Railway unmanned aerial vehicle intelligent cooperative inspection method based on industrial vision

Abstract

This invention discloses an intelligent collaborative inspection method for railway drones based on industrial vision, comprising the following steps: S1, acquiring an inspection parameter set; S2, constructing a task matrix and generating an inspection plan; S3, generating candidate routes; S4, identifying occluded and deviated sections and generating collaborative routes; S5, generating an inspection visual flow; S6, inputting an improved Mask DINO model, introducing a track structure guidance mechanism, and outputting target segmentation results; S7, employing an improved SuperGlue graph matching algorithm, introducing a track domain verification mechanism, and generating multi-drone defect results; S8, performing defect level classification and generating inspection handling instructions. This invention can improve the efficiency of collaborative inspection of drones in complex scenarios along railway lines, the accuracy of cross-view defect matching, and the precision of defect location, while reducing route conflict rates and repeated defect alarm rates.

Description

Technical Field

[0001] This invention relates to the field of intelligent railway inspection technology, and in particular to an intelligent collaborative inspection method for railway drones based on industrial vision. Background Technology

[0002] With the continuous expansion of railway lines and the increasing demand for intelligent operation and maintenance of railway infrastructure, unmanned aerial vehicle (UAV) intelligent inspection technology for trackside components, overhead contact line equipment, bridge structures, and slope areas along railway lines has attracted widespread attention. Existing railway inspection methods mainly rely on manual foot patrols, onboard inspection equipment, or single UAV visual inspections for defect identification. However, these methods generally suffer from the following problems in practical applications:

[0003] The railway inspection area spans a large area, with complex distribution of trackside components, and multiple drones' inspection paths are prone to overlap. Current drone inspection scheduling methods typically allocate tasks based solely on fixed routes or simple area divisions, lacking the ability to coordinate and handle line mileage, flight segment occupancy, and airspace conflicts. This leads to overlapping flight paths, missed observations, and chaotic inspection sequences among drones, affecting the continuity and efficiency of railway inspections. Furthermore, bridge, tunnel, overhead contact line, and slope areas along railway lines contain numerous obstructing structures and complex spatial forms. Existing inspection route planning methods struggle to dynamically adjust observation direction and position based on the spatial distribution of trackside components, easily resulting in insufficient visibility of trackside components, deviation in defective area observations, and loss of detail in distant defects, leading to poor integrity of industrial visual data. Furthermore, the accuracy of defect area identification decreases; for long, continuous targets such as rails, sleepers, overhead contact lines, and bridge edges along railway lines, traditional target detection and instance segmentation methods typically use general query vectors for target matching, lacking the ability to jointly express the continuous relationship of railway component boundaries, line mileage relationships, and spatial extension relationships. This leads to problems such as component boundary breaks, missegmentation of defect areas, and inconsistent segmentation results across viewpoints in complex backgrounds; industrial vision data acquired by multiple UAVs have differences in viewpoint, positional offset, and temporal differences. Existing graph matching technologies struggle to simultaneously integrate the correlation between line mileage, defect contours, and image positions, easily resulting in repeated defect identification, drifting defect localization, and cross-viewpoint defect matching errors, affecting the accuracy of railway defect localization and the reliability of inspection and handling results.

[0004] Therefore, how to provide a railway drone intelligent collaborative inspection method based on industrial vision is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent collaborative inspection method for railway drones based on industrial vision. This invention fully utilizes drone collaborative inspection, industrial vision, an improved Mask DINO model, and an improved SuperGlue graph matching algorithm. It describes in detail the collaborative processing methods of railway inspection task data, railway line data, and drone operation data. It also details the processes of inspection section division, task matrix construction, candidate route generation, extraction of visible areas of trackside components, collaborative route generation, target segmentation result generation, multi-drone defect result generation, and inspection handling instruction generation. This method has the advantages of high inspection collaboration, low route conflict rate, strong adaptability to complex railway scenarios, high accuracy of cross-view defect matching, and high defect location accuracy.

[0006] An intelligent collaborative inspection method for railway drones based on industrial vision, according to an embodiment of the present invention, includes the following steps:

[0007] S1. Acquire railway inspection task data, railway line data, and drone operation data; perform time correlation and line mileage correlation to generate inspection parameter set.

[0008] S2. Divide the inspection sections based on the inspection parameter set, match the UAV with the inspection sections, construct a task matrix, and generate an inspection plan.

[0009] S3. Generate an initial route based on the inspection plan, map the initial route to the airspace grid along the railway, perform route conflict detection and segment coordination, and generate candidate routes;

[0010] S4. Extract the visible area of ​​trackside components based on railway line data, identify occluded and deviated sections, perform view compensation and segment rearrangement on candidate routes, and generate cooperative routes.

[0011] S5. Control the drone to acquire industrial visual data along the railway line according to the collaborative flight path and generate an inspection visual flow.

[0012] S6. Input the inspection visual flow into the improved Mask DINO model, introduce a track structure query mechanism in the mask query generation module, extract the track structure boundary sequence, construct the track structure main domain, and output the target segmentation result.

[0013] S7. Based on the target segmentation results and the inspection visual flow, an improved SuperGlue graph matching algorithm is adopted, and a trace domain verification mechanism is introduced to generate multi-machine defect results;

[0014] S8. Perform spatial back projection and line mileage mapping on the multi-machine defect results to generate defect location results, and perform defect level classification based on the defect location results to generate inspection and handling instructions.

[0015] Optionally, S1 specifically includes:

[0016] Collect railway inspection task data, railway line data, and drone operation data to generate raw inspection data;

[0017] The raw inspection data is formatted and fields are merged to generate standard inspection data; a unified time stamp is added to the standard inspection data to generate time-series inspection data.

[0018] Based on the line mileage in the railway line data, mileage correlation is performed on the time-series inspection data to generate mileage inspection data;

[0019] The inspection tasks, railway lines, and drone status in the mileage inspection data are matched to generate an inspection parameter set.

[0020] Optionally, S2 specifically includes:

[0021] The railway line and inspection task are obtained from the inspection parameter set, and the continuous interval of the line mileage is divided into sections to generate a section sequence.

[0022] The drone's location, flight attitude, battery status, and communication status are obtained from the inspection parameter set. The spatial and temporal correspondence between the drone and the inspection section is calculated, and a matching sequence is generated.

[0023] The drone number, inspection section number, and corresponding number of times in the matching sequence are counted. The drone number is used as the matrix row index, the inspection section number is used as the matrix column index, and the corresponding number of times is written into the corresponding matrix position to generate a task matrix.

[0024] Traverse the matrix elements in the task matrix, extract the UAV number and inspection section number corresponding to the matrix element, arrange the inspection sections and inspection order corresponding to the UAV, and generate an inspection plan.

[0025] Optionally, generating the initial route based on the inspection plan specifically involves:

[0026] The drone number, inspection section, and inspection sequence are obtained from the inspection plan to generate flight path arrangement data.

[0027] The inspection parameters are used to obtain the line mileage, track centerline coordinates and trackside component locations corresponding to the inspection section, and the section spatial data is generated.

[0028] The route arrangement data is matched with the segment spatial data to determine the start inspection position, end inspection position and inspection direction of the UAV, and generate route endpoint data.

[0029] The initial route is generated by connecting the starting inspection position, trackside component position, and ending inspection position in the route endpoint data according to the inspection direction.

[0030] Optionally, the step of mapping the initial route to the airspace grid along the railway line, performing route conflict detection and segment coordination, and generating candidate routes specifically involves:

[0031] The coordinates of the track centerline corresponding to the railway line are obtained from the inspection parameter set, and a spatial grid along the railway line is generated along the track centerline coordinates.

[0032] The initial route is decomposed into multiple segments, and the spatial location corresponding to each segment is written into the airspace grid along the railway to generate route grid data.

[0033] The UAV number, flight segment number, and time stamp corresponding to the same airspace grid in the route grid data are counted to generate flight segment occupancy data;

[0034] Conflict marking is performed on flight segments with the same airspace grid and the same time marker in the flight segment occupancy data to generate conflict flight segment data;

[0035] The conflicting flight segment data is subjected to altitude layer separation and time staggering coordination to generate candidate routes.

[0036] Optionally, the step of extracting the visible area of ​​trackside components based on railway line data and identifying occluded sections and deviation sections specifically includes:

[0037] The location of trackside components, coordinates of the track centerline, and track mileage are obtained from railway line data to generate a component location sequence; the spatial location, direction, and time marker of flight segments are obtained from candidate flight routes to generate a flight segment observation sequence.

[0038] The component location sequence is spatially mapped to the flight segment observation sequence to generate component observation data;

[0039] Based on the observation data of the components, the observation azimuth and observation distance of the trackside components relative to the candidate route are calculated, and the visible field of view of the trackside components is generated.

[0040] Spatial comparison is performed between the visible field of trackside components and the observation sequence of flight segments to generate visible field comparison data.

[0041] Extract the line mileage intervals where there is an occlusion relationship between the trackside components and the spatial location of the flight segment from the visual field comparison data, and generate occlusion sections.

[0042] Extract the line mileage intervals where there is a deviation between the observation direction of the trackside components and the flight segment direction from the visual field comparison data, and generate the deviation section.

[0043] Optionally, the step of performing perspective compensation and segment rearrangement on candidate routes to generate cooperative routes specifically involves:

[0044] Based on the obstructed sections, deviation sections, and candidate routes, route correction data is generated. Lateral observation segments are added to the obstructed sections in the route correction data to generate obstruction compensation routes.

[0045] Adjust the direction of the flight segment in the deviation section of the occlusion compensation route to generate a view compensation route;

[0046] The segments in the perspective compensation route are sorted according to the route mileage order to generate a rearranged segment sequence;

[0047] Perform segment connection and time stamp updates on the rearranged segment sequence to generate cooperative routes.

[0048] Optionally, the improved Mask DINO model specifically includes a feature encoding module, a mask query module, a decoding prediction module, and a segmentation output module;

[0049] The feature encoding module converts the industrial visual data along the railway line in the inspection visual stream into an image frame sequence, and extracts multi-scale image features from the image frame sequence to generate inspection image features.

[0050] The mask query module receives inspection image features and introduces a track structure guidance mechanism. This mechanism extracts rail extension boundaries, sleeper arrangement boundaries, catenary suspension boundaries, bridge edge boundaries, tunnel entrance ring boundaries, and slope toe boundaries from the inspection image features, and arranges them according to line mileage to generate a track structure boundary sequence. Boundary points belonging to the same railway component in the track structure boundary sequence are connected according to their adjacent positions to generate rail extension zones, sleeper arrangement zones, catenary suspension zones, bridge edge zones, tunnel entrance boundary zones, and slope toe zones. The rail extension zone, sleeper arrangement zone, catenary suspension zone, bridge edge zone, tunnel entrance boundary zone, and slope toe zone are designated as the main track structure domains. The main domain center position, main domain extension direction, main domain boundary span, and main domain category identifier are extracted from the main track structure domains. These are then encoded into domain kernel vectors. The domain kernel vectors are written into the query vector set of the mask query module, and the general query vectors in the query vector set are replaced with the query vectors corresponding to the main track structure domains to generate track structure query vectors.

[0051] The decoding and prediction module performs attention interaction between the track structure query vector and the inspection image features to generate target query features, and generates component segmentation results and defect segmentation results based on the target query features;

[0052] The segmentation output module performs category labeling, contour closure, and pixel merging on the component segmentation results and defect segmentation results to generate the target segmentation result.

[0053] Optionally, the improved SuperGlue graph matching algorithm is specifically as follows:

[0054] The defect segmentation results in the target segmentation results are correlated with the image frames, time markers and line mileage in the inspection visual flow to generate defect observation data;

[0055] Defect feature data is generated by extracting defect contours, defect categories, route mileage locations, image frame locations, and UAV numbers from the defect observation data.

[0056] Defect map nodes are constructed by using the defect contour in the defect feature data as node boundaries, the defect category as node attributes, and the line mileage location and image frame location as node coordinates.

[0057] The defect map nodes corresponding to different UAV numbers are paired, and the mileage difference, category consistency value, contour overlap value and image frame position difference between the paired defect map nodes are calculated to generate node pairing features.

[0058] In the improved SuperGlue graph matching algorithm, a trace domain verification mechanism is introduced. This mechanism writes the node pairing features into the edge features during the graph attention matching process and updates the node descriptors of the defect graph nodes based on the edge features, generating cross-view matching features. Based on the cross-view matching features, the matching confidence between defect graph nodes under different UAV perspectives is calculated to generate a defect matching matrix. From the defect matching matrix, the defect graph node pairing results corresponding to the same route mileage location and the same defect category are extracted to generate homogeneous defect data. The duplicate defect regions in the homogeneous defect data are merged, and the position of the drifting defect regions in the homogeneous defect data is corrected to generate multi-UAV defect results.

[0059] Optionally, S8 specifically includes:

[0060] Defect contours, image frame positions, UAV numbers, and route mileage positions are extracted from the multi-machine defect results to generate defect projection data.

[0061] The image frame positions in the defect projection data are transformed to the UAV imaging coordinate system to generate defect imaging coordinates; the defect imaging coordinates are then back-projected to the railway line spatial coordinate system to generate defect spatial coordinates.

[0062] The defect spatial coordinates are mapped to the line mileage location to generate the defect location result;

[0063] Based on the defect category, defect outline range, and line mileage location in the defect location results, the defect level is classified and a defect level result is generated.

[0064] An inspection and handling instruction is generated based on the defect level result and the defect location result.

[0065] The beneficial effects of this invention are:

[0066] (1) By jointly constructing the inspection parameter set, task matrix, airspace grid along the railway line and collaborative route, in order to address the problems of large span of inspection area along the railway line, crossover of UAV routes and chaotic inspection time sequence, route conflict detection, altitude layer separation and time staggered coordination are adopted to reduce the airspace conflict rate in the collaborative inspection process of multiple UAVs and improve the continuity and coordination of inspection along the railway line.

[0067] (2) In the improved Mask DINO model, a track structure query mechanism is introduced. In response to the problem of strong continuity of the boundaries of rails, sleepers and catenary areas along the railway and complex background structure, the track structure boundary sequence, track structure main domain and domain kernel vector are written into the query generation process to improve the stability of component segmentation, defect boundary identification capability and cross-view target segmentation consistency under complex background.

[0068] (3) In the improved SuperGlue graph matching algorithm, a trace domain verification mechanism is introduced. In response to the problems of repeated defect identification, drift defect location and cross-view matching error in the inspection process of multiple UAVs, the mileage difference, category consistency value, contour overlap value and image frame position difference are written into the graph attention matching process to improve the matching accuracy of multi-UAV defect results, defect location accuracy and inspection handling results reliability. Attached Figure Description

[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0070] Figure 1 This is a flowchart of an intelligent collaborative inspection method for railway drones based on industrial vision proposed in this invention;

[0071] Figure 2 This is a schematic diagram of the improved Mask DINO model proposed in this invention;

[0072] Figure 3 This is a data flow diagram of an intelligent collaborative inspection method for railway drones based on industrial vision proposed in this invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0074] refer to Figures 1-3 A method for intelligent collaborative inspection of railways using unmanned aerial vehicles (UAVs) based on industrial vision includes the following steps:

[0075] S1. Acquire railway inspection task data, railway line data, and drone operation data; perform time correlation and line mileage correlation to generate inspection parameter set.

[0076] S2. Divide the inspection sections based on the inspection parameter set, match the UAVs with the inspection sections, construct the task matrix, and generate the inspection plan.

[0077] S3. Generate an initial route based on the inspection plan, map the initial route to the airspace grid along the railway, perform route conflict detection and segment coordination, and generate candidate routes;

[0078] S4. Extract the visible area of ​​trackside components based on railway line data, identify occluded and deviated sections, perform view compensation and segment rearrangement on candidate routes, and generate cooperative routes.

[0079] S5. Control the drone to acquire industrial visual data along the railway line according to the collaborative flight path and generate an inspection visual flow.

[0080] S6. Input the inspection visual flow into the improved Mask DINO model, introduce the track structure query mechanism in the mask query generation module, extract the track structure boundary sequence, construct the track structure main domain, and output the target segmentation result.

[0081] S7. Based on the target segmentation results and the inspection visual flow, an improved SuperGlue graph matching algorithm is adopted, and a trace domain verification mechanism is introduced to generate multi-machine defect results;

[0082] S8. Perform spatial back projection and line mileage mapping on the multi-machine defect results to generate defect location results, and perform defect level classification based on the defect location results to generate inspection and handling instructions.

[0083] In this embodiment, S1 specifically refers to:

[0084] Railway inspection task data is collected from railway inspection plans, emergency inspection instructions and inspection work orders; railway line data is collected from railway line basic data and trackside component data; and UAV operation data is collected from UAV flight control records and load status records to generate raw inspection data.

[0085] The inspection task number, inspection start and end mileage, inspection object, inspection time period, line mileage, trackside component location, track centerline coordinates, UAV number, flight position, flight attitude, power status, load status and communication status in the original inspection data are converted into a unified field format, duplicate fields are deleted, and data records corresponding to the same inspection task are merged to generate standard inspection data.

[0086] Establish a unified time stamp based on the data collection time and task start and end time in the inspection standard data, and classify the inspection tasks, railway lines and drone status under the same unified time stamp into the same time sequence record to generate time sequence inspection data.

[0087] Based on the line mileage in the railway line data, the inspection start and end mileage, trackside component location, track centerline coordinates, UAV flight position and UAV flight attitude in the time-series inspection data are mapped to the same line mileage interval to generate mileage inspection data.

[0088] The inspection task number, inspection object, line mileage section, trackside component location, UAV number, flight position, flight attitude, power status, load status, and communication status from the mileage inspection data are written into the same data record to generate an inspection parameter set.

[0089] In this embodiment, S2 specifically refers to:

[0090] Extract the line mileage interval, inspection object, inspection time period and trackside component location from the inspection parameter set. Group the data records with continuous start and end mileage and the same inspection object in the line mileage interval into the same inspection section. Arrange the inspection sections in ascending order of line mileage to generate a section sequence.

[0091] Assign an inspection section number to each inspection section in the section sequence, and record the starting mileage, ending mileage, inspection object, inspection time period and trackside component location corresponding to the inspection section to generate a numbered section sequence;

[0092] Extract the UAV number, flight position, flight attitude, battery status and communication status from the inspection parameter set, map the UAV flight position to the line mileage interval in the numbered segment sequence, and generate UAV segment corresponding data;

[0093] The data corresponding to the drone segment is aligned with the numbered segment sequence according to the inspection period, and the inspection segment number corresponding to the drone within the inspection period is retained to generate time-series data.

[0094] Write the UAV number, inspection section number, flight position, flight attitude, battery status and communication status from the time-series corresponding data into the same data record to generate a matching sequence;

[0095] The number of data records in the matching sequence that simultaneously contain the same UAV number and the same inspection section number is counted to obtain the corresponding number of occurrences.

[0096] The drone number is used as the row index of the task matrix, the inspection section number is used as the column index of the task matrix, and the corresponding number of times is written into the intersection of the corresponding row index and the corresponding column index to generate the task matrix.

[0097] The inspection segments are arranged according to the inspection segment number corresponding to the same UAV number in the task matrix, and the inspection order is determined according to the increasing order of the starting mileage of the inspection segments, generating an inspection plan that includes the UAV number, inspection segment number and inspection order.

[0098] In this embodiment, an initial route is generated based on the inspection plan, specifically as follows:

[0099] Extract the UAV number, inspection section number and inspection sequence from the inspection plan, arrange the inspection section numbers corresponding to the same UAV number according to the inspection sequence, and generate flight path arrangement data.

[0100] Extract the starting mileage, ending mileage, track centerline coordinates, and trackside component locations corresponding to the inspection section number from the inspection parameter set. Arrange the track centerline coordinates between the starting mileage and ending mileage in ascending order of line mileage to generate section spatial data.

[0101] By matching the inspection section number in the route arrangement data with the inspection section number in the section space data, the starting mileage, ending mileage, track centerline coordinates, and trackside component positions corresponding to the UAV number are determined, and route matching data is generated.

[0102] In the route matching data, the track centerline coordinates corresponding to the starting mileage are used as the starting inspection position, the track centerline coordinates corresponding to the ending mileage are used as the ending inspection position, and the direction from the starting inspection position to the ending inspection position is used as the inspection direction to generate route endpoint data.

[0103] Arrange the starting inspection position, trackside component position, and ending inspection position in the route endpoint data according to the inspection direction. Take the starting inspection position as the route start point, the ending inspection position as the route end point, and the trackside component position located between the starting inspection position and the ending inspection position as the route midpoint to generate a route point sequence. Connect the points of adjacent positions in the route point sequence in sequence to generate the initial route.

[0104] In this implementation, the initial route is mapped to the airspace grid along the railway line, route conflict detection and segment coordination are performed, and candidate routes are generated, specifically as follows:

[0105] The track centerline coordinates and line mileage are extracted from the inspection parameter set. The track centerline coordinates are arranged in ascending order of line mileage to generate a line coordinate sequence.

[0106] Vertical grids are divided along the extension direction of the line coordinate sequence, horizontal grids are divided along the lateral direction of the line coordinate sequence, and altitude grids are divided along the flight altitude direction of the UAV. The vertical grids, horizontal grids, and altitude grids are combined into an airspace grid along the railway line.

[0107] The lines connecting adjacent route points in the initial route are taken as flight segments. Each flight segment is marked with the UAV number, flight segment number, starting point coordinates, ending point coordinates, flight altitude and time stamp to generate flight segment data.

[0108] The starting point coordinates, ending point coordinates, and flight altitude in the flight segment data are mapped to the airspace grid along the railway line. The airspace grid number occupied by each flight segment is recorded to generate flight route grid data.

[0109] In the statistical route grid data, the UAV number and flight segment number corresponding to the same airspace grid number and the same time mark are used to generate flight segment occupancy data;

[0110] Mark multiple segments with the same airspace grid number and the same time stamp in the segment occupancy data as conflicting segments to generate conflicting segment data;

[0111] Different flight altitudes are assigned to multiple flight segments in the conflict flight segment data, the flight altitudes corresponding to the conflict flight segments are updated, and altitude coordination data is generated.

[0112] For segments in the highly coordinated data that still have the same airspace grid number and the same time stamp, change the time stamp to generate time coordinated data; update the segment flight altitude and time stamp in the initial route according to the time coordinated data to generate candidate routes.

[0113] In this embodiment, the visible area of ​​trackside components is extracted based on railway line data, and occluded sections and deviation sections are identified, specifically as follows:

[0114] Extract the location of trackside components, the coordinates of the track centerline, and the mileage of the railway line data. Arrange the locations of trackside components in ascending order of mileage to generate a component location sequence.

[0115] Extract the spatial location, direction, altitude and time markers of the flight segments from the candidate routes, arrange the spatial locations of the flight segments in chronological order of the time markers, and generate the flight segment observation sequence.

[0116] Spatially pair the trackside component locations in the component location sequence with the segment spatial locations in the segment observation sequence, record the trackside component locations, segment spatial locations, segment directions, flight altitudes, timestamps, and route mileage, and generate component observation data;

[0117] Extract the trackside component location and segment spatial location corresponding to the same line mileage from the component observation data, take the direction from the segment spatial location to the trackside component location as the observation azimuth, and take the spatial distance between the segment spatial location and the trackside component location as the observation distance to generate the trackside component visible field.

[0118] Spatial comparison is performed between the observation azimuth, observation distance and line mileage in the visible field of trackside components and the segment direction, segment spatial location and time marker in the segment observation sequence to generate visible field comparison data.

[0119] In the visual field comparison data, the line connecting the location of the trackside component and the spatial location of the flight segment is used as the observation ray, and the bridge edge, tunnel entrance boundary, slope outline and trackside component outline in the railway line data are used as the occlusion boundary; if the observation ray intersects with the occlusion boundary, the line mileage interval corresponding to the intersection position is marked as the occlusion section.

[0120] In the visual field comparison data, the directional difference between the flight segment direction and the observation azimuth is taken as the observation directional deviation; if the flight segment direction corresponding to the observation directional deviation deviates from the position of the trackside component, the line mileage interval corresponding to the deviation position is marked as the deviation section.

[0121] In this implementation, perspective compensation and segment rearrangement are performed on candidate routes to generate cooperative routes, specifically as follows:

[0122] Write the corresponding segment number, segment spatial location, segment direction, flight altitude, and time stamp from the obscured segment, deviation segment, and candidate route into the same data record to generate route correction data;

[0123] For the obstructed sections in the route correction data, select the location of the trackside component corresponding to the obstructed section and the spatial location of the adjacent segments in the candidate route, set lateral observation points in the transverse direction along the track centerline, and connect the lateral observation points with the spatial locations of the adjacent segments to generate lateral observation segments.

[0124] Insert the lateral observation segments into the candidate route positions corresponding to the obscured sections, and retain the segments in the candidate routes that do not correspond to the obscured sections to generate obscuration compensation routes.

[0125] For the deviation section in the occlusion compensation route, the position of the trackside component and the spatial position of the segment corresponding to the deviation section are selected, and the direction from the spatial position of the segment to the position of the trackside component is used as the corrected segment direction.

[0126] Replace the segment direction of the segment corresponding to the deviation segment with the corrected segment direction, and retain the segment direction of the segment that does not correspond to the deviation segment in the occlusion compensation route to generate the view compensation route.

[0127] The segments are arranged in ascending order of the route mileage corresponding to the segments in the perspective compensation route. If the same UAV corresponds to multiple inspection sections, the segments in the inspection section are arranged in the inspection order in the inspection plan to generate a rearranged segment sequence.

[0128] Connect the end point coordinates and start point coordinates of adjacent segments in the rearranged segment sequence to form continuous segments, and assign time stamps to the continuous segments according to the arrangement order of the rearranged segment sequence to generate cooperative routes.

[0129] In this embodiment, the improved Mask DINO model specifically includes a feature encoding module, a mask query module, a decoding prediction module, and a segmentation output module;

[0130] The feature encoding module receives the inspection visual stream, arranges the railway-side industrial visual data in the inspection visual stream into an image frame sequence according to time stamps, scales the image frames in the image frame sequence to the same image size, and forms three-channel image data according to the red, green and blue channels; extracts texture features, edge features and semantic features at different scales from the three-channel image data, and arranges the texture features, edge features and semantic features at different scales according to the image coordinate correspondence to generate inspection image features;

[0131] The query generation module receives inspection image features and introduces a track structure guidance mechanism. This mechanism extracts rail extension boundaries, sleeper arrangement boundaries, catenary suspension boundaries, bridge edge boundaries, tunnel entrance ring boundaries, and slope toe boundaries from the inspection image features. It arranges the boundary points corresponding to the same railway component according to their image coordinate adjacency and writes the corresponding line mileage into the boundary point record, generating a track structure boundary sequence. In this sequence, adjacent boundary points corresponding to the same railway component are connected sequentially to form component boundary lines. These boundary lines are then expanded into strip-shaped areas along their extension direction, resulting in rail extension strips, sleeper arrangement strips, catenary suspension strips, bridge edge strips, tunnel entrance boundary strips, and slope toe strips. The overhead contact line suspension zone, bridge edge zone, tunnel entrance boundary zone, and slope toe zone are arranged in order of line mileage to generate the main track structure domain. The main domain's center position, extension direction, boundary span, and category identifier are extracted from the main track structure domain. The main domain's center position is converted into image coordinate encoding, the extension direction into direction encoding, the boundary span into scale encoding, and the category identifier into category encoding. The image coordinate encoding, direction encoding, scale encoding, and category encoding are combined in a fixed order to generate a domain kernel vector. The domain kernel vector is written into the query vector set of the query generation module. General query vectors in the query vector set that correspond to the main track structure domain are replaced with the domain kernel vector. General query vectors that do not correspond to the main track structure domain are retained, generating a track structure query vector.

[0132] The decoding and prediction module receives the track structure query vector and the inspection image features. It uses the track structure query vector as the target query and the inspection image features as the image feature keys to establish an attention correspondence between the target query and the image feature keys, generating target query features. The target query features are then mapped to component category results, defect category results, component mask results, and defect mask results, respectively. The component category results and component mask results are combined to form a component segmentation result, and the defect category results and defect mask results are combined to form a defect segmentation result.

[0133] The segmentation output module organizes the component segmentation results and defect segmentation results according to the line mileage, image frame position and category identifier. It connects adjacent pixels in the broken pixel regions in the component mask results and defect mask results, and merges overlapping pixel regions according to the category identifier to generate the target segmentation result.

[0134] In this embodiment, both the improved Mask DINO model and the standard Mask DINO model employ feature encoding, query generation, decoding prediction, and segmentation output structures. Both generate target segmentation results through the attention correspondence between image features and the target query, and both possess instance segmentation and defect segmentation capabilities. The improved Mask DINO model introduces a track structure guidance mechanism in its query generation module, incorporating the rail extension boundaries, sleeper arrangement boundaries, catenary suspension boundaries, bridge edge boundaries, tunnel entrance ring boundaries, and slope toe boundaries of the railway line into the query generation process. This integrates the spatial extension relationships, line mileage relationships, and boundary continuity relationships of railway line components into the target query generation process. The track structure guidance mechanism connects boundary points to form rail extension zones, sleeper arrangement zones, catenary suspension zones, bridge edge zones, tunnel entrance boundary zones, and slope toe zones, establishing a correspondence between the target query vector and the spatial distribution structure of railway components. The improved Mask... The DINO model constructs a domain kernel vector by defining the main domain center location, main domain extension direction, main domain boundary span, and main domain category identifier. This domain kernel vector replaces the generic query vector in the query vector set, ensuring that the target query vector includes railway component boundary structure information and track spatial location information. The improved Mask DINO model incorporates track mileage, component boundaries, and image coordinates into the track structure boundary sequence, preserving both the continuous features of the railway line and the spatial adjacency features of components during query generation. Furthermore, the improved Mask DINO model constrains the target query range through the track structure main domain, prioritizing component and defect boundaries in rail, sleeper, catenary, bridge, tunnel entrance, and slope regions, reducing interference from background areas. Finally, the improved Mask DINO model enhances feature association capabilities in continuous regions of railway components by extending component boundary lines into strip-shaped areas, improving boundary recognition capabilities in long rail regions, repeating sleeper regions, and catenary suspension regions. The Mask DINO model further enhances these capabilities by extending component boundary lines into strip-shaped regions. The DINO model enhances the spatial correspondence between target query features and railway components by establishing an attention correspondence between track structure query vectors and inspection image features, thereby improving the stability of defect segmentation in occluded areas, complex lighting areas, and long-distance inspection areas. The improved Mask DINO model enhances the continuous association of railway components between different image frames by arranging the main track structure domain in the order of line mileage, thereby improving the consistency of cross-view segmentation and the accuracy of defect area localization in multi-UAV inspection scenarios.

[0135] In this embodiment, the SuperGlue graph matching algorithm is improved as follows:

[0136] The defect segmentation results in the target segmentation results are mapped to the image frames in the inspection visual flow according to the image frame positions, and the time stamp and line mileage corresponding to the defect segmentation results are written into the same data record to generate defect observation data.

[0137] Defect contours, defect categories, line mileage locations, image frame locations, and UAV numbers are extracted from defect observation data. The defect contours are arranged in the order of contour points, the line mileage locations are recorded according to the line mileage values, and the image frame locations are recorded according to the image coordinates to generate defect feature data.

[0138] Defect map nodes are generated by using the defect contour in the defect feature data as node boundaries, the defect category as node attributes, the line mileage location and image frame location as node coordinates, and the UAV number as node source. Defect map nodes corresponding to different UAV numbers are paired up, and the line mileage location, defect category, defect contour and image frame location corresponding to the paired defect map nodes are extracted to generate node pairing data.

[0139] The difference between two line mileage locations in the node pairing data is taken as the mileage difference. Two defects of the same category are considered to be of the same category, and two defects of different categories are considered to be of different categories. The ratio of the overlapping area to the union area of ​​the two defect contours in the image coordinates is taken as the contour overlap value. The coordinate distance between two image frame locations is taken as the image frame location difference. Node pairing features are generated.

[0140] The trace domain verification mechanism incorporates mileage difference, category consistency value, contour overlap value, and image frame position difference into the edge features of the graph attention matching process. This allows the edge features to represent the correspondences between defect graph nodes from different UAV perspectives, including route mileage correspondence, defect category correspondence, defect contour correspondence, and image position correspondence. During graph attention matching, the node attributes and coordinates of the defect graph nodes are used as node descriptors. The edge features are added to the message passing calculation of the node descriptors, enabling each defect graph node to receive the mileage difference, category consistency value, contour overlap value, and image frame position difference from paired defect graph nodes, update its node descriptor, and generate cross-view matching features. These cross-view matching features are then input into the matching layer to obtain the matching confidence between defect graph nodes from different UAV perspectives. The matching confidence is then calculated according to the defect graph nodes... Points are numbered and arranged to generate a defect matching matrix. Defect map nodes corresponding to the same line mileage location and defect category are selected from the defect matching matrix for pairing. The defect contour, image frame position, UAV number, and matching confidence score corresponding to the defect map node pairing results are written into the same data record to generate homogeneous defect data. For duplicate defect areas in the homogeneous defect data, defect contours corresponding to the same line mileage location and the same defect category are merged into a joint defect contour to generate a merged defect area. For drifting defect areas in the homogeneous defect data, the defect map node positions corresponding to different UAV numbers are merged according to the matching confidence score to obtain the corrected defect position. The merged defect area, corrected defect position, defect category, line mileage location, and UAV number are written into the same data record to generate multi-UAV defect results.

[0141] In this embodiment, the improved SuperGlue graph matching algorithm maps defect segmentation results to image frames, time markers, and line mileage, enabling defect observation data to simultaneously possess spatial and temporal location information, thus improving data consistency among inspection results from different UAVs. By constructing defect graph nodes from defect contours, defect categories, line mileage locations, and image frame locations, these nodes possess both contour structure features and line spatial features, enhancing node representation capabilities in cross-view defect matching. Furthermore, by constructing mileage difference, category consistency value, contour overlap value, and image frame location difference, node pairing features simultaneously characterize line correspondence, defect category correspondence, and contour correspondence, improving the correlation between defect regions from different UAV perspectives. Simultaneously, a trace domain verification mechanism further enhances node pairing. Features are written into the edge features of the graph attention matching process, enabling node descriptors to integrate the correlation information between line mileage, defect contours, and image locations during message passing, thus improving the stability of cross-view matching in complex railway scenarios. A defect matching matrix is ​​generated by matching confidence, enabling defect graph nodes from different UAV perspectives to form a unified correspondence, improving the ability to identify repetitive defect areas. At the same time, merged defect areas are generated by jointly identifying defect contours, reducing redundant results caused by multiple UAVs repeatedly identifying the same defect. Furthermore, the location of drifting defect areas is merged by matching confidence, improving the accuracy of defect location when there is a shift in line mileage. In addition, the defect results from multiple UAVs are fused with defect contours and defect locations from different UAV perspectives, improving the completeness of defect identification along the railway line and the consistency of inspection results.

[0142] In this embodiment, S8 specifically refers to:

[0143] Extract defect contours, defect categories, image frame positions, UAV numbers, route mileage positions, and corrected defect positions from the multi-aircraft defect results. Arrange the defect contours in the order of contour points, record the image frame positions according to image coordinates, and record the route mileage positions according to route mileage values ​​to generate defect projection data.

[0144] The image frame positions in the defect projection data are converted into pixel coordinates in the UAV imaging coordinate system. The pixel coordinates and the contour point coordinates corresponding to the defect contour are written into the same data record to generate defect imaging coordinates.

[0145] Extract the flight position, flight attitude, camera focal length, principal point coordinates and pixel size corresponding to the UAV number from the inspection visual flow, convert the pixel coordinates in the defect imaging coordinates into the camera light direction, and generate the defect projection ray;

[0146] The defect projection ray is spatially matched with the track centerline coordinates and trackside component positions in the railway line spatial coordinate system, and the intersection position is used as the defect spatial coordinates. The defect spatial coordinates are then matched with the line mileage position in the defect projection data. The defect type, defect outline, UAV number, defect spatial coordinates, and line mileage position are written into the same data record to generate the defect location result.

[0147] Extract the defect category, defect outline range, line mileage location and defect spatial coordinates from the defect location results, convert the pixel area corresponding to the defect outline range into a spatial range in the railway line spatial coordinate system, and generate defect range data.

[0148] The defect location results are classified into levels according to the defect type, defect range data, and line mileage location to generate defect level results; the defect level results, defect location results, defect type, line mileage location, and defect spatial coordinates are written into the inspection and handling data to generate inspection and handling instructions.

[0149] Example 1: To verify the feasibility of this invention in practice, it was applied to a UAV-assisted inspection task on a mountainous conventional railway line and an adjacent high-speed railway connecting line of a railway maintenance section. The total length of the test section was 42.6 km, including 7 bridges, 12 tunnel entrances, 18 high slopes, 326 catenary supports, and approximately 4100 key trackside components. The test site presented challenges such as mountain obstruction, alternating bridges and tunnels, dense trackside components, significant changes in UAV perspective, and concentrated inspection tasks. Traditional single-drone inspection methods are prone to issues such as repeated flight paths, missed component images, incomplete defect boundary identification, and duplicate alarms from multiple perspectives. For comparison, a manually assisted single-drone inspection method was used as the control group, and this invention was used as the test group. The inspection targets included rail surface anomalies, sleeper damage, catenary suspension anomalies, bridge edge damage, tunnel entrance spalling, and slope slippage areas.

[0150] Six drones were deployed on-site, each equipped with a visible light industrial camera and a positioning module, with an image resolution of 3840×2160, a frame rate of 25fps, and an inspection altitude set from 35m to 70m. The system first acquires railway inspection task data, railway line data, and drone operation data to generate an inspection parameter set. Then, it divides the line into continuous sections based on mileage, matching drones to these sections to form an inspection plan. After the inspection plan is generated, the system maps the initial flight path to a grid of airspace along the railway line, marking conflicts between flight segments within the same airspace grid and at the same time, and generating candidate flight paths through altitude layer separation and time-staggered coordination. To address occlusion issues at bridge edges, tunnel entrances, and slope toes, the system extracts the visible area of ​​trackside components, identifies occluded and deviated sections, adds lateral observation segments to candidate flight paths, and adjusts the direction of these segments to generate collaborative flight paths. After the UAVs acquire industrial visual data along the railway line according to the collaborative flight path, the improved Mask DINO model extracts the track structure boundary sequence through the track structure query mechanism, constructs the track structure main domain, and outputs the target segmentation results; the improved SuperGlue graph matching algorithm matches the defect graph nodes under different UAV perspectives through the track domain verification mechanism to generate multi-UAV defect results; finally, spatial back projection and line mileage mapping are performed on the multi-UAV defect results to output inspection and handling instructions.

[0151] Table 1. Comparison of Efficiency and Quality of Collaborative Inspection of Routes

[0152] Inspection route length 42.6km 42.6km Consistent Time to complete inspection 6.8h 3.1h Shortened by 54.4% Average effective operating time of drones 29.5 min / flight 38.2 min / flight Increased by 29.5% Overlapping mileage coverage on routes 7.4km 1.6km Reduced by 78.4% Number of route conflict alerts 18 times 3 times Reduced by 83.3% Trackside component integrity observation rate 86.7% 96.8% Increased by 10.1% Points missed by the bridge and tunnel slopes 31 places 6 places Reduced by 80.6% Average generation time of inspection and handling instructions 24.6min 8.9min Shortened by 63.8%

[0153] As shown in Table 1, under the same line length and the same inspection objects, this invention transforms the UAV inspection task from simple segmented flight to collaborative inspection based on line mileage and airspace grid by continuously generating inspection parameter sets, task matrices, candidate routes, and collaborative routes. The manual-assisted single-aircraft inspection method requires 6.8 hours to complete the inspection of 42.6 km of line, mainly due to limited single-aircraft endurance, numerous repeated flights, and the need for return flights to retake photos in bridge and tunnel slope areas. This invention completes the same task in only 3.1 hours, reducing the repeated coverage mileage of the route from 7.4 km to 1.6 km. The number of route conflict alarms decreased from 18 to 3, indicating that the airspace grid along the railway line, altitude layer separation, and time-staggered coordination can effectively reduce the risk of overlapping flight segments by multiple aircraft. The complete observation rate of trackside components increased to 96.8%, indicating that the introduction of visible areas, obstructed sections, and deviated sections of trackside components can improve the observation quality of bridge and tunnel slope areas and reduce data loss at the defect identification front end.

[0154] Table 2 Comparison of Industrial Visual Defect Recognition and Localization Results

[0155] Actual number of defects marked 126 places 126 places Consistent Correctly identify the number of defects 106 places 119 places 13 new locations added Defect identification accuracy 84.1% 94.4% Increased by 10.3% Number of defects missed 20 places 7 places Reduced by 65.0% Number of repeat defect alarms 27 articles 5 items Reduced by 81.5% Average overlap rate of defect profiles 76.3% 89.6% Increased by 13.3% Average error of line mileage positioning 2.85m 0.92m Reduced by 67.7% Defect level classification consistency rate 82.5% 93.7% Increased by 11.2%

[0156] As shown in Table 2, this invention demonstrates higher stability in defect identification, defect contouring, and track mileage localization. While conventional target detection and image stitching methods can identify large foreign objects and obvious damage, they are easily affected by long-distance viewing angles, repetitive textures, and occluded backgrounds when dealing with small rail cracks, abnormal overhead contact line suspension, tunnel entrance debris, and slope slippage areas, leading to numerous missed detections and duplicate alarms. This invention improves the track structure query mechanism in the Mask DINO model by incorporating the main track structure domains, such as rail extensions, sleeper arrangement zones, and overhead contact line suspension zones, into the query generation process. This strengthens the correspondence between the target segmentation results and railway component boundaries, increasing the average overlap rate of defect contours to 89.6%. Simultaneously, the improved SuperGlue graph matching algorithm incorporates mileage difference, category consistency value, contour overlap value, and image frame position difference into the graph attention matching process through a trace domain verification mechanism. This reduces the number of duplicate defect alarms from 27 to 5, and the average error in track mileage localization from 2.85m to 0.92m, indicating that multi-machine defect results are more suitable for directly generating inspection and handling instructions.

[0157] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A railway UAV intelligent collaborative inspection method based on industrial vision, characterized in that, Includes the following steps: S1. Acquire railway inspection task data, railway line data, and drone operation data; perform time correlation and line mileage correlation to generate inspection parameter set. S2. Divide the inspection sections based on the inspection parameter set, match the UAV with the inspection sections, construct a task matrix, and generate an inspection plan. S3. Generate an initial route based on the inspection plan, map the initial route to the airspace grid along the railway, perform route conflict detection and segment coordination, and generate candidate routes; S4. Extract the visible area of ​​trackside components based on railway line data, identify occluded and deviated sections, perform view compensation and segment rearrangement on candidate routes, and generate cooperative routes. S5. Control the drone to acquire industrial visual data along the railway line according to the collaborative flight path and generate an inspection visual flow. S6. Input the inspection visual flow into the improved Mask DINO model, introduce a track structure query mechanism in the mask query generation module, extract the track structure boundary sequence, construct the track structure main domain, and output the target segmentation result. S7. Based on the target segmentation results and the inspection visual flow, an improved SuperGlue graph matching algorithm is adopted, and a trace domain verification mechanism is introduced to generate multi-machine defect results; S8. Perform spatial back projection and line mileage mapping on the multi-machine defect results to generate defect location results, and perform defect level classification based on the defect location results to generate inspection and handling instructions.

2. The intelligent collaborative inspection method for railway drones based on industrial vision according to claim 1, characterized in that, Specifically, S1 is: Collect railway inspection task data, railway line data, and drone operation data to generate raw inspection data; The raw inspection data is standardized in data format and merged into fields to generate standard inspection data. A unified time stamp is added to the standard inspection data to generate time-series inspection data; Based on the line mileage in the railway line data, mileage correlation is performed on the time-series inspection data to generate mileage inspection data; The inspection tasks, railway lines, and drone status in the mileage inspection data are matched to generate an inspection parameter set.

3. The intelligent collaborative inspection method for railway drones based on industrial vision according to claim 1, characterized in that, Specifically, S2 is: The railway line and inspection task are obtained from the inspection parameter set, and the continuous interval of the line mileage is divided into sections to generate a section sequence. The drone's location, flight attitude, battery status, and communication status are obtained from the inspection parameter set. The spatial and temporal correspondence between the drone and the inspection section is calculated, and a matching sequence is generated. The drone number, inspection section number, and corresponding number of times in the matching sequence are counted. The drone number is used as the matrix row index, the inspection section number is used as the matrix column index, and the corresponding number of times is written into the corresponding matrix position to generate a task matrix. Traverse the matrix elements in the task matrix, extract the UAV number and inspection section number corresponding to the matrix element, arrange the inspection sections and inspection order corresponding to the UAV, and generate an inspection plan.

4. The intelligent collaborative inspection method for railway drones based on industrial vision according to claim 1, characterized in that, The initial route is generated based on the inspection plan, specifically as follows: The drone number, inspection section, and inspection sequence are obtained from the inspection plan to generate flight path arrangement data. The inspection parameters are used to obtain the line mileage, track centerline coordinates and trackside component locations corresponding to the inspection section, and the section spatial data is generated. The route arrangement data is matched with the segment spatial data to determine the start inspection position, end inspection position and inspection direction of the UAV, and generate route endpoint data. The initial route is generated by connecting the starting inspection position, trackside component position, and ending inspection position in the route endpoint data according to the inspection direction.

5. The intelligent collaborative inspection method for railway drones based on industrial vision according to claim 1, characterized in that, The process of mapping the initial route to the airspace grid along the railway line, performing route conflict detection and segment coordination, and generating candidate routes specifically involves: The coordinates of the track centerline corresponding to the railway line are obtained from the inspection parameter set, and a spatial grid along the railway line is generated along the track centerline coordinates. The initial route is decomposed into multiple segments, and the spatial location corresponding to each segment is written into the airspace grid along the railway to generate route grid data. The UAV number, flight segment number, and time stamp corresponding to the same airspace grid in the route grid data are counted to generate flight segment occupancy data; Conflict marking is performed on flight segments with the same airspace grid and the same time marker in the flight segment occupancy data to generate conflict flight segment data; The conflicting flight segment data is subjected to altitude layer separation and time staggering coordination to generate candidate routes.

6. The intelligent collaborative inspection method for railway drones based on industrial vision according to claim 1, characterized in that, The process of extracting the visible area of ​​trackside components based on railway line data and identifying occluded and off-center sections specifically involves: The location of trackside components, coordinates of the track centerline, and track mileage are obtained from railway line data to generate a component location sequence; the spatial location, direction, and time marker of flight segments are obtained from candidate flight routes to generate a flight segment observation sequence. The component location sequence is spatially mapped to the flight segment observation sequence to generate component observation data; Based on the observation data of the components, the observation azimuth and observation distance of the trackside components relative to the candidate route are calculated, and the visible field of view of the trackside components is generated. Spatial comparison is performed between the visible field of trackside components and the observation sequence of flight segments to generate visible field comparison data. Extract the line mileage intervals where there is an occlusion relationship between the trackside components and the spatial location of the flight segment from the visual field comparison data, and generate occlusion sections. Extract the line mileage intervals where there is a deviation between the observation direction of the trackside components and the flight segment direction from the visual field comparison data, and generate the deviation section.

7. The intelligent collaborative inspection method for railway drones based on industrial vision according to claim 1, characterized in that, The process of performing perspective compensation and segment rearrangement on candidate routes to generate collaborative routes specifically involves: Based on the obstructed sections, deviation sections, and candidate routes, route correction data is generated. Lateral observation segments are added to the obstructed sections in the route correction data to generate obstruction compensation routes. Adjust the direction of the flight segment in the deviation section of the occlusion compensation route to generate a view compensation route; The segments in the perspective compensation route are sorted according to the route mileage order to generate a rearranged segment sequence; Perform segment connection and time stamp updates on the rearranged segment sequence to generate cooperative routes.

8. The intelligent collaborative inspection method for railway drones based on industrial vision according to claim 1, characterized in that, The improved Mask DINO model specifically includes a feature encoding module, a mask query module, a decoding prediction module, and a segmentation output module; The feature encoding module converts the industrial visual data along the railway line in the inspection visual stream into an image frame sequence, and extracts multi-scale image features from the image frame sequence to generate inspection image features. The mask query module receives inspection image features and introduces a track structure guidance mechanism. This mechanism extracts rail extension boundaries, sleeper arrangement boundaries, catenary suspension boundaries, bridge edge boundaries, tunnel entrance ring boundaries, and slope toe boundaries from the inspection image features, and arranges them according to line mileage to generate a track structure boundary sequence. Boundary points belonging to the same railway component in the track structure boundary sequence are connected according to their adjacent positions to generate rail extension zones, sleeper arrangement zones, catenary suspension zones, bridge edge zones, tunnel entrance boundary zones, and slope toe zones. The rail extension zone, sleeper arrangement zone, catenary suspension zone, bridge edge zone, tunnel entrance boundary zone, and slope toe zone are designated as the main track structure domains. The main domain center position, main domain extension direction, main domain boundary span, and main domain category identifier are extracted from the main track structure domains. These are then encoded into domain kernel vectors. The domain kernel vectors are written into the query vector set of the mask query module, and the general query vectors in the query vector set are replaced with the query vectors corresponding to the main track structure domains to generate track structure query vectors. The decoding and prediction module performs attention interaction between the track structure query vector and the inspection image features to generate target query features, and generates component segmentation results and defect segmentation results based on the target query features; The segmentation output module performs category labeling, contour closure, and pixel merging on the component segmentation results and defect segmentation results to generate the target segmentation result.

9. A method for intelligent collaborative inspection of railway drones based on industrial vision according to claim 1, characterized in that, The improved SuperGlue graph matching algorithm is as follows: The defect segmentation results in the target segmentation results are correlated with the image frames, time markers and line mileage in the inspection visual flow to generate defect observation data; Defect feature data is generated by extracting defect contours, defect categories, route mileage locations, image frame locations, and UAV numbers from the defect observation data. Defect map nodes are constructed by using the defect contour in the defect feature data as node boundaries, the defect category as node attributes, and the line mileage location and image frame location as node coordinates. The defect map nodes corresponding to different UAV numbers are paired, and the mileage difference, category consistency value, contour overlap value and image frame position difference between the paired defect map nodes are calculated to generate node pairing features. In the improved SuperGlue graph matching algorithm, a trace domain verification mechanism is introduced. The trace domain verification mechanism writes the node pairing features into the edge features in the graph attention matching process, and updates the node descriptors of the defective graph nodes according to the edge features to generate cross-view matching features. Based on the cross-view matching features, the matching confidence between defect map nodes under different UAV views is calculated, and a defect matching matrix is ​​generated. Extract the defect map node pairing results corresponding to the same line mileage location and the same defect category from the defect matching matrix to generate source defect data; The duplicate defect regions in the same source defect data are merged, and the position of the drift defect regions in the same source defect data is corrected to generate multi-machine defect results.

10. A railway drone intelligent collaborative inspection method based on industrial vision according to claim 1, characterized in that, Specifically, S8 is: Defect contours, image frame positions, UAV numbers, and route mileage positions are extracted from the multi-machine defect results to generate defect projection data. The image frame positions in the defect projection data are transformed to the UAV imaging coordinate system to generate defect imaging coordinates; the defect imaging coordinates are then back-projected to the railway line spatial coordinate system to generate defect spatial coordinates. The defect spatial coordinates are mapped to the line mileage location to generate the defect location result; Based on the defect category, defect outline range, and line mileage location in the defect location results, the defect level is classified and a defect level result is generated. An inspection and handling instruction is generated based on the defect level result and the defect location result.

Images