Space-air-ground cooperative railway external environment inspection method

Abstract

The application discloses a space-air-ground cooperative railway external environment inspection method, which comprises the following steps: S1, constructing a geographic information data resource pool; satellite remote sensing identifying static hidden dangers; tower high point video identifying emergent hidden dangers; laying out a UAV warehouse; S2, positioning and calibrating the static hidden dangers and the emergent hidden dangers; S3, artificially judging the hidden dangers to determine whether they are general hidden dangers, emergent hidden dangers or major hidden dangers; if the hidden dangers are general hidden dangers, generating an inspection task instruction; if the hidden dangers are emergent or major hidden dangers, generating a UAV task instruction; S4, the UAV warehouse receiving the UAV task instruction, planning an optimal route by using an improved A* algorithm, and taking evidence by the UAV; S5, a front-line inspection personnel receiving the inspection task instruction, arriving at the hidden danger position, and taking evidence on the spot; based on the UAV or artificial evidence taking result, the hidden dangers are treated, the satellite remote sensing or high point video is used to confirm that the hidden dangers are eliminated, and the whole-process data is archived to the geographic information data resource pool. The method realizes cost reduction and efficiency increase and data closed loop.

Description

Technical Field

[0001] This invention relates to the field of railway external environment safety monitoring technology, specifically to a method for railway external environment inspection that combines air, space, and ground operations. Background Technology

[0002] External environmental hazards along railway lines (such as illegal construction and occupation, hard floating objects, illegal construction, pedestrians on the tracks, etc.) are key factors affecting railway traffic safety. The traditional inspection model, which mainly relies on "manual patrols + fixed video surveillance," has the following drawbacks:

[0003] 1. Incomplete hazard identification: Manual patrols are limited by terrain (such as steep slopes and dense forests) and manpower costs, and cannot cover the entire area within 500m to the left and right of the line; fixed video surveillance focuses on equipment on the railway line, which is insufficient for the coverage of large-scale sudden hazards along the line (such as illegal burning of vegetation and pedestrians on the track), and the old equipment has poor image quality and frequent lag, making it difficult to support intelligent identification.

[0004] 2. Untimely response to hidden dangers: The traditional model relies on manual discovery of hidden dangers and reporting them level by level, which is a cumbersome process. It cannot quickly trigger the handling of sudden hidden dangers (such as light floating objects blown onto the overhead contact line), and it is easy to miss the best control opportunity.

[0005] 3. Low efficiency in location and evidence collection: When reporting hazards, the mileage and location are often estimated manually, which is inaccurate and requires repeated searches when collecting evidence manually or by drone. Drone inspections rely heavily on manual route planning, which is greatly restricted by airspace and terrain, and cannot quickly respond to the needs of collecting evidence of hazards in multiple locations.

[0006] 4. Insufficient collaboration between air, space, and ground technologies: In existing technologies, satellite remote sensing is only used for large-scale static hazard investigation, high-point video is only used for local real-time monitoring, and drones are only used for fixed-point inspections. The data from these three technologies are isolated and have not formed a collaborative closed loop of "static monitoring - dynamic early warning - rapid evidence collection," thus failing to fully leverage their technological advantages.

[0007] Existing technologies attempt to supplement security measures by introducing single technical means (such as adding drone inspections), but have not broken through the core bottlenecks of "data silos" and "positioning-route planning," and still cannot achieve full-area, intelligent and efficient inspection of the railway's external environment. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides a method for inspecting the external environment of railways using a combined air-space-ground approach.

[0009] Therefore, the present invention adopts the following technical solution:

[0010] A method for inspecting the external environment of railways using a combined air-space-ground approach includes the following steps:

[0011] S1: Construct a geographic information data resource pool; identify static hazards through satellite remote sensing; identify sudden hazards through high-point video of iron towers; deploy a drone hangar.

[0012] S2, locates and calibrates the static and sudden hazards identified in S1 to obtain the precise latitude and longitude coordinates of the hazards; and pushes the hazard information to the dispatch and command center;

[0013] S3 involves manual assessment of potential hazards to confirm their authenticity, eliminate false identifications, and determine whether a hazard is classified as a general hazard, an emergency hazard, or a major hazard. If it is a general hazard, the handling method is manual inspection and evidence collection, generating an inspection task instruction containing hazard coordinates, navigation path, and evidence collection list, which is then pushed to frontline inspection personnel. If it is an emergency or major hazard, the handling method is automatic evidence collection by drone, generating a drone task instruction containing hazard coordinates and evidence collection requirements.

[0014] S4: After receiving the drone mission instructions generated by S3, the drone library uses the improved A* algorithm to plan the optimal route, control the drone to fly automatically, collect evidence, and return to base.

[0015] In S5, frontline inspection personnel receive the inspection task instruction generated in S3, arrive at the location of the hazard, take photos / videos of the hazard on-site for evidence collection, upload them to the geographic information data resource pool, and fill in a site situation description. Based on the evidence collection results from drones or manual inspections, the dispatch and command center formulates a hazard remediation plan for remediation. After the remediation is completed, the hazard is confirmed to be eliminated through satellite remote sensing review or high-point video revisit, and the hazard status is marked as "cancelled". The entire process data is archived to the geographic information data resource pool, forming a closed loop.

[0016] The specific operation of step S1 above is as follows:

[0017] S11 integrates railway alignment data, line plane coordinate data, high-definition aerial imagery, real-world 3D models of covered sections, railway construction site elevation data, GIS maps, and railway-specific airspace rules to form a unified geographic information data resource pool; wherein, the railway alignment data includes the azimuth angle of the line alignment. The coordinates of the line centerline and the curve radius; the line plane coordinate data includes the standard mileage-latitude and longitude correspondence.

[0018] S12, the satellite periodically acquires remote sensing images of a 1km radius outside the railway fence; based on a pre-set hazard feature database, a convolutional neural network algorithm is used to detect changes in the remote sensing images from two different periods, outputting the preliminary location and type of static hazards, and storing them together with the acquired remote sensing images in the geographic information data resource pool; wherein, the hazard feature database contains 10 types of static hazard features, namely, plastic film, prefabricated steel houses, illegal construction sites, hazardous chemical storage sites, illegal dumping grounds / slag heaps, non-standard poles / tall facilities, hazards of road-rail bridges, illegal mining / quarries, illegal facilities in river channels, and dangerous trees / tall vegetation;

[0019] S13, High-point video recognition of towers to identify potential hazards:

[0020] Intelligent spherical cameras are deployed on the iron towers on both sides of the railway line; based on the railway line location data, automatic cruise points along the line are set in the camera backend to achieve uninterrupted cruise, and the real-time video stream of the high point captured is stored in the geographic information data resource pool along with the deployed camera number and camera calibration parameters.

[0021] A lightweight AI recognition model is installed on the camera to identify potential hazards in real time. When a potential hazard is identified, an early warning is triggered and the warning frame is automatically captured, along with the camera number and real-time patrol point information, and stored in the geographic information data resource pool. The potential hazards include pedestrians on the road, illegal burning of vegetation, and intrusion of large machinery.

[0022] S14. Deploy drone hangars at key construction sites along the railway line, equipped with multi-rotor drones; and store the drone hangar location information and airspace electronic fences in the geographic information data resource pool.

[0023] The specific operation of step S2 above is as follows:

[0024] S21, Satellite remote sensing hazard location calibration:

[0025] Based on the preliminary location of static hazards identified by satellite remote sensing in S12, combined with high-resolution aerial imagery, a feature matching algorithm is used to locate the specific location of the static hazards, which is then corrected to precise latitude and longitude coordinates and stored in the geographic information data resource pool.

[0026] S22, High Point Video Warning Information Mileage-Latitude / Longitude Conversion:

[0027] Based on the associated camera number and real-time cruise point information obtained from S13, the calibration parameters of the camera and the railway mileage of the corresponding cruise point are retrieved from the geographic information data resource pool. ;

[0028] If the railway mileage of the cruise point If it is completely consistent with the standard mileage, then the corresponding precise latitude and longitude are directly retrieved as the initial latitude and longitude coordinates. ); if the railway mileage of the cruise point If the distance falls exactly between two sets of standard mileages, then linear interpolation is used for calculation. The corresponding initial latitude and longitude coordinates ( );

[0029] The initial latitude and longitude coordinates ( The location of the corresponding patrol point on the centerline of the railway line is used to correct the lateral offset distance of the hazard relative to the line, based on the camera calibration parameters and the early warning frame, thus obtaining the precise latitude and longitude coordinates of the hazard. (and automatically labeled to the GIS map in the geographic information data resource pool).

[0030] In step S22 above, linear interpolation is used to calculate... The corresponding initial latitude and longitude coordinates ( The formula for () is as follows:

[0031] (1)

[0032] in, , Railway mileage for cruise points The tens mileage portion, The corresponding latitude and longitude are ( ); The next set of standard mileage is , Corresponding latitude and longitude ( ).

[0033] In step S22 above, the precise latitude and longitude coordinates of the hidden danger are obtained. The specific steps are as follows:

[0034] (1) Based on camera calibration parameters, the pixel offset between the sudden hidden danger and the center line of the line in the early warning frame. Based on the camera imaging principle, the actual lateral offset distance of the sudden hidden danger relative to the center of the line in the early warning frame is calculated. The calculation formula is:

[0035] (2)

[0036] in, To hang the camera high, It is the angle of inclination. Focal length This is a pixel scale for the camera;

[0037] (2) Combined with the azimuth of the route , lateral offset distance Convert to latitude and longitude increments ( ):

[0038] (3)

[0039] In the formula, 111319.9 represents the number of meters corresponding to 1 degree of latitude. Used to correct for the scaling difference between longitude and latitude;

[0040] (3) Calculate the precise latitude and longitude coordinates of any potential hazard using the following formula:

[0041] (4).

[0042] The specific operation of step S4 above is as follows:

[0043] S41, based on the traditional A* algorithm, considers the railway scenario and designs an improved A* algorithm;

[0044] S42, based on the improved A* algorithm designed by S41, generates 3 to 5 candidate evidence collection points and their corresponding candidate routes, sorts them according to comprehensive cost, obtains the optimal evidence collection point and the optimal route, and generates a route task package.

[0045] S43, the drone library automatically completes the drone self-check, loads the smoothed flight path mission package, takes off according to the planned route and takes pictures according to the certification requirements, and transmits the data back to the geographic information data resource pool in real time; after the certification is completed, it automatically returns to the base, and the geographic information data resource pool generates a hidden danger certification report containing flight path trajectory, certification images and mileage correlation information, and pushes it to the dispatch and command center.

[0046] Step S41 above includes:

[0047] S411, based on the dynamic constraints of railway safety taboo zones, sets taboo zone constraint penalty items. :

[0048] (5)

[0049] In the formula, Intrusion distance;

[0050] The Level 1 Prohibited Zone is a 50m radius on either side of the centerline of the line. Entry into this zone is strictly prohibited. If the train encroaches on this zone, it will be deemed impassable. This applies at all times. The Level 2 Warning Zone is a 50m radius outside the Level 1 Prohibited Zone. Entry into this zone is restricted. Only a brief stay of ≤3 minutes is allowed for evidence collection, and the altitude must be ≥15m. The Temporary Prohibited Zone is the section that the train is about to pass through. This zone temporarily expands the Level 1 Prohibited Zone to 100m, and drones are prohibited from entering it.

[0051] S412, based on railway construction site elevation adaptation, sets elevation constraint penalty items. :

[0052] (6)

[0053] In the formula, , is the height of the route node and the highest elevation of the work site The difference;

[0054] S413 sets airspace constraint penalties based on airspace rules and real-time updated no-fly zone information. :

[0055] (7)

[0056] The airspace rules include static rules and dynamic rules. The static rules include no-fly zones and restricted flight altitudes for specific sections. The no-fly zone includes the latitude and longitude boundaries over stations and marshalling yards, and hub sections. 80m, suburban section The limit is 120m; the dynamic rules include temporary no-fly notices and airspace clearance authorizations.

[0057] S414, Estimated Cost Optimization:

[0058] In the improved A* algorithm, the estimated cost is... Reconstruction into distance cost plus evidence collection and adaptation cost Composite functions:

[0059] (8)

[0060] In the formula, For the current node The straight-line distance to the potential hazard Distance-weighted, forensic adaptation cost The formula is as follows:

[0061] (9)

[0062] In the formula, Distance between evidence collection points and potential hazards;

[0063] S415, setting railway business priority weights, yields the formula for calculating the comprehensive cost of the improved A* algorithm:

[0064] (10)

[0065] In the formula, For safety weights, For the weight of evidence collection; Efficiency is the weighting factor; for urgent hidden dangers, , , .

[0066] Step S42 above includes:

[0067] S421, load the basic coordinate data, constraint data, and evidence collection requirement data, and generate 3-5 candidate evidence collection points:

[0068] The basic coordinate data includes the precise latitude and longitude coordinates of the drone hangar, the precise latitude and longitude coordinates of the potential hazards, railway alignment data, and line plane coordinate data;

[0069] The constraint data includes:

[0070] (1) Safety forbidden zone data: the coordinates of the first-level forbidden zone, the second-level warning zone, and the temporary forbidden zone;

[0071] (2) Elevation constraint data: Elevation of railway construction sites and height of terrain obstacles in the real-world 3D model;

[0072] (3) Airspace rule data: no-fly zones, restricted flight altitudes for certain sections, and temporary no-fly notices;

[0073] The evidence collection requirements data include evidence collection point constraint rules and shooting parameter requirements, wherein:

[0074] The constraint rules for evidence collection points include distance constraints between the evidence collection point and the potential hazard. Angle constraints and safety constraints: The evidence collection point must be outside the first-level prohibited area; Shooting parameter requirements: Number of photos ≥ 3, video length ≥ 30s, shooting accuracy;

[0075] Based on the aforementioned evidence collection point constraint rules, potential evidence collection point areas M are circled in the GIS map; based on potential evidence collection point areas M, spatial stratified sampling + multi-constraint compliance verification algorithm is used to generate 3 to 5 candidate evidence collection points.

[0076] S422, 3D raster map construction and attribute annotation:

[0077] Based on the constraint data described in S421, the inspection area is divided into 1 ×1 ×1 The three-dimensional grid map marks impassable grids and labels the work site elevation and flight restriction altitude corresponding to each grid.

[0078] S423, take the candidate evidence collection points obtained in S421 as the endpoints in sequence. Iterative path search is performed to obtain the corresponding candidate paths, and the optimal path is selected as the optimal route based on the comprehensive cost of the improved A* algorithm.

[0079] S424 uses a Bézier curve to smooth the optimal route obtained from S423, requiring a turning angle ≤ 15°. To avoid frequent changes in direction by the drone; after smoothing, the flight path is re-verified to ensure it meets the constraint data. If it violates the constraint, adjustments are made in steps S421 to S423; if it meets the constraint, a flight path task package is generated.

[0080] In step S421 above, the specific operation for generating candidate evidence collection points is as follows:

[0081] The potential evidence collection point area M is segmented into multiple feature sub-regions based on its distance from the hazard, elevation difference, and proximity to the obstacle; random sampling is performed within each feature sub-region to obtain 6 to 8 initial candidate points;

[0082] The constraint data and evidence collection point constraint rules are used to perform multi-constraint compliance verification, including distance, security, elevation, angle, and spatial constraints, to eliminate initial candidate points that violate regulations.

[0083] Based on the principle of optimizing security costs, evidence collection and adaptation costs, and efficiency costs, 3 to 5 candidate evidence collection points are selected.

[0084] The 3D coordinates of candidate evidence collection points are calibrated using the real-scene 3D model for subsequent iterative path search.

[0085] Step S423 above includes:

[0086] Initialize the list of nodes to be evaluated, using the precise latitude and longitude coordinates of the drone hangar as the starting point. Add nodes to the list of nodes to be evaluated and calculate their total cost; initialize the list of evaluated nodes, initially empty;

[0087] Search starting point The eight 3D neighboring nodes of the inaccessible grid are excluded, their comprehensive costs are calculated, and they are added to the list of nodes to be evaluated. The node with the lowest comprehensive cost is selected. Move to the list of evaluated nodes; continue searching for the 8 3D neighboring nodes of the node with the lowest total cost, excluding the inaccessible graticule, calculate its total cost, and add it to the list of nodes to be evaluated;

[0088] Repeat the process of selecting the node with the lowest overall cost from the list of nodes to be evaluated. Move to the list of evaluated nodes; continue searching for the 8 3D neighboring nodes of the node with the lowest overall cost, excluding impassable grids, calculating their overall cost, and adding them to the list of nodes to be evaluated, until the endpoint is found among the neighboring nodes of the current node. At that time, the search ends, and the zigzag route is obtained;

[0089] Compare the candidate broken line routes corresponding to multiple candidate evidence collection points, and select the route with the lowest overall cost of the improved A* algorithm corresponding to the candidate evidence collection point as the optimal route.

[0090] Compared with the prior art, the present invention has the following beneficial effects:

[0091] 1. Mileage-Latitude and Longitude Intelligent Conversion Positioning: The method of this invention achieves a breakthrough from "fuzzy mileage positioning" to "precise latitude and longitude positioning" by correcting through linear interpolation and perspective transformation. The positioning accuracy of the hidden danger is ≤5m, which solves the problems of poor accuracy and low efficiency of traditional manual positioning.

[0092] 2. Multi-Constraint UAV Intelligent Flight Route Planning: This invention integrates multi-source data such as real-world 3D data, electronic fences, and railway construction site elevation data. An improved A* algorithm is proposed for flight route planning, achieving a UAV flight route planning time of ≤1 minute. This satisfies airspace and terrain constraints while ensuring evidence collection accuracy and flight safety, avoiding the tediousness and risks of manual flight route planning. The automatic UAV flight route planning avoids obstacles and railway lines, preventing flight accidents and traffic interference, thus improving inspection safety.

[0093] 3. Integrated Air-Ground-Space Collaborative Closed-Loop: This invention breaks down data silos between satellite remote sensing (static monitoring), high-point video (dynamic early warning), and unmanned aerial vehicle (UAV) databases (rapid evidence collection). By combining manual assessment with a geographic information data resource pool, it forms a closed-loop process of "identification-location-assessment-evidence collection-rectification-deactivation," significantly improving inspection efficiency. Evidence collection response time is ≤15 minutes, more than three times more efficient than traditional manual inspections, preventing the escalation of potential hazards.

[0094] 4. Cost reduction and efficiency improvement and data closed loop: The method of this invention reduces the workload of manual patrol by 60% and lowers labor costs; the data of the whole process is archived to the geographic information data resource pool, which supports the analysis of hidden danger trends and provides data support for the long-term management and control of the railway external environment. Attached Figure Description

[0095] Figure 1 This is an overall flowchart of the railway external environment inspection method in this embodiment of the invention;

[0096] Figure 2 This is a block diagram of the improved A* algorithm optimization logic in this invention;

[0097] Figure 3This is a flowchart of the route planning using the improved A* algorithm in this invention. Detailed Implementation

[0098] The technical solution of the invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. Obviously, the following embodiments are only some embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0099] This invention provides a method for inspecting the external environment of railways using a combined air-space-ground approach, such as... Figure 1 As shown, it includes the following steps:

[0100] S1. Construct a geographic information data resource pool; identify static hazards through satellite remote sensing; identify sudden hazards through high-point video of iron towers; deploy a drone hangar; including the following steps:

[0101] S11, Construct a geographic information data resource pool.

[0102] Integrate railway alignment data (including alignment direction and azimuth) The database includes core geometric parameters such as centerline coordinates and curve radius, plane coordinate data of the railway line (including the standard mileage-latitude and longitude correspondence, i.e., the latitude and longitude corresponding to every 10m mileage), high-definition aerial imagery, real-scene 3D models of the covered sections, elevation data of railway construction sites (such as roadbed height and bridge clearance), GIS maps, and railway-specific airspace rules, forming a unified geographic information data resource pool that supports the storage, correlation query, and collaborative calling of multi-source (air, ground, and space) data.

[0103] S12, satellite remote sensing identifies static potential hazards.

[0104] Sub-meter resolution satellites with a resolution of 0.5m to 0.75m were selected to acquire remote sensing images of a 1km radius outside the railway fence every 3 months. Based on a pre-set hazard feature library, a convolutional neural network (CNN) algorithm was used to detect changes between the two sets of remote sensing images. The algorithm outputs the preliminary location of static hazards (including the estimated railway mileage range, such as K262+100-K262+200) and the type of static hazard, which are stored together with the acquired remote sensing images in the geographic information data resource pool.

[0105] The hazard feature database contains 10 categories of static hazard features, namely: plastic film, prefabricated steel houses, illegal construction sites, hazardous chemical storage sites, illegal dumping grounds, unregulated poles / tall facilities, hazards on public-railway bridges, illegal mining / quarries, illegal river facilities, and dangerous trees / tall vegetation.

[0106] S13, video recognition of high-point towers to identify potential hazards.

[0107] Intelligent dome cameras with 4 megapixels and 53x zoom are deployed on iron towers 100m to 500m apart and 1.5km to 2km apart on both sides of the railway line, covering a line-of-sight range of 500m to the left and right of the line. Based on the railway line location data, automatic patrol points along the line are set in the camera backend to achieve 24 / 7 uninterrupted patrol. The real-time video streams of the high points captured, along with the deployed camera numbers and camera calibration parameters (including camera height, tilt angle, and focal length), and the railway mileage of the patrol points (such as K262+150) are stored in the geographic information data resource pool.

[0108] Lightweight AI recognition models (such as YOLOv8-tiny) are installed on cameras to identify potential hazards such as pedestrians on the road, illegal burning of fields, and intrusion of large machinery in real time. When a potential hazard is identified, an alert is triggered and the alert frame is automatically captured, along with the associated camera number and real-time patrol point information, and stored in the geographic information data resource pool.

[0109] S14, deploying a drone hangar.

[0110] Unmanned aerial vehicle (UAV) hangars will be set up at key construction sites along the railway line (such as bridges, tunnel entrances, and high-risk slopes), with at least one hangar every 10km. Each hangar will be equipped with a multi-rotor UAV (with a flight time of ≥40min, equipped with a 4K optical camera and a Beidou positioning module), and the location information of the UAV hangars and the airspace electronic fence (no-fly zone, restricted flight altitude) will be stored in the geographic information data resource pool.

[0111] S2, precise location and push notification of potential hazards.

[0112] S21, Satellite remote sensing hazard location calibration.

[0113] Based on the preliminary location of static hazards identified by satellite remote sensing in S12, and combined with high-resolution aerial images (5cm resolution) in the geographic information data resource pool, the specific location of static hazards is located using a feature matching algorithm (such as SIFT), and then corrected to precise latitude and longitude coordinates and stored in the geographic information data resource pool.

[0114] S22, high-point video early warning information mileage-latitude / longitude conversion.

[0115] Based on the associated camera number and real-time cruise point information obtained from S13, the calibration parameters of the camera and the railway mileage of the corresponding cruise point are retrieved from the geographic information data resource pool. The pre-stored route plane coordinate data in the geographic information data resource pool is labeled with one set of precise latitude and longitude every 10m of mileage, as shown in Table 1:

[0116] Table 1

[0117]

[0118] If the railway mileage of the cruise point If the standard mileage is used (e.g., K262+150), then the precise latitude and longitude corresponding to that mileage will be directly retrieved as the initial latitude and longitude coordinates. If the railway mileage of the cruise point... (For example, K262+153) falls exactly between two sets of standard mileages (such as K262+150 and K262+160), and the railway mileage needs to be calculated using linear interpolation. The corresponding initial latitude and longitude coordinates are used to achieve a preliminary conversion from mileage to latitude and longitude, with an error ≤3m, meeting the basic accuracy requirements for subsequent offset correction; linear interpolation is used to calculate railway mileage. The specific steps for obtaining the initial latitude and longitude coordinates are as follows:

[0119] Set the railway mileage of the cruise point The tens mileage portion is (Corresponding to standard mileage, such as K262+150), the next set of standard mileage is... (e.g., K262+160) and The difference is Retrieve Corresponding latitude and longitude ( ), Corresponding latitude and longitude ( ), calculated using linear interpolation formula The corresponding initial latitude and longitude coordinates ( The calculation formula is as follows:

[0120] (1)

[0121] Initial latitude and longitude coordinates ( The location of the corresponding patrol point on the centerline of the railway line needs to be determined by combining the camera calibration parameters and the warning image frames to correct the lateral offset distance of the hazard relative to the line, thus obtaining the precise latitude and longitude coordinates of the hazard. (Error ≤ 5m) and automatically labeled to the GIS map in the geographic information data resource pool. The specific operation is as follows:

[0122] (1) Based on camera calibration parameters, the pixel offset between the sudden hidden danger and the center line of the line in the early warning frame. (Automatically extracted using image recognition algorithms) Based on the camera imaging principle, calculate the actual lateral offset distance of the sudden hidden danger relative to the center of the line in the early warning frame. (Unit: m), the calculation formula is:

[0123] (2)

[0124] in, To hang the camera high, It is the angle of inclination. Focal length This is the camera pixel scale (unit: pixels / m, provided by factory calibration parameters, such as 10 pixels / m).

[0125] (2) Combined with the azimuth of the route (Retrieved from the geographic information data resource pool, e.g., 90° represents the east-west direction of the route), the lateral offset distance will be... Convert to latitude and longitude increments :

[0126] (3)

[0127] Where 111319.9 represents the number of meters corresponding to 1 degree of latitude. Used to correct for the scaling difference between longitude and latitude.

[0128] (3) Calculate the precise latitude and longitude coordinates of any potential hazard using the following formula:

[0129] (4)

[0130] Finally, the precise latitude and longitude coordinates of the sudden hazard were obtained. With a comprehensive error of ≤5m, it can be directly used for UAV automatic navigation or manual inspection positioning.

[0131] S23, Hazard Information Push. Hazard information (including latitude and longitude coordinates, hazard type, early warning screen frames / satellite imagery, and mileage) is pushed to the dispatch and command center of the corresponding station segment in real time through the geographic information data resource pool.

[0132] S3, Hazard assessment and task assignment.

[0133] Professionals at the dispatch and command center can view the details of potential hazards through a visualization platform, conduct manual assessments, confirm the authenticity of the hazards (excluding false identifications), and determine the hazard level as: general hazard, emergency hazard, or major hazard.

[0134] For general hazards, the handling method is manual inspection and evidence collection, generating inspection task instructions that include hazard coordinates, navigation path, and evidence collection list, and pushing them to the mobile APP of front-line inspection personnel.

[0135] In the case of an emergency or major hazard, the method of handling it is to use drones to automatically collect evidence and generate drone mission instructions that include the coordinates of the hazard and the evidence collection requirements (such as shooting angle and number of photos).

[0136] S4, Intelligent Flight Route Planning and Automatic Evidence Collection for Drones Based on an Improved A* Algorithm. After receiving mission instructions from the drone, the drone library integrates drone location information, hazard latitude and longitude coordinates, a real-world 3D model, airspace electronic fence data, and railway construction site elevation data. It then uses the improved A* algorithm to plan the optimal flight route, controlling the drone to automatically fly, collect evidence, and return to base. The specific operation is as follows:

[0137] S41, based on the traditional A* algorithm, considering the railway scenario, designs an improved A* algorithm, such as... Figure 2 As shown, the details are as follows:

[0138] S411, based on the dynamic constraints of railway safety taboo zones, sets taboo zone constraint penalty items. .

[0139] Based on railway alignment data in the geographic information data resource pool, multi-level safety taboo zones are marked, stored in the geographic information data resource pool, and linked with railway traffic scheduling data to achieve dynamic adjustments.

[0140] The Level 1 Prohibited Zone is a 50m radius on both sides of the centerline of the railway line. Entry into this area is strictly prohibited. Any intrusion into the railway line is considered impassable. This applies at all times and complies with the safety protection zone requirements of the "Railway Safety Management Regulations".

[0141] The Level II warning zone is 50m outside the Level I prohibited zone (50-100m on both sides of the route). Entry is restricted within this area, and only short stays (≤3min) are allowed for evidence collection, with an altitude ≥15m. This zone is used for routine inspections to avoid excessively restricting the flight range.

[0142] The temporary restricted area is the section that the train is about to pass through (≤10 minutes). This area is temporarily expanded to 100 meters from the first-level restricted area to prohibit drones from entering, dynamically adapting to the railway operation status and avoiding conflicts with trains.

[0143] Set forbidden zones and related penalties. The penalty value changes dynamically with the depth of intrusion and the level of the forbidden zone, as shown in the following formula:

[0144] (5)

[0145] In the formula, The depth of the intrusion is the penalty value; the deeper the intrusion, the greater the penalty.

[0146] S412, based on railway construction site elevation adaptation, sets elevation constraint penalty items. .

[0147] Elevation data of railway construction sites are retrieved from the geographic information data resource pool and combined with a real-world 3D model (including terrain obstacle heights) to establish a joint elevation constraint between the construction site and the terrain. An elevation constraint penalty is then set. The height of each candidate route node Real-time verification is performed, and a penalty is triggered if the requirements are not met, as shown in the following formula:

[0148] (6)

[0149] In the formula, , is the height of the route node and the highest elevation of the work site The difference, Insufficient safety redundancy will result in greater penalties for larger redundancies.

[0150] S413 sets airspace constraint penalties based on airspace rules and real-time updated no-fly zone information. .

[0151] Railway-specific airspace rules are imported from the geographic information data resource pool, including static and dynamic rules. Static rules include no-fly zones (latitude and longitude boundaries above stations and marshalling yards) and section-specific flight altitude restrictions (80m for hub sections and 120m for suburban sections). Dynamic rules include temporary no-fly notices (such as during key events or construction periods) and airspace lifting authorizations (such as maintenance sections of engineering departments). Airspace constraint penalties are set. The candidate routes undergo multi-dimensional airspace verification, and violations trigger penalties, as shown in the following formula:

[0152] (7)

[0153] By interacting in real time with the railway airspace management platform through a geographic information data resource pool, temporary no-fly notifications are received. For example, if a certain section is temporarily closed to flying for 4 hours, the algorithm automatically marks that area as a temporary no-fly zone. To avoid planning ineffective routes.

[0154] S414, estimated cost optimization.

[0155] In the traditional A* algorithm, the overall cost is the actual distance. (Current node) Distance from the starting point and estimated cost (Current node) The sum of distances from the endpoint, where, Only straight-line distance is considered; the effectiveness of evidence collection is not taken into account.

[0156] In the improved A* algorithm provided by this invention, the estimated cost is... Reconstruction into distance cost plus evidence collection and adaptation cost The composite function ensures that the evidence collection task can be completed upon arrival. The formula is as follows:

[0157] (8)

[0158] In the formula, For the current node straight-line distance to the hidden danger ( ), Distance-weighted, forensic adaptation cost The formula is as follows:

[0159] (9)

[0160] In the formula, Distance between evidence collection points and potential hazards.

[0161] S415 sets the priority weights for railway business, resulting in the formula for calculating the comprehensive cost of the improved A* algorithm.

[0162] Based on the railway's operational requirements of "safety first, certification second, and efficiency third," differentiated weights are assigned to each cost item to ensure matching priorities. Safety costs include prohibited zone constraints and penalties. and airspace constraint penalty item Set security weights Highest priority; cost of evidence collection and adaptation. Set evidence collection weights Efficiency costs include actual distance. and elevation constraint penalty item Set efficiency weights .

[0163] The formula for calculating the comprehensive cost of the improved A* algorithm is obtained as follows:

[0164] (10)

[0165] For urgent hazards (e.g., pedestrians on the road), safety weights can be applied: Evidence weighting: Efficiency weight: Configure settings to ensure the fastest and safest evidence collection, balancing security and efficiency.

[0166] The estimated cost after reconstruction described in step S414 This is a core component of the comprehensive cost of the improved A* algorithm described above, representing the reconstructed estimated cost. In This is indirectly reflected through efficiency costs; evidence collection and adaptation costs. As an independent cost item, it is directly included in the comprehensive cost calculation to achieve the dual objectives of "optimal distance + effective evidence collection".

[0167] This function directly incorporates the estimated reconstruction cost. The core connotation, among which correspond The cost of evidence collection and adaptation in the process correspond The distance cost in the process is weighted to achieve priority matching of "security-evidence-efficiency" to ensure reconstruction. The design goals were fully realized in the comprehensive cost calculation.

[0168] S42, based on the improved A* algorithm designed in S41, plans the optimal route for the UAV: ​​It generates 3-5 candidate evidence collection points and their corresponding candidate routes, sorts them by comprehensive cost, and obtains the optimal evidence collection point and optimal route. If the optimal evidence collection point has temporary obstacles (such as construction machinery), it can quickly switch to the second-best option to avoid evidence collection failure. Figure 3 As shown, the specific steps are as follows:

[0169] S421, Data initialization, requires loading basic coordinate data, constraint data, and evidence collection requirement data, and generating candidate evidence collection points.

[0170] The basic coordinate data includes: starting point Coordinates: Precise latitude and longitude coordinates of the drone hangar; precise latitude and longitude coordinates of potential hazards; Geographic benchmark data: railway alignment data and line plane coordinate data, providing a geographic reference framework for path search.

[0171] The constraint data includes:

[0172] (1) Safety forbidden zone data: the coordinates of the first-level forbidden zone, the second-level warning zone, and the temporary forbidden zone;

[0173] (2) Elevation constraint data: Elevation of railway construction sites and height of terrain obstacles (such as trees and building heights) in the real-world 3D model;

[0174] (3) Airspace rule data: no-fly zones (latitude and longitude boundaries of stations and marshalling yards), restricted flight altitudes for sections (80m for hubs and 120m for suburban areas), and temporary no-fly notices (such as a 4-hour no-fly zone for construction sections).

[0175] The data required for evidence collection includes: evidence collection point constraint rules: distance constraints between the evidence collection point and the potential hazard ( The requirements include: angle constraints (side shot, front lighting, avoid backlighting / overhead view > 60°), safety constraints (the evidence collection point must be outside the first-level prohibited area); shooting parameter requirements: number of photos (≥3), video length (≥30s), shooting resolution (4K high definition), to provide standards for subsequent determination of the validity of the evidence collection point.

[0176] Based on the aforementioned evidence collection point constraint rules, potential evidence collection point areas M (e.g., a rectangular area 20-30m north of the potential hazard, unobstructed, and with direct sunlight) are circled on the GIS map.

[0177] Based on the potential evidence collection point region M, a spatial stratified sampling + multi-constraint compliance verification algorithm is used to generate 3-5 candidate evidence collection points. Specifically, the potential evidence collection point region M is stratified into multiple feature sub-regions based on distance from potential hazards, elevation differences, and obstacle proximity. Random sampling is performed within each feature sub-region to obtain 6-8 initial candidate points. Multi-constraint compliance verification, including distance, safety, elevation, angle, and spatial constraints, is performed using the constraint data and evidence collection point constraint rules to eliminate non-compliant initial candidate points. Based on the principle of optimal safety cost, evidence collection adaptation cost, and efficiency cost, 3-5 candidate evidence collection points are retained. The 3D coordinates of the candidate evidence collection points are calibrated using a real-world 3D model for subsequent iterative path search.

[0178] S422, 3D raster map construction and attribute annotation, specifically: based on the constraint data described in S421, the inspection area is divided into 1 ×1 ×1 The three-dimensional grid map is marked with "inaccessible grids" (restricted areas, no-fly zones, terrain obstacles), and the elevation of the work site and the restricted flight altitude are marked for each grid.

[0179] S423, take the candidate evidence collection points obtained in S421 as the endpoints in sequence. The algorithm performs iterative path search to obtain corresponding candidate paths, and selects the optimal path as the optimal route based on the comprehensive cost of the improved A* algorithm.

[0180] Initialize the list of nodes to be evaluated, starting from the node. Add nodes to the list of nodes to be evaluated and calculate their total cost; initialize the list of evaluated nodes, which is initially empty.

[0181] Search starting point The eight 3D neighboring nodes of the inaccessible grid are excluded, their comprehensive costs are calculated, and they are added to the list of nodes to be evaluated. The node with the lowest comprehensive cost is selected. Move to the list of evaluated nodes; continue searching for the 8 3D neighboring nodes of the node with the lowest total cost, excluding the inaccessible graticule, calculate its total cost, and add it to the list of nodes to be evaluated;

[0182] Repeat the process of selecting the node with the lowest overall cost from the list of nodes to be evaluated. Move to the list of evaluated nodes; continue searching for the 8 3D neighboring nodes of the node with the lowest overall cost, excluding impassable grids, calculating their overall cost, and adding them to the list of nodes to be evaluated, until the endpoint is found among the neighboring nodes of the current node. At that point, the search ends, and the broken line route is obtained.

[0183] Compare the candidate broken line routes corresponding to multiple candidate evidence collection points, and select the route with the lowest overall cost of the improved A* algorithm corresponding to the candidate evidence collection point as the optimal route.

[0184] S424, Flight Line Smoothing and Verification.

[0185] The optimal route obtained from S423 is smoothed using Bézier curves, requiring a turning angle ≤ 15°. To avoid frequent changes in direction by drones; after smoothing, the flight path is re-verified (secondary verification) to ensure that all constraints (restricted areas, elevation, airspace) are met. If there are any violations, adjustments S421~S423 are executed; if they are met, a flight path task package is generated.

[0186] S43, automated evidence collection by drone.

[0187] The drone warehouse automatically completes the drone self-check, loads the smoothed flight path mission package, and takes off according to the planned flight path.

[0188] During flight (including the journey to the evidence collection point and the return flight after evidence collection), the drone uses BeiDou positioning and a real-world 3D model to correct its position in real time. For fixed obstacles, it deviates horizontally by 3-5 meters and rises 2-3 meters to bypass them along a preset trajectory. When encountering temporary obstacles, it triggers local replanning. If it approaches a restricted area, it automatically deviates 2-3 meters to the outside to maintain a safe distance, ensuring that the flight path deviation is ≤ Upon reaching the evidence collection point, the drone hovers stably (hovering accuracy ≤ 0.5m), ceases detour operations, and takes photos as required (e.g., ≥ 3 photos from different angles, ≥ 30 seconds of video), transmitting the data back to the geographic information data resource pool in real time. After evidence collection is complete, the drone automatically returns to base, and the geographic information data resource pool generates a hazard evidence collection report (including flight path, evidence images, and mileage correlation information), which is then pushed to the dispatch and command center.

[0189] S5, manual inspection and evidence collection and closed-loop management, details are as follows:

[0190] S51, Manual Inspection Navigation and Evidence Collection: Frontline inspection personnel receive inspection task instructions through a mobile APP, and use the real-time navigation function based on the GIS map (supporting offline mode) to reach the location of the hazard; take photos / videos of the hazard on-site for evidence collection, upload them to the geographic information data resource pool, and fill in a description of the on-site situation (such as the scale of the hazard and suggestions for handling).

[0191] S52, Closed-loop management of hidden dangers: The dispatch and command center formulates a hidden danger rectification plan based on the results of drone or manual evidence collection; after the rectification is completed, the elimination of hidden dangers is confirmed by satellite remote sensing or high-point video revisit, and the status of hidden dangers is marked as "cancelled". All process data (including identification, analysis, evidence collection and rectification records) are archived to the geographic information data resource pool to form a closed loop.

[0192] Example

[0193] In this embodiment, taking the sudden hazard of "pedestrians crossing the road" as an example, the implementation process of the present invention is described in detail:

[0194] (1) Video analysis of high-point towers identifies potential hazards:

[0195] A camera (numbered T-001) on a tower along a high-speed railway patrolled to K262+150 according to a preset patrol point. Based on the real-time video stream from the high point, the YOLOv8-tiny model identified the sudden hazard of "pedestrians on the track" in real time, triggered an early warning, captured the warning frame, associated it with the camera number and the real-time patrol point mileage (K262+150), and transmitted it to the geographic information data resource pool for storage via a 50M VPN dedicated line.

[0196] (2) Accurate location and notification of potential hazards:

[0197] Retrieve the calibration parameters (hanging height 45m, tilt angle 30°) of camera T-001 from the geographic information data resource pool, along with the latitude and longitude corresponding to K262+150. =116.7554°E, =36.9652°N); the lateral offset distance of the hazard (pedestrian) relative to the center of the road is 12m. After correction, the final accurate latitude and longitude coordinates of the hazard are obtained. =116.7560°E, =36.9655°N), and marked it on the GIS map. The hazard information (including latitude and longitude coordinates, hazard type, early warning screen frame and mileage) is pushed to the dispatch and command center of the corresponding station section through the geographic information data resource pool.

[0198] (3) Hazard assessment and task assignment:

[0199] Once the hazard is confirmed as "urgent" (pedestrians on the road endangering traffic safety), select "automatic drone evidence collection"; generate drone mission instructions: hazard coordinates (116.7560°E, 36.9655°N), evidence collection requirements (take 5 photos from 3 angles (side shots), 1 30-second video), and push them to the nearest drone hangar (located at K260+000, 2.1km away from the hazard).

[0200] (4) Drone route planning and certification:

[0201] Starting point S: Drone hangar (K260+000, coordinates 116.7300°E, 36.9502°N); Hazard coordinates (116.7560°E, 36.9655°); Real-world 3D model: No tall buildings around the hazard; Restricted area: 50m on both sides of the route is a Level 1 restricted area; Work site elevation: K262+150 is the roadbed, H=5m, flight path height must be ≥8m; Airspace rules: This section is suburban, restricted flight altitude is 120m, no no-fly zone; Evidence collection requirements: 20-30m distance, side shot, 5 photos + 30s video.

[0202] Traditional A* algorithm path: The distance from the starting point S of the drone hangar to the coordinates of the potential hazard is 2100m, but it needs to fly 40m north of the route (Level 1 forbidden zone), which is directly excluded.

[0203] Candidate paths are obtained by path search based on the improved A* algorithm provided in this invention:

[0204] Candidate Path 1: Fly 300m north of the route. The candidate evidence collection point is 2300m from the starting point at an altitude of 11m (≥5+3=8m). The candidate evidence collection point (116.7558°E, 36.9660°N) is 25m from the potential hazard. Calculate the comprehensive cost corresponding to the candidate evidence collection point:

[0205] =2300, =0, =0, =0, =0, =3×(0+0)+2×0+1×(2300+0)=2300

[0206] Candidate Path 2: Fly 200m south of the route. The candidate evidence collection point is 2200m from the starting point at an altitude of 11m. However, the candidate evidence collection point (116.7563°E, 36.9640°N) is 28m away from the potential hazard (P_4=20×1=20). Calculate the comprehensive cost corresponding to the candidate evidence collection point:

[0207] =2200, =0, =0, =0, =0, =3×(0+0)+2×0+1×(2200+0)=2200

[0208] Candidate route 2 is better. Route smoothing and verification are performed on candidate route 2. If it meets the requirements, a route task package is generated.

[0209] The drone flies along the planned route at a speed of 8 m / s for about 4 minutes. After reaching the evidence collection point, it takes 5 side photos (clearly showing the relative positions of pedestrians and fences) and 1 30-second video. The data is transmitted back to the dispatch center in real time, and the drone returns to base after 12 minutes, meeting the emergency hazard response requirement of ≤15 minutes.

[0210] (5) Hazard Management. Based on the evidence collection results, the dispatch center notifies the nearest engineering section emergency personnel to rush to the scene and drive away pedestrians within 15 minutes; 1 hour later, the hazard is confirmed to be eliminated through high-point video revisit, the hazard status is marked as "cancelled", and the entire process data is archived to the geographic information data resource pool.

Claims

1. A method for inspecting the external environment of railways using a combined air-space-ground approach, characterized in that: Includes the following steps: S1: Construct a geographic information data resource pool; identify static hazards through satellite remote sensing; identify sudden hazards through high-point video of iron towers; deploy a drone hangar. S2, locates and calibrates the static and sudden hazards identified in S1 to obtain the precise latitude and longitude coordinates of the hazards; and pushes the hazard information to the dispatch and command center; S3. Conduct manual assessment of potential hazards to confirm their authenticity, eliminate false identifications, and determine whether the hazard is classified as a general hazard, an emergency hazard, or a major hazard. For general hazards, the handling method is manual inspection and evidence collection, generating an inspection task instruction that includes the hazard coordinates, navigation path, and evidence collection list, and pushing it to the front-line inspection personnel; In the case of an emergency or major hazard, the handling method is to use drones to automatically collect evidence and generate drone mission instructions that include the coordinates of the hazard and the evidence collection requirements. S4: After receiving the drone mission instructions generated by S3, the drone library uses the improved A* algorithm to plan the optimal route, control the drone to fly automatically, collect evidence, and return to base. In S5, frontline inspection personnel receive the inspection task instruction generated in S3, arrive at the location of the hazard, take photos / videos of the hazard on-site for evidence collection, upload them to the geographic information data resource pool, and fill in a site situation description. Based on the evidence collection results from drones or manual inspections, the dispatch and command center formulates a hazard remediation plan for remediation. After the remediation is completed, the hazard is confirmed to be eliminated through satellite remote sensing review or high-point video revisit, and the hazard status is marked as "cancelled". The entire process data is archived to the geographic information data resource pool, forming a closed loop.

2. The railway external environment inspection method according to claim 1, characterized in that, The specific operation of step S1 is as follows: S11 integrates railway alignment data, line plane coordinate data, high-definition aerial imagery, real-world 3D models of covered sections, railway construction site elevation data, GIS maps, and railway-specific airspace rules to form a unified geographic information data resource pool; wherein, the railway alignment data includes the azimuth angle of the line alignment. The coordinates of the line centerline and the curve radius; the line plane coordinate data includes the standard mileage-latitude and longitude correspondence. S12, the satellite periodically acquires remote sensing images of a 1km radius outside the railway fence; based on a pre-set hazard feature database, a convolutional neural network algorithm is used to detect changes in the remote sensing images from two different periods, outputting the preliminary location and type of static hazards, and storing them together with the acquired remote sensing images in the geographic information data resource pool; wherein, the hazard feature database contains 10 types of static hazard features, namely, plastic film, prefabricated steel houses, illegal construction sites, hazardous chemical storage sites, illegal dumping grounds / slag heaps, non-standard poles / tall facilities, hazards of road-rail bridges, illegal mining / quarries, illegal facilities in river channels, and dangerous trees / tall vegetation; S13, High-point video recognition of towers to identify potential hazards: Intelligent spherical cameras are deployed on the iron towers on both sides of the railway line; based on the railway line location data, automatic cruise points along the line are set in the camera backend to achieve uninterrupted cruise, and the real-time video stream of the high point captured is stored in the geographic information data resource pool along with the deployed camera number and camera calibration parameters. A lightweight AI recognition model is installed on the camera to identify potential hazards in real time. When a potential hazard is identified, an early warning is triggered and the warning frame is automatically captured, along with the camera number and real-time patrol point information, and stored in the geographic information data resource pool. The potential hazards include pedestrians on the road, illegal burning of vegetation, and intrusion of large machinery. S14. Deploy drone hangars at key construction sites along the railway line, equipped with multi-rotor drones; and store the drone hangar location information and airspace electronic fences in the geographic information data resource pool.

3. The railway external environment inspection method according to claim 2, characterized in that, The specific operation of step S2 is as follows: S21, Satellite remote sensing hazard location calibration: Based on the preliminary location of static hazards identified by satellite remote sensing in S12, combined with high-resolution aerial imagery, a feature matching algorithm is used to locate the specific location of the static hazards, which is then corrected to precise latitude and longitude coordinates and stored in the geographic information data resource pool. S22, High Point Video Warning Information Mileage-Latitude / Longitude Conversion: Based on the associated camera number and real-time cruise point information obtained from S13, the calibration parameters of the camera and the railway mileage of the corresponding cruise point are retrieved from the geographic information data resource pool. ; If the railway mileage of the cruise point If it is completely consistent with the standard mileage, then the corresponding precise latitude and longitude are directly retrieved as the initial latitude and longitude coordinates. ); if the railway mileage of the cruise point If the distance falls exactly between two sets of standard mileages, then linear interpolation is used for calculation. The corresponding initial latitude and longitude coordinates ( ); The initial latitude and longitude coordinates ( The location of the corresponding patrol point on the centerline of the railway line is used to correct the lateral offset distance of the hazard relative to the line, based on the camera calibration parameters and the early warning frame, thus obtaining the precise latitude and longitude coordinates of the hazard. (and automatically labeled to the GIS map in the geographic information data resource pool).

4. The railway external environment inspection method according to claim 3, characterized in that, In step S22, linear interpolation is used to calculate... The corresponding initial latitude and longitude coordinates ( The formula for () is as follows: （1） in, , For the railway mileage of the cruise point The tens mileage portion, The corresponding latitude and longitude are ( ); The next set of standard mileage is , Corresponding latitude and longitude ( ).

5. The railway external environment inspection method according to claim 4, characterized in that, In step S22, the precise latitude and longitude coordinates of the hidden danger are obtained. The specific steps are as follows: (1) Based on camera calibration parameters, the pixel offset between the sudden hidden danger and the center line of the line in the early warning frame. Based on the camera imaging principle, the actual lateral offset distance of the sudden hidden danger relative to the center of the line in the early warning frame is calculated. The calculation formula is: （2） in, To hang the camera high, It is the angle of inclination. Focal length This is a pixel scale for the camera; (2) Combined with the azimuth of the route , lateral offset distance Convert to latitude and longitude increments ( ): （3） In the formula, 111319.9 represents the number of meters corresponding to 1 degree of latitude. Used to correct for the scaling difference between longitude and latitude; (3) Calculate the precise latitude and longitude coordinates of any potential hazard using the following formula: （4）。 6. The railway external environment inspection method according to claim 5, characterized in that, The specific operation of step S4 is as follows: S41, based on the traditional A* algorithm, considers the railway scenario and designs an improved A* algorithm; S42, based on the improved A* algorithm designed by S41, generates 3 to 5 candidate evidence collection points and their corresponding candidate routes, sorts them according to comprehensive cost, obtains the optimal evidence collection point and the optimal route, and generates a route task package. S43, the drone library automatically completes the drone self-check, loads the smoothed flight path mission package, takes off according to the planned route and takes pictures according to the certification requirements, and transmits the data back to the geographic information data resource pool in real time; after the certification is completed, it automatically returns to the base, and the geographic information data resource pool generates a hidden danger certification report containing flight path trajectory, certification images and mileage correlation information, and pushes it to the dispatch and command center.

7. The railway external environment inspection method according to claim 6, characterized in that, Step S41 includes: S411, based on the dynamic constraints of railway safety taboo zones, sets taboo zone constraint penalty items. : （5） In the formula, Intrusion distance; The Level 1 Prohibited Zone is a 50m radius on either side of the centerline of the line. Entry into this zone is strictly prohibited. If the train encroaches on this zone, it will be deemed impassable. This applies at all times. The Level 2 Warning Zone is a 50m radius outside the Level 1 Prohibited Zone. Entry into this zone is restricted. Only a brief stay of ≤3 minutes is allowed for evidence collection, and the altitude must be ≥15m. The Temporary Prohibited Zone is the section that the train is about to pass through. This zone temporarily expands the Level 1 Prohibited Zone to 100m, and drones are prohibited from entering it. S412, based on railway construction site elevation adaptation, sets elevation constraint penalty items. : （6） In the formula, , is the height of the route node and the highest elevation of the work site The difference; S413 sets airspace constraint penalties based on airspace rules and real-time updated no-fly zone information. : （7） The airspace rules include static rules and dynamic rules. The static rules include no-fly zones and restricted flight altitudes for specific sections. The no-fly zone includes the latitude and longitude boundaries over stations and marshalling yards, and hub sections. 80m, suburban section The limit is 120m; the dynamic rules include temporary no-fly notices and airspace clearance authorizations. S414, Estimated Cost Optimization: In the improved A* algorithm, the estimated cost is... Reconstruction into distance cost plus evidence collection and adaptation cost Composite functions: （8） In the formula, For the current node The straight-line distance to the potential hazard Distance-weighted, forensic adaptation cost The formula is as follows: （9） In the formula, Distance between evidence collection points and potential hazards; S415, setting railway business priority weights, yields the formula for calculating the comprehensive cost of the improved A* algorithm: （10） In the formula, For safety weights, For the weight of evidence collection; Efficiency is the weighting factor; for urgent hidden dangers, , , .

8. The railway external environment inspection method according to claim 7, characterized in that, Step S42 includes: S421, load the basic coordinate data, constraint data, and evidence collection requirement data, and generate 3-5 candidate evidence collection points: The basic coordinate data includes the precise latitude and longitude coordinates of the drone hangar, the precise latitude and longitude coordinates of the potential hazards, railway alignment data, and line plane coordinate data; The constraint data includes: (1) Safety forbidden zone data: the coordinates of the first-level forbidden zone, the second-level warning zone, and the temporary forbidden zone; (2) Elevation constraint data: Elevation of railway construction sites and height of terrain obstacles in the real-world 3D model; (3) Airspace rule data: no-fly zones, restricted flight altitudes for certain sections, and temporary no-fly notices; The evidence collection requirements data include evidence collection point constraint rules and shooting parameter requirements, wherein: The constraint rules for evidence collection points include distance constraints between the evidence collection point and the potential hazard. Angle constraints and safety constraints: The evidence collection point must be outside the first-level prohibited area; Shooting parameter requirements: Number of photos ≥ 3, video length ≥ 30s, shooting accuracy; Based on the aforementioned evidence collection point constraint rules, potential evidence collection point areas M are circled in the GIS map; based on potential evidence collection point areas M, spatial stratified sampling + multi-constraint compliance verification algorithm is used to generate 3 to 5 candidate evidence collection points. S422, 3D raster map construction and attribute annotation: Based on the constraint data described in S421, the inspection area is divided into 1 ×1 ×1 The three-dimensional grid map marks impassable grids and labels the work site elevation and flight restriction altitude corresponding to each grid. S423, take the candidate evidence collection points obtained in S421 as the endpoints in sequence. Iterative path search is performed to obtain the corresponding candidate paths, and the optimal path is selected as the optimal route based on the comprehensive cost of the improved A* algorithm. S424 uses a Bézier curve to smooth the optimal route obtained from S423, requiring a turning angle ≤ 15°. To avoid frequent changes in direction by the drone; after smoothing, the flight path is re-verified to ensure it meets the constraint data. If it violates the constraint, adjustments are made in steps S421 to S423; if it meets the constraint, a flight path task package is generated.

9. The railway external environment inspection method according to claim 8, characterized in that, In step S421, the specific operation for generating candidate evidence collection points is as follows: The potential evidence collection point area M is segmented into multiple feature sub-regions based on its distance from the hazard, elevation difference, and proximity to the obstacle; random sampling is performed within each feature sub-region to obtain 6 to 8 initial candidate points; The constraint data and evidence collection point constraint rules are used to perform multi-constraint compliance verification, including distance, security, elevation, angle, and spatial constraints, to eliminate initial candidate points that violate regulations. Based on the principle of optimizing security costs, evidence collection and adaptation costs, and efficiency costs, 3 to 5 candidate evidence collection points are selected. The 3D coordinates of candidate evidence collection points are calibrated using the real-scene 3D model for subsequent iterative path search.

10. The railway external environment inspection method according to claim 9, characterized in that, Step S423 includes: Initialize the list of nodes to be evaluated, using the precise latitude and longitude coordinates of the drone hangar as the starting point. Add nodes to the list of nodes to be evaluated and calculate their total cost; initialize the list of evaluated nodes, initially empty; Search starting point The eight 3D neighboring nodes of the inaccessible grid are excluded, their comprehensive costs are calculated, and they are added to the list of nodes to be evaluated. The node with the lowest comprehensive cost is selected. Move to the list of evaluated nodes; continue searching for the 8 3D neighboring nodes of the node with the lowest total cost, excluding the inaccessible graticule, calculate its total cost, and add it to the list of nodes to be evaluated; Repeat the process of selecting the node with the lowest overall cost from the list of nodes to be evaluated. Move to the list of evaluated nodes; continue searching for the 8 3D neighboring nodes of the node with the lowest overall cost, excluding impassable grids, calculating their overall cost, and adding them to the list of nodes to be evaluated, until the endpoint is found among the neighboring nodes of the current node. At that time, the search ends, and the zigzag route is obtained; Compare the candidate broken line routes corresponding to multiple candidate evidence collection points, and select the route with the lowest overall cost of the improved A* algorithm corresponding to the candidate evidence collection point as the optimal route.

Images