100] like Figure 9 As shown, at the starting point of the track cross-section, the track profile is square. In step S2-1, based on the predetermined track cross-section starting point matching template, it is determined whether there is a profile at a predetermined position (the middle between the two wing rails) in the profile data. If a profile is determined to exist, the profile is cut off according to the predetermined profile cutting algorithm to obtain the profile of the upper end face of the track (roughly a horizontal straight line segment). It is then determined whether the cut-off profile is a horizontal straight line segment with a predetermined length. If it is, the starting point of the track cross-section is identified, and the mileage information corresponding to the profile data frame is used as the mileage information of the starting point of the track cross-section.

[0101] Step S2-2: Based on the mileage information of the starting point of the track section, obtain the profile data of a predetermined number of consecutive frames after the starting point of the track section, and stack them to generate a point cloud.

[0102] Each frame of point data is data in the XZ plane coordinate system, and each point contains corresponding mileage information. According to the mileage information, multiple frames of point data are stacked along the Y-axis and converted into a three-dimensional point cloud according to the corresponding algorithm.

[0103] Step S2-3: Based on the generated point cloud, identify the theoretical starting point of the orbit by fitting a plane and obtain the corresponding mileage information.

[0104] Figure 10 This is a flowchart illustrating the identification of the theoretical starting point of the orbital path using a plane fitting method in this embodiment.

[0105] As Figure 10 shown, step S2-3 specifically includes the following sub-steps:

[0106] Step S2-3-1: based on the point cloud, fitting a plane between the start point of the rail profile and the theoretical start point of the rail profile by the RANSAC algorithm.

[0107] Step S2-3-2: based on the point cloud, fitting a slope after the theoretical start point of the rail profile by the RANSAC algorithm.

[0108] Step S2-3-3: calculating the Y-axis coordinate of the point on the intersection line of the fitted slope and the fitted plane, and obtaining the corresponding mileage information based on the Y-axis coordinate of the point on the intersection line as the mileage information of the theoretical start point of the rail profile.

[0109] Alternatively, the center line of the fitted slope, the center line of the fitted plane (both are center lines along the Y-axis direction) can also be calculated respectively, and the Y-axis coordinate of the intersection point of the two center lines is calculated, and the corresponding mileage information is obtained.

[0110] Wherein, the plane fitted by the RANSAC algorithm is specifically:

[0111] At least three points (the minimum number of points that can fit a plane model) are randomly sampled in the corresponding three-dimensional point cloud as initial points. Let the initial equation of the plane model be AX+BY+CZ+D=0, and the plane model is fitted based on a plurality of initial points to obtain the values of the plane model parameters A, B, C, D. In the three-dimensional point cloud, traverse the points other than the initial points, and calculate the distance d of the traversed point to the plane model:

[0112]

[0113] In the formula, x0, y0, z0 are the coordinates of the traversed point. The points with a distance d less than a preset distance threshold are in the inlier group, and the proportion of the number of inlier points to the total number of points in the three-dimensional point cloud is calculated. When the proportion is greater than a preset proportion threshold, the updated plane model parameters are obtained based on the inlier points, and the fitting of the corresponding plane is completed.

[0114] Step S2-4: based on the mileage information of the theoretical start point of the rail profile, extracting the profile data within a predetermined position range before the theoretical start point of the rail profile.

[0115] The start point of the wing rail on both sides of the rail is located at a predetermined distance before the theoretical start point of the rail, therefore, based on the identified theoretical start point of the rail, the approximate position range of the start point of the wing rail is inferred, and only a small amount of profile data within the position range is extracted to further identify and locate the start point of the wing rail.

[0116] Step S2-5: The extracted profile data is matched with the profile of the wing rail starting point frame by frame, the starting point of the wing rail is identified, and the corresponding mileage information is obtained.

[0117] Before the starting point of the wing rail, the profile of the wing rail (at the gap between the connecting rail and the wing rail) is not present in the collected profile, and at the starting point of the wing rail, the profiles of the two wing rails are further present in the collected profile. The profile of the wing rail is similar to that of the basic rail, and the distance between the two wing rails at the starting point of the wing rail is a predetermined value, so the profile data frame of the starting point of the wing rail can be identified based on the profile of the wing rail and the distance between the two wing rails.

[0118] Specifically, based on the preset wing rail starting point matching template, it is first determined whether there is a profile present in the two predetermined position ranges (corresponding to the positions of the two wing rail profiles at the starting point of the wing rail); when it is determined that there is a profile, the profile in the two predetermined position ranges is intercepted according to a predetermined profile interception algorithm, and the rail head part profile in the two wing rail profiles is intercepted; then, the intercepted rail head part profile is matched with the standard wing rail starting point profile, the similarity is calculated, and if the calculated similarity reaches a predetermined similarity threshold, it is determined that the starting point position of the wing rail is identified in the current profile data frame, and the mileage information corresponding to the frame profile data is taken as the mileage information of the starting point of the wing rail.

[0119] Two-dimensional image extraction step S3: based on the mileage information of the starting point of the turnout rail piece and the preset extension length of the turnout rail piece, a plurality of two-dimensional images corresponding to the turnout rail piece are extracted as the to-be-detected images.

[0120] Based on the mileage value of the starting point of the center rail, the predetermined extension length of the center rail, the mileage value of the starting point of the wing rail, and the predetermined extension length of the wing rail, the mileage range from the starting point to the terminal point of the center rail and the mileage range from the starting point to the terminal point of the wing rail can be calculated, and all two-dimensional images in the mileage range are extracted, which cover the entire center rail and wing rail.

[0121] Drop block area identification step S4: the extracted to-be-detected images are sequentially input into a rail surface drop block identification model, and the model outputs the drop block position information of the identified rail surface drop block.

[0122] Figure 11 is a block diagram of the rail surface drop block identification model in this embodiment.

[0123] As Figure 11As shown, in this embodiment, the track surface block recognition model is an improved version of the YOLOv5 neural network model. It includes a Backbone module, a PAN+Bi-FPN module, and a Head module. Compared to the existing YOLOv5 convolutional neural network model, the improvement lies in the addition of an upsampling layer after the Backbone module to increase the resolution of the feature map, thereby improving the recall rate of small-sized blocks. The original FPN structure is replaced with a PAN+Bi-FPN structure to deepen the network and improve the model's feature fusion performance.

[0124] When training the model, a large number of images of rail surface spalling, along with corresponding spalling labels and rail surface labels, were first prepared as training data. The rail surface spalling images were two-dimensional images of the rails acquired in the same manner as described above, and the spalling labels and rail surface labels were manually annotated by experienced inspection personnel. Then, data augmentation was performed on the training data to improve the robustness and generalization ability of the trained model. Finally, the augmented training data was used to train the rail surface spalling region recognition model.

[0125] In this embodiment, the data augmentation algorithm includes geometric transformations of the two-dimensional orbital image, such as cropping, scaling, translation, flipping, and rotation, to obtain datasets of different physical sizes. Then, color-layer data transformations are applied to the orbital image, including noise addition, color perturbation, and blurring. Through this data augmentation algorithm, a new dataset several times larger than the original training dataset is obtained.

[0126] Inference is performed using a trained track surface rockfall region recognition model. The input is a 2D image of the track captured by a camera. The model fuses the extracted features using its PAN and Bi-FPN structures, and then outputs the fused features to the Head module. The Head module includes a multi-scale detection head. The Head module outputs the position coordinates and confidence score (denoted as the first confidence score) of the rockfall anchor frame identifying the rockfall region in the 2D image, and the position coordinates and confidence score (denoted as the second confidence score) of the track surface anchor frame identifying the rail component's track surface region. Specifically, the anchor frame is a rectangular frame, and its position coordinates include the coordinates of its upper left and lower right corners; all coordinates here refer to the pixel coordinates of the 2D image.

[0127] It is understandable that the drop block anchor frame should be located within the rail surface anchor frame. Therefore, if a drop block anchor frame is located outside the rail surface anchor frame in the output, then that drop block anchor frame will be filtered out.

[0128] Figure 12 This is a flowchart of the block drop area identification step in this embodiment.

[0129] like Figure 12As shown, the block area identification step S4 specifically includes the following steps:

[0130] Step S4-1: Preprocessing the extracted to-be-detected image.

[0131] Step S4-2: Inputting the preprocessed to-be-detected image into the rail surface block identification model, which outputs the block anchor frame and its first confidence, the rail surface anchor frame and its second confidence.

[0132] Step S4-3: Removing the block anchor frame outside the rail surface anchor frame with a first confidence lower than a preset confidence threshold.

[0133] Block effective point cloud generation step S5: based on the block position information, the calibration parameters of the camera and the line laser sensor, extracting the multi-frame profile data corresponding to the rail surface block from the profile data, and generating the block effective point cloud.

[0134] Figure 13 is the flow chart of the block effective point cloud generation step in this embodiment.

[0135] As shown, the block effective point cloud generation step S5 specifically includes the following steps: Figure 13

[0136] Step S5-1: Based on the mileage information corresponding to a frame of to-be-detected image, extracting the corresponding multi-frame profile data from all collected profile data.

[0137] Step S5-2: Stacking the extracted multi-frame profile data along the Y axis according to its mileage information, and converting it into a rail piece point cloud through a corresponding conversion algorithm.

[0138] Step S5-3: Based on the distance between the line laser sensor and the rail surface of the turnout rail piece, the rail piece point cloud is cropped to obtain the rail surface point cloud.

[0139] Step S5-4: Converting the pixel coordinates of the block anchor frame into camera coordinates based on the camera intrinsic parameters, and converting the camera coordinates of the block anchor frame into world coordinates based on the calibration relationship between the camera and the line laser sensor.

[0140] That is, the coordinates of the pixel points are corrected through the camera intrinsic parameter matrix to obtain their real pixel positions, and then the pixel positions are projected into the world coordinate system of the three-dimensional point cloud.

[0141] Wherein, the camera intrinsic parameter matrix is:

[0142]

[0143] The coordinates (u, v) of the pixel points in the anchor frame are converted into camera coordinates (X, Y, Z) according to the intrinsic parameter matrix:

[0144] ​

[0145] The camera coordinates (X, Y, Z) can be obtained as follows:

[0146]

[0147]

[0148] Furthermore, camera coordinates P C World coordinates P are generated using rotation matrix R and offset matrix t. W :

[0149] P C =RP W +t

[0150] The rotation matrix R and the offset matrix t can be obtained from the calibration parameters.

[0151] The world coordinates P are solved using the following matrix. W :

[0152]

[0153]

[0154] Step S5-5: Based on the world coordinates of the dropped anchor frame, extract the part of the point cloud corresponding to the dropped block from the point cloud of the track surface, and use it as the effective point cloud of the dropped block.

[0155] This step extracts a small-scale point cloud corresponding to the area where the rail fell off, which greatly reduces the amount of data compared to the point cloud of the entire rail section.

[0156] Step S6 for obtaining the size of the dropped block: Based on the location information of the dropped block and the feature size of the dropped block on the cloud computing orbital plane at the valid point of the dropped block.

[0157] In this embodiment, the length and depth of the rail surface block are calculated.

[0158] Figure 14 This is a flowchart of the block size acquisition step in this embodiment.

[0159] like Figure 14 As shown, step S6, which involves obtaining the size of the dropped block, specifically includes the following steps:

[0160] Step S6-1: Based on the world coordinates of the drop block anchor frame, calculate the length and width of the drop block anchor frame, and take the maximum value as the length of the track surface drop block.

[0161] Step S6-2: Based on the effective point cloud of the dropped blocks, fit the orbital plane using the RANSAC algorithm.

[0162] Step S6-3: Calculate the tilt angle of the fitted rail surface plane along the Y axis, and based on the tilt angle, perform horizontal correction on the effective point cloud of the block.

[0163] Step S6-4: Traverse the effective point cloud of the block after horizontal correction, find the maximum value and the minimum value in the Z axis coordinate values of the traversed points, and calculate the difference between the maximum value and the minimum value as the depth of the rail surface block.

[0164] Wherein, the method of fitting the rail surface plane by RANSAC algorithm is similar to the above, that is, set the plane model equation, get enough proportion of in-group points, and update the parameters A, B, C, D of the plane model. After that, in step S6-3, the tilt angle of the fitted rail surface plane along the Y axis is calculated according to the following formula:

[0165]

[0166] Based on the above tilt angle θ, the rotation matrix of the effective point cloud of the block along the Y axis is obtained:

[0167]

[0168] And the effective point cloud P raw of the block is corrected horizontally by the rotation matrix:

[0169] P rotate = R y (θ)P raw

[0170] Block degree determination step S7: based on the length and depth of the rail surface block and the preset rail block severity standard, the severity of the rail surface block is determined.

[0171] Wherein, the rail block severity standard includes block severity level and corresponding threshold information, and in this embodiment, the standard is specifically shown in the following table 1.

[0172] Table 1: Rail surface block severity threshold information table

[0173]

[0174]

[0175] Therefore, based on the obtained set of length and depth data of the block area, it can be determined that the severity of the block area falls within which threshold range in the table.

[0176] In this embodiment, the two-dimensional image corresponding to each identified rail surface block, the mileage information, the rail surface point cloud, the effective point cloud of the block, the length and depth of the rail surface block, and the severity are stored as turnout rail block information. This facilitates the inspection and further analysis of the automatically identified rail surface block and its information by the detection personnel. For example, the detection personnel can filter out the block information with severe severity from the stored block information for priority processing. Or the detection personnel can call up the two-dimensional image and point cloud data of the block area for review. Or the detection personnel can locate the corresponding position on the rail piece according to the mileage information corresponding to the block area and go to the position for further manual detection.

[0177] Figure 15 is a block diagram of the turnout rail block detection device in this embodiment.

[0178] As shown in Figure 15 , this embodiment also provides a turnout rail block detection device 60 corresponding to the above method, which includes a model storage unit 601, a detection information storage unit 602, a turnout block detection communication unit 603, a turnout rail recognition unit 604, a two-dimensional image extraction unit 605, a rail surface block recognition unit 606, an effective block point cloud generation unit 607, a block feature size calculation unit 608, a turnout block information storage unit 609, and a turnout block detection control unit 610.

[0179] The model storage unit 601 stores the above rail surface block recognition model (neural network model). The detection information storage unit 602 stores the parameters required for the detection of the rail surface block of the turnout rail. The turnout block detection communication unit 603 is used for communication with other devices, including obtaining the collected two-dimensional image, profile data, and mileage information from the data acquisition device of the track detection vehicle 100. The turnout rail recognition unit 604 identifies the starting point of the turnout rail from the two-dimensional image or profile data according to the method of step S2. The two-dimensional image extraction unit 605 extracts multiple frames of two-dimensional images corresponding to the turnout rail for block detection according to the method of step S3. The rail surface block recognition unit 606 inputs the extracted two-dimensional image into the rail surface block recognition model to identify the rail surface block and obtain its position information according to the method of step S4. The effective block point cloud generation unit 607 extracts the effective block point cloud based on the position information of the identified rail surface block according to the method of step S5. The block feature size calculation unit 608 calculates the feature size of the rail surface block according to the method of step S6. The turnout block information storage unit 609 is used to store the above turnout rail block information. The turnout block detection control unit 610 is used to control the work of the above functional units.

[0180] As one of the examples, the rail surface spalling detection device 60 is a notebook computer provided with a corresponding detection program, which is placed on the notebook computer placing rack of the track detection vehicle 10 and connected with the vehicle-mounted data acquisition device through a cable. The detection personnel can conveniently obtain the detection result through the notebook computer.

[0181] In this embodiment, the parts not described in detail are the known technologies in the art.

[0182] Effects of the embodiments

[0183] According to the turnout rail piece rail surface spalling detection method and device provided in the embodiment, the method comprises a turnout rail piece identification step, a two-dimensional image extraction step, a rail surface spalling identification step, a spalling effective point cloud generation step, and a spalling feature size calculation step. Since the profile features of the turnout rail piece are identified based on the profile data, the starting point of the turnout rail piece is identified, and the corresponding two-dimensional image is extracted based on the starting point of the turnout rail piece and the predetermined extension length to identify the rail surface spalling, the identified rail surface spalling can correspond to the corresponding type of turnout rail piece. Since the camera image and the track profile data collected by the line laser sensor are combined, not only the rail surface spalling can be efficiently identified from the camera image by using the neural network, but also the corresponding profile data can be further extracted to generate the effective point cloud, and the depth and other feature sizes of the spalling can be calculated based on the effective point cloud, so that more comprehensive detection data of the rail surface spalling can be obtained, and the depth and other sizes of the spalling do not need to be manually measured, which can be applied to actual field detection to improve the detection efficiency. In addition, since only the part of the profile data corresponding to the identified rail surface spalling is extracted to generate the point cloud for further analysis, the calculation amount is relatively small, which is also very beneficial to the application of field detection.

[0184] In the embodiment, the track detection vehicle carrying the camera, the line laser sensor, and the mileage encoder is used to collect the rail data on site, and the detection personnel only need to start the device and move the vehicle along the rail to collect the data, which is convenient to operate, and the vehicle can obtain the required multiple data by walking once. Further, the camera is arranged in the shell of the detection vehicle, and the shell is also provided with an illumination unit, so that the influence of the change of external light conditions on the camera shooting can be reduced, and sufficient illumination can be provided for the camera shooting, thereby improving the quality of the two-dimensional image collected by the camera.

[0185] Further, the notebook computer is also placed on the handle of the track detection vehicle, the notebook computer stores the preset detection program, and is connected with the camera and the sensor through the data acquisition device, so that the detection personnel can conveniently obtain the real-time data collection situation and the rail surface spalling detection result through the notebook computer, and timely find out the problems and make adjustments or re-measurements.

[0186] In the embodiment, the improved YOLOv5 neural network model is used for automatic identification of the block area. An upsampling layer is added at the end of the Backbone module in the model, so that the resolution of the feature map can be improved, thereby improving the recall rate of small size block; the FPN structure in the original model is optimized to PAN+Bi-FPN structure, which deepens the network and improves the feature fusion performance of the model. Therefore, the model is used for rail surface block identification, and the identification accuracy is higher.

[0187] In the embodiment, the theoretical starting point of the center rail and the starting point of the wing rail are identified based on the profile matching and plane fitting, so that the multi-frame two-dimensional images covering the rail pieces can be extracted based on the identified starting points of the rail pieces and the preset extension length of the rail pieces, for block detection. The detected block can be corresponded to the corresponding type of rail piece, so that the automatic detection of the rail surface block of the multiple rail pieces in the turnout area can be realized. In the profile matching, the interception and matching identification algorithm involved in each frame of profile data is relatively simple and has small calculation amount, so that the starting point of the corresponding rail piece can be quickly identified, and the identified starting point position is accurate.

[0188] In the embodiment, only the small range of rail surface point cloud is extracted as the effective point cloud according to the block anchor frame from all the collected profile data, for the calculation of the block length and depth. It can be seen that the amount of point cloud data involved is small, so the calculation amount is small, the detection result can be quickly obtained in the field detection, and the time of field detection is reduced.

[0189] Further, the severity of the block is automatically determined based on the length and depth of the block area and the preset severity standard, and the two-dimensional image, point cloud data, calculated size, severity and other information corresponding to the block area are correspondingly stored. Therefore, the detection personnel can conveniently recheck and confirm or further analyze the identified rail surface block, and can for example screen out the block area of the heavy injury grade for priority processing, so as to further improve the efficiency and accuracy of the rail surface block detection, and the degree of automation is high, which can reduce the burden of the detection personnel.

[0190] The above embodiments are only used to illustrate the specific embodiments of the present application, and the present application is not limited to the description range of the above embodiments.

[0191] In the above embodiments, the center rail and the wing rails on both sides thereof at the turnout are taken as examples for specific description. It can be understood that the method can also be used for other steel rails at the turnout.

[0192] In the above embodiment, the starting point of the turnout rail piece is identified from the profile data of multiple frames based on the profile features of the turnout rail piece, so that the identified rail surface block can be corresponded to the turnout rail piece. In an alternative, the starting point or type of the turnout rail piece can also be identified from the two-dimensional image of the camera based on the structural features of the turnout rail piece in the image, for example, inputting the two-dimensional image into a trained turnout rail piece classification model (neural network model), and the model outputs the type and probability score of the rail piece in the two-dimensional image. For example, the starting point of the switch rail is identified from the two-dimensional image, and the mileage information corresponding to the frame of the two-dimensional image is taken as the mileage information of the starting point of the switch rail, and the corresponding technical effects can also be achieved.

[0193] In the above embodiment, the slight injury, light injury, heavy injury grades of the rail surface block and the corresponding threshold information are set. In an alternative, more grades and corresponding threshold information can also be set according to actual detection needs.

[0194] In the above embodiment, for the convenience of description, the two-dimensional image extraction, rail surface block identification and other steps are performed after the turnout data acquisition step is completed. It can be understood that, in the case of sufficient computing power, the two-dimensional image extraction, rail surface block identification and other steps can also be performed on the data that has been collected during the data collection process, thereby further improving the real-time performance of the detection.

Claims

1. A method for detecting a rail surface spalling of a turnout rail, for detecting a rail surface spalling of a turnout rail, characterized by, The method is based on two-dimensional images collected by a camera, profile data collected by a line laser sensor, and corresponding mileage information for detection, and the method comprises: A turnout rail piece identification step: based on the profile features of the starting point of the turnout rail piece, the starting point of the turnout rail piece is identified from the two-dimensional images or multiple frames of profile data, and the corresponding mileage information is obtained; A two-dimensional image extraction step: based on the mileage information of the starting point of the turnout rail piece and the predetermined extension length of the turnout rail piece, a plurality of frames of two-dimensional images corresponding to the turnout rail piece are extracted as detection images; A rail surface blockage identification step: the detection images are sequentially input into a rail surface blockage identification model, and the model outputs blockage position information of the rail surface blockage, wherein the rail surface blockage identification model is a trained neural network model; A blockage effective point cloud generation step: based on the blockage position information and the calibration parameters of the camera and the line laser sensor, multiple frames of profile data corresponding to the rail surface blockage are extracted from multiple frames of profile data, and are stacked to generate a blockage effective point cloud; A blockage feature size calculation step: based on the blockage position information and the blockage effective point cloud, the feature size of the rail surface blockage is calculated.

2. The method of claim 1, Its features are: Wherein, The turnout rail piece comprises a center rail and wing rails arranged on both sides of the center rail, The turnout rail piece identification step comprises: Based on the mileage information, the profile matching identification of the starting point of the center rail section is performed on the collected profile data frame by frame, the starting point of the center rail section is identified, and the corresponding mileage information is obtained; Based on the mileage information of the starting point of the center rail section, a predetermined number of frames of profile data connected after the starting point of the center rail section are obtained, which are stacked to generate a point cloud; Based on the point cloud, the center rail theoretical starting point is identified by fitting a plane, and the corresponding mileage information is obtained; Based on the mileage information of the center rail theoretical starting point, profile data within a predetermined position range before the center rail theoretical starting point is extracted; The profile matching identification of the starting point of the wing rail is performed on the extracted profile data frame by frame, the starting point of the wing rail is identified, and the corresponding mileage information is obtained, In the two-dimensional image extraction step, based on the mileage information of the center rail theoretical starting point, the extension length of the predetermined center rail, the mileage information of the starting point of the wing rail, and the extension length of the predetermined wing rail, a plurality of frames of two-dimensional images covering the center rail and the wing rail are extracted.

3. The method of detecting a rail surface chunk of a turnout rail according to claim 2, Its features are: Wherein, The identification of the center rail theoretical starting point by fitting a plane comprises: Based on the point cloud, a plane between the starting point of the center rail section and the center rail theoretical starting point is fitted by RANSAC algorithm; Based on the point cloud, an inclined surface after the center rail theoretical starting point is fitted by RANSAC algorithm; The coordinates of the points on the intersection line of the fitted plane and the fitted inclined surface are calculated, and the corresponding mileage information is obtained based on the coordinates, or the center lines of the fitted plane and the fitted inclined surface are calculated respectively, and the coordinates of the intersection point of the two center lines are calculated, and the corresponding mileage information is obtained based on the coordinates.

4. The turnout rail surface chunk detection method according to claim 2, characterized in that: wherein the matching of the profile of the point rail section starting point comprises: based on a predetermined point rail section starting point matching template, judging whether there is a profile at a predetermined position in the profile data, when judging that there is a profile, performing profile cutting on the profile according to a predetermined profile cutting algorithm to obtain the profile of the upper end surface of the point rail, judging whether the cut profile is a straight line segment with a predetermined length, when judging that it is, identifying the point rail section starting point, and taking the mileage information corresponding to the profile data of the frame as the mileage information of the point rail section starting point, the matching of the profile of the wing rail starting point comprises: based on a predetermined wing rail starting point matching template, judging whether there is a profile at a predetermined position in the profile data, when judging that there is a profile, performing profile cutting on the profile according to a predetermined profile cutting algorithm to obtain the profile of the rail head part of the two side wing rails, matching the cut profile with a predetermined standard wing rail starting point profile, calculating the similarity, and when the similarity reaches a predetermined similarity threshold, judging that the wing rail starting point is identified, and taking the mileage information corresponding to the profile data of the frame as the mileage information of the wing rail starting point.

5. The method of claim 1, characterized in that: wherein, the rail surface chunk area recognition model is an improved YOLOv5 neural network model, which comprises: a Backbone module for extracting features of the two-dimensional image, and an upsampling layer is additionally provided after the Backbone module; a PAN+Bi-FPN module for fusing the extracted features to obtain fused features; and a Head module for identifying chunk areas and rail surface areas in the two-dimensional image based on the fused features, and outputting pixel coordinates of corresponding chunk anchor boxes and their first confidence, and pixel coordinates of rail surface anchor boxes and their second confidence.

6. The turnout rail surface chunk detection method according to claim 5, characterized in that: wherein the rail surface chunk identification step further comprises: removing the chunk anchor boxes with a first confidence lower than a preset confidence threshold and the chunk anchor boxes located outside the rail surface anchor boxes from the identified chunk anchor boxes.

7. The method of detecting rail surface defects of a switch rail according to claim 5, characterized in that: wherein, the chunk effective point cloud generation step comprises: based on the mileage information corresponding to the image to be detected, extracting corresponding multiple frames from the profile data and stacking them to generate a rail point cloud; based on the distance between the line laser sensor and the rail surface of the turnout rail, the rail point cloud is cropped to obtain a rail surface point cloud; based on the camera intrinsic parameter of the camera, the pixel coordinates of the chunk anchor box are converted into camera coordinates, and based on the calibration relationship between the camera and the line laser sensor, the camera coordinates of the chunk anchor box are converted into world coordinates; based on the world coordinates of the chunk anchor box, the chunk effective point cloud corresponding to the rail surface chunk is extracted from the rail surface point cloud.

8. The method of detecting rail surface defects of a switch rail according to claim 5, characterized in that: wherein, the chunk feature size calculation step comprises: based on the world coordinates of the chunk anchor box, the length and width thereof are calculated, and the maximum value among them is taken as the length of the rail surface chunk. Fitting a track surface plane by a RANSAC algorithm based on the block-out effective point cloud; Calculating the tilt angle of the fitted track surface plane, and performing horizontal correction on the block-out effective point cloud based on the tilt angle; Traversing the block-out effective point cloud after horizontal correction, finding the maximum and minimum values of the Z-axis coordinate values of the traversed points, and calculating the difference between the maximum and minimum values as the depth of the track surface block-out.

9. The method of claim 1, wherein, After the block-out feature size calculation step, further comprising: A block-out severity determination step, which determines the severity of the track surface block-out based on the feature size and a predetermined track surface block-out severity standard, Wherein the track surface block-out severity standard includes multiple severity levels and corresponding threshold information.

10. A track surface spalling detection device for turnout track components, used to detect spalling on the track surface of turnout track components, characterized in that, Based on the two-dimensional images collected by the camera, the profile data collected by the line laser sensor, and the corresponding mileage information, the device includes: A turnout rail piece recognition unit for identifying the starting point of the turnout rail piece from multiple frames of the two-dimensional images or multiple frames of the profile data based on the profile features of the starting point of the turnout rail piece and obtaining the corresponding mileage information; A two-dimensional image extraction unit for extracting a number of frames of the two-dimensional images corresponding to the turnout rail piece based on the mileage information of the starting point of the turnout rail piece and the predetermined extension length of the turnout rail piece; A track surface block-out recognition unit for sequentially inputting the extracted two-dimensional images into a track surface block-out recognition model, which outputs block-out position information of the track surface block-out, wherein the track surface block-out recognition model is a trained neural network model; A block-out effective point cloud generation unit for extracting multiple frames corresponding to the track surface block-out from multiple frames of the profile data based on the block-out position information and the calibration parameters of the camera and the line laser sensor, and stacking them to generate a block-out effective point cloud; and A block-out feature size calculation unit for calculating the feature size of the track surface block-out based on the block-out effective point cloud.

Images