A method for identifying damage of a highway guardrail by a UAV aerial photography

Abstract

The present application relates to the technical field of intelligent inspection of highway facilities, and particularly discloses a method for identifying damage of a highway isolation fence by aerial photography of a UAV, which first generates a flight route that is adapted to the linear distribution of the isolation fence based on GIS geographic information and a digital elevation model of the highway isolation fence; then controls the UAV to fly autonomously along the route, collects high-resolution sequence images containing the isolation fence posts and mesh at preset equidistant shooting points, and synchronously records positioning and orientation data; subsequently detects the post inclination angle and mesh damage area using a pre-trained deep learning model, and marks the abnormalities; finally fuses the positioning and orientation data with the identification results, calculates the three-dimensional geographic coordinates of the abnormal points through a space forward intersection algorithm, and generates an isolation fence abnormality report with geographic coordinates. The present application solves the problems of low efficiency and poor accuracy of traditional inspection, improves the accuracy of abnormality identification and positioning, and provides a precise basis for maintenance of the isolation fence.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for highway facilities, specifically a method for identifying damage to highway guardrails using drone aerial photography. Background Technology

[0002] Highway guardrails are crucial facilities for ensuring highway traffic safety, primarily used to prevent pedestrians, non-motorized vehicles, and small animals from entering highway traffic areas. Regular inspections are necessary to check for abnormalities such as tilted posts and damaged mesh panels to maintain their protective function. Current common guardrail inspection methods have significant shortcomings. Manual inspection requires staff to check each section on-site, which is not only inefficient and difficult to cover long stretches of highway guardrails, but also susceptible to factors such as complex terrain and inclement weather, leading to a high rate of missed or false detections. Traditional fixed monitoring equipment, such as cameras deployed along the route, has a fixed coverage area and blind spots, unable to flexibly adapt to the linear distribution characteristics of the guardrails, and struggles to comprehensively capture anomalies along the entire route. Early drone inspection solutions lacked precise flight path planning capabilities and did not incorporate GIS geographic information and digital elevation models of the guardrails, easily resulting in problems such as improper flight altitude and flight path deviation, making it difficult to capture images that completely and clearly cover the guardrail posts and mesh components. Furthermore, anomaly identification relies heavily on manual judgment, with low levels of intelligence, failing to quickly and accurately determine the spatial location of anomalies, thus affecting the timeliness and targeted nature of subsequent maintenance work. Existing inspection methods are no longer sufficient to meet the actual needs of efficient and accurate inspection of highway guardrails. Therefore, there is an urgent need for an optimized method for identifying anomalies in guardrails using drone aerial photography. Summary of the Invention

[0003] The purpose of this invention is to provide a method for identifying damage to highway guardrails using drone aerial photography, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying damage to highway guardrails using drone aerial photography includes the following steps: S1. Data Acquisition Flight Route Planning: Based on the GIS geographic information and digital elevation model of the highway guardrail, a terrain-following flight route adapted to the linear distribution of the guardrail is generated. The planning parameters of the terrain-following flight route include: the horizontal distance between the UAV and the guardrail, the photography angle, the flight altitude and the equidistant shooting points. The flight altitude is dynamically adjusted based on the digital elevation model to maintain a constant altitude relative to the ground. S2. Automated Image Data Acquisition: Control the UAV to fly autonomously along the terrain-following flight path, and acquire images of the fence at preset equidistant shooting points and set shooting angles, obtaining high-resolution sequence images including fence pillars and mesh; at the same time, record the UAV's positioning and attitude determination system data at each shooting point, including latitude and longitude, altitude, pitch angle, roll angle and yaw angle. S3. Intelligent Recognition and Anomaly Detection of Isolation Fence Components: Input the sequence of images into a pre-trained deep learning target recognition model, and simultaneously perform the following recognition and calculations: Target detection is performed on the fence posts in the image, the position of the posts is located, and the tilt angle of the posts relative to the direction of gravity is calculated based on the image coordinate system regression. If the tilt angle exceeds a preset threshold, it is marked as an abnormal tilt of the posts. Target detection is performed on the mesh area in the image, and pixel-level semantic segmentation is performed within the detected mesh area to identify the mesh damage area. If the area of ​​the damage area exceeds a preset threshold, it is marked as mesh damage abnormality. S4. 3D Geographic Calibration of Abnormal Results: The positioning and attitude determination system data recorded in step S2 is integrated with the identification results in step S3. The 3D geographic coordinates of the column tilting abnormality and the mesh damage abnormality are calculated by the spatial forward intersection algorithm. The identified abnormal status is calibrated to the geographic coordinate system to generate an isolation fence abnormality report with geographic coordinates. The report includes the abnormality type, location coordinates and severity.

[0005] As a preferred approach, in step S1, based on the GIS geographic information and digital elevation model of the highway guardrail, a terrain-following flight path adapted to the linear distribution of the guardrail is generated, including the following sub-steps: S1-1. Obtain GIS geographic information of highway guardrails and extract linear distribution centerline data of guardrails; S1-2. Based on the linear distribution centerline data, determine the preset horizontal distance between the UAV and the isolation fence, thereby generating a preliminary flight path parallel to the centerline of the isolation fence; S1-3. Based on the preset photography angle and reference flight altitude, and combined with digital elevation model data, perform elevation correction on the initial flight path and dynamically adjust the flight altitude to maintain a constant altitude of the UAV relative to the ground. S1-4. On the corrected flight path, equidistant shooting points are generated based on preset equidistant intervals, thus forming the final terrain-following flight path.

[0006] As a preferred embodiment, step S2 specifically includes the following sub-steps: S2-1. Load the terrain-following flight path data generated in step S1, control the UAV to fly autonomously along the path, and monitor the UAV's position information in real time to ensure that it follows the predetermined path. S2-2. When the UAV flies to the equidistant shooting point on the terrain-following flight path, the onboard camera equipment is automatically triggered based on the position monitoring results to capture images of the fence at a preset shooting angle, thereby obtaining a series of high-resolution sequential images to ensure that each image covers the fence posts and mesh area. S2-3. At the moment of each image acquisition, synchronously acquire and record the UAV's positioning and attitude determination system data, including latitude and longitude, altitude, pitch angle, roll angle and yaw angle, to form attitude and position information corresponding to each image; S2-4. The acquired high-resolution sequence images are spatiotemporally correlated and stored with the corresponding positioning and attitude determination system data to construct a structured dataset, providing input for subsequent intelligent identification and anomaly detection of the isolation fence components.

[0007] As a preferred approach, the sequence of images is input into a pre-trained deep learning target recognition model to detect the fence pillars in the images, locate the pillar positions, and calculate their tilt angle relative to the direction of gravity based on image coordinate system regression. If the tilt angle exceeds a preset threshold, it is marked as an abnormal tilt of the pillar, including the following sub-steps: S3-1. Using a pre-trained deep learning target detection model, target detection of isolation fence pillars is performed on the input sequence images, and the bounding box coordinates of each pillar in the image are output to locate the position of the pillar. S3-2. Based on the bounding box coordinates, crop out the column region sub-image from the original image and use it as input for subsequent tilt angle calculation; S3-3. Perform edge detection and contour extraction on the sub-image of the column region to identify the main axis of the column; S3-4. In the image coordinate system, calculate the angle between the main axis and the direction of gravity. The direction of gravity is determined based on the vertical direction of the image, thereby regressing to obtain the tilt angle of the column relative to the direction of gravity. S3-5. Compare the calculated tilt angle with the preset threshold. If the tilt angle exceeds the threshold, mark the column as tilting abnormally and record the abnormal information.

[0008] As a preferred embodiment, in step S3, target detection is performed on the mesh region in the image, and pixel-level semantic segmentation is performed within the detected mesh region to identify the mesh damage area. If the area of ​​the damage area exceeds a preset threshold, it is marked as mesh damage abnormality, including the following sub-steps: S3-6. Using a pre-trained deep learning target detection model, target detection is performed on the input sequence images in the isolation fence region, and the bounding box coordinates of each mesh region in the image are output. S3-7. Based on the bounding box coordinates, extract the mesh region sub-images from the original image and use them as input for subsequent semantic segmentation; S3-8. Using a pre-trained pixel-level semantic segmentation model, the sub-image of the mesh region is segmented, and the pixels are classified into complete meshes and damaged meshes, thereby generating a binary mask for the damaged mesh region. S3-9. Based on the binary mask, calculate the pixel area of ​​the damaged area of ​​the mesh, and combine the camera parameters and shooting distance to convert the pixel area into the actual physical area; S3-10. Compare the calculated actual physical area with the preset threshold. If the damaged area exceeds the threshold, mark the mesh as abnormal and record the abnormal information.

[0009] As a preferred approach, the positioning and attitude determination system data recorded in step S2 is integrated with the identification results in step S3. The three-dimensional geographic coordinates of the column tilt anomaly and the mesh damage anomaly are calculated using a spatial forward intersection algorithm. The identified anomaly states are then mapped to the geographic coordinate system to generate an isolation fence anomaly report with geographic coordinates. The report includes the anomaly type, location coordinates, and severity, and includes the following sub-steps: S4-1. Extract abnormal information from the identification results of step S3, including the abnormality type and the position coordinates of the abnormal point in the image; wherein, for the column tilting abnormality, the position coordinates of the abnormal point are the coordinates of the center point of the bottom of the column in the image; for the mesh damage abnormality, the position coordinates of the abnormal point are the coordinates of the center point of the damaged area in the image. S4-2. Based on the image identifier of the abnormal point in step S3, obtain the camera position and attitude data of the corresponding image from the positioning and attitude system data recorded in step S2, including latitude and longitude, altitude, pitch angle, roll angle and yaw angle. S4-3. For each anomaly, based on the image sequence and geographic information, find at least two other images containing the anomaly from the sequence images collected in step S2, and obtain the location coordinates of the anomaly in these images and its corresponding camera position and attitude data. S4-4. Using the multiple camera position and attitude data obtained in step S4-3 and the position coordinates of the anomaly points in the corresponding images, calculate the three-dimensional geographic coordinates of the anomaly points using the spatial forward intersection algorithm. S4-5. Associate the three-dimensional geographic coordinates calculated in step S4-4 with the anomaly type and severity in step S3, and label the anomaly status to the geographic coordinate system. S4-6. Summarize all calibrated abnormal states and generate an isolation fence anomaly report. The report includes the anomaly type, three-dimensional geographic coordinates, and severity.

[0010] As can be seen from the above technical solution provided by the present invention, the beneficial effects of the method for identifying damage to highway guardrails by drone aerial photography provided by the present invention are: The flight path planning is accurate and reliable. Based on the GIS geographic information and digital elevation model of the highway guardrail, the flight path is generated to simulate the terrain. The flight altitude can be dynamically adjusted to adapt to the terrain undulations. This ensures that the images captured by the UAV completely cover the guardrail posts and mesh area, avoiding image loss or blurring caused by flight path deviation or terrain influence, and providing a high-quality data foundation for subsequent identification. The image acquisition is highly automated, controlling the drone to fly autonomously along a preset route and automatically acquire high-resolution images at equidistant shooting points. At the same time, it records positioning and attitude data and builds a structured dataset, which greatly improves the acquisition efficiency compared with manual inspection and can ensure the accurate correlation between images and spatial location information, reducing human operation errors. It has high anomaly recognition accuracy. It uses a pre-trained deep learning model to intelligently detect the fence posts and mesh panels. It can accurately locate the posts and calculate the tilt angle, identify the damaged areas of the mesh panels and convert them into actual areas. It can effectively distinguish between normal and abnormal states, reduce the false judgment rate and false negative rate of manual identification, and realize the automated and accurate detection of fence anomalies. The anomaly identification and maintenance guidance are clear. It integrates positioning and attitude determination data with identification results, calculates the three-dimensional geographic coordinates of the anomaly point through a spatial forward intersection algorithm, calibrates the anomaly status to the geographic coordinate system, and generates a report containing the anomaly type, location coordinates, and severity. This provides maintenance personnel with accurate anomaly location guidance, avoids blind troubleshooting, and improves the efficiency and targeting of fence maintenance operations. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the steps of a method for identifying damage to highway guardrails using drone aerial photography according to the present invention. Detailed Implementation

[0012] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0013] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0014] like Figure 1As shown, this embodiment of the invention provides a method for identifying damage to highway guardrails using drone aerial photography, comprising the following steps: S1. Data Acquisition Flight Route Planning: Based on the GIS geographic information and digital elevation model of the highway guardrail, a terrain-following flight route adapted to the linear distribution of the guardrail is generated. The planning parameters of the terrain-following flight route include: the horizontal distance between the UAV and the guardrail, the photography angle, the flight altitude and the equidistant shooting points. The flight altitude is dynamically adjusted based on the digital elevation model to maintain a constant altitude relative to the ground. S2. Automated Image Data Acquisition: Control the UAV to fly autonomously along the terrain-following flight path, and acquire images of the fence at preset equidistant shooting points and set shooting angles, obtaining high-resolution sequence images including fence pillars and mesh; at the same time, record the UAV's positioning and attitude determination system data at each shooting point, including latitude and longitude, altitude, pitch angle, roll angle and yaw angle. S3. Intelligent Recognition and Anomaly Detection of Isolation Fence Components: Input the sequence of images into a pre-trained deep learning target recognition model, and simultaneously perform the following recognition and calculations: Target detection is performed on the fence posts in the image, the position of the posts is located, and the tilt angle of the posts relative to the direction of gravity is calculated based on the image coordinate system regression. If the tilt angle exceeds a preset threshold, it is marked as an abnormal tilt of the posts. Target detection is performed on the mesh area in the image, and pixel-level semantic segmentation is performed within the detected mesh area to identify the mesh damage area. If the area of ​​the damage area exceeds a preset threshold, it is marked as mesh damage abnormality. S4. 3D Geographic Calibration of Abnormal Results: The positioning and attitude determination system data recorded in step S2 is integrated with the identification results in step S3. The 3D geographic coordinates of the column tilting abnormality and the mesh damage abnormality are calculated by the spatial forward intersection algorithm. The identified abnormal status is calibrated to the geographic coordinate system to generate an isolation fence abnormality report with geographic coordinates. The report includes the abnormality type, location coordinates and severity.

[0015] In this embodiment, step S1 involves combining the GIS geographic information and digital elevation model of the highway guardrail to plan a terrain-following flight path that matches the linear distribution of the guardrail. This ensures that subsequent image acquisition by the UAV can completely and clearly cover the guardrail area, avoiding data loss or image quality issues caused by terrain undulations or flight path deviations. This provides an accurate and high-quality image data foundation for subsequent guardrail component identification and anomaly detection. The detailed steps are as follows: Step S1-1: Obtain GIS geographic information of highway guardrails and extract linear distribution centerline data: GIS geographic information acquisition: Retrieve the GIS geographic information corresponding to the target highway guardrail from the highway facility geographic information system database of the highway management department; this information must include the actual location, overall orientation, total length, segment nodes, and basic geographic attributes of the surrounding terrain of the guardrail to ensure that the information can fully reflect the spatial distribution characteristics of the guardrail; Linear distribution centerline data extraction: The acquired fence GIS information is processed using the line feature analysis tool of professional geographic information processing software (such as ArcGIS). If the fence is stored as a polygon feature, the geometric centerline of the polygon feature is calculated using the polygon feature centerline generation function. If it is stored as a line feature, the lines are directly smoothed and simplified to remove local minor offset interference. Through the above operations, linear distribution centerline data that accurately reflects the overall orientation of the fence is extracted. This data will serve as the core benchmark for subsequent flight path planning. Step S1-2: Determine the lateral horizontal distance based on the linear distribution centerline to generate a preliminary flight path: Preset horizontal distance determination: Based on the optical parameters of the drone's onboard camera (such as focal length and field of view), the preset shooting angle, and the actual width of the fence, the preset horizontal distance between the drone and the fence is calculated and determined. This distance must meet two core requirements: first, ensure that the captured image can completely include the fence's pillars and mesh area without edge cropping; second, ensure that the pixel ratio of the fence components in the image is moderate, facilitating subsequent intelligent recognition. Typically, this preset horizontal distance is set within the range of 3 to 5 meters. Preliminary flight path generation: Based on the linear distribution centerline extracted in step S1-1, in the geographic information processing software, the centerline is offset by a preset horizontal distance away from the fence along a direction perpendicular to the centerline, forming a new line feature; this line feature is the preliminary flight path parallel to the linear distribution centerline of the fence, and its direction is consistent with the fence, ensuring that the UAV flight trajectory can follow the fence and avoid missing the shooting area due to flight path deviation; Steps S1-3: Correct the initial flight path by combining the flight altitude referenced by the photographic angle with the digital elevation model: Reference flight altitude and camera angle settings: Based on the resolution of the UAV's onboard camera, the desired image ground sampling distance (i.e., the actual ground distance represented by a unit pixel), and the recognition accuracy requirements of the fence components, a reference flight altitude is set. Typically, the reference flight altitude is set between 5 and 8 meters to ensure that the image clearly presents the outline details of the pillar and the texture features of the mesh. At the same time, the camera angle is set, generally using a downward angle of 30 to 45 degrees. This angle can clearly capture both the side profile of the pillar and the planar distribution of the mesh, avoiding the pillar obscuring the mesh due to an excessively large angle or the mesh details becoming blurred due to an excessively small angle. Digital Elevation Model Data Application and Elevation Correction: Obtain digital elevation model data corresponding to the UAV flight area, which contains ground elevation information for each point in the area; overlay and analyze the preliminary flight path generated in steps S1-2 with the digital elevation model data in geographic information processing software, and obtain the ground elevation value corresponding to each point on the preliminary flight path through the software's elevation interpolation calculation function; calculate the flight altitude value to be adjusted for each point based on the preset benchmark flight altitude, that is, the actual flight altitude is equal to the benchmark flight altitude plus the difference between the ground elevation of the point and the average ground elevation of the area, and dynamically adjust the flight altitude of each point on the preliminary flight path through this calculation, so that the height of the UAV relative to the ground remains constant during the flight, avoiding excessively high or low flight altitudes due to terrain undulations (such as hills and depressions), and ensuring the clarity and consistency of image acquisition; Steps S1-4: Generate equidistant shooting points on the corrected path to form the final terrain-following flight path: Preset equidistant interval determination: Based on the image resolution of the UAV's onboard camera equipment, the overlap rate requirements of adjacent images, and the continuity characteristics of the isolation fence components, the preset interval of equidistant shooting points is determined; to ensure that subsequent image stitching is complete and that adjacent images can be effectively correlated, the overlap rate of images corresponding to adjacent shooting points needs to reach 30% to 50%; combined with the ground coverage width of the image at the reference flight altitude, the preset equidistant interval is calculated, which is usually set in the range of 2 to 4 meters; Equidistant shooting point generation and final flight path integration: Using the point feature batch generation tool of the flight path planning software, equidistant shooting points are generated sequentially at preset equidistant intervals along the corrected flight path. Each shooting point corresponds to a specific image acquisition location coordinate. The corrected flight path is then linked and integrated with all the generated equidistant shooting points, and parameters such as flight altitude and shooting angle corresponding to each shooting point are labeled to form the final terrain-following flight path. This flight path is suitable for both the linear distribution of the guardrail and the terrain undulations, while also clearly defining the specific location and parameters of each image acquisition, providing a precise execution basis for subsequent automated image acquisition by the UAV.

[0016] In this embodiment, step S2 involves automatically controlling a drone to fly stably along a preset terrain-following flight path, acquiring high-resolution sequential images of the fence at designated equidistant shooting points according to set parameters, and simultaneously recording the drone's positioning and attitude data. This achieves precise correlation between images and spatial location information, providing a structured, high-quality basic dataset for subsequent intelligent identification of fence components, anomaly detection, and 3D geographic calibration. The detailed steps are as follows: Step S2-1: Load terrain-following flight path data and control the UAV for autonomous flight and position monitoring: Loading terrain-following flight path data: The terrain-following flight path data generated in step S1 (including the three-dimensional coordinates, flight altitude, shooting point position and other parameters of each point on the path) is imported into the UAV ground station system through the data transmission interface. The ground station software parses and verifies the data. After confirming that the data format is correct and the parameters are not missing, the path data is sent to the UAV flight control system. Autonomous flight control of UAVs: After receiving the flight path data, the UAV flight control system combines its own power system (motors, propellers) and navigation module (GNSS positioning, inertial measurement unit) to perform autonomous flight according to the preset path and altitude parameters. During flight, the flight control system adjusts the motor speed in real time to control the flight speed and attitude of the UAV, ensuring that the UAV moves smoothly along the flight path and avoiding unstable flight states such as sudden acceleration and sharp turns. Real-time position monitoring and path correction: The ground station system receives the UAV's current position data (latitude, longitude, and altitude) in real time via wireless communication links (such as 4G, 5G, or dedicated data radios) and compares it with the position parameters of the preset flight path. If the deviation between the UAV's current position and the preset flight path exceeds a set threshold (usually the lateral deviation does not exceed 0.5 meters and the longitudinal deviation does not exceed 0.3 meters), the ground station system immediately sends a correction command to the flight control system. The flight control system adjusts the flight direction and altitude according to the command to bring the UAV back to the preset flight path and ensure the accuracy of the flight path. Step S2-2: Automatically trigger the camera device to acquire a sequence of high-resolution images at equidistant shooting points: Shooting point arrival determination: During the flight of the drone, the flight control system matches the current position coordinates with the coordinates of the preset equidistant shooting points in real time; when the distance between the two is less than the determination threshold (usually 0.2 meters), the flight control system determines that the drone has arrived at the equidistant shooting point and generates a shooting trigger signal; Automatic triggering and parameter control of camera equipment: The shooting trigger signal is transmitted to the onboard camera equipment through the internal communication bus of the UAV. The triggering equipment starts image acquisition according to preset parameters. The shooting angle (such as 30 to 45 degrees of downward angle), shutter speed, aperture size, ISO sensitivity and other parameters of the camera equipment are all set in advance and remain stable during the acquisition process to ensure the exposure consistency of the sequence of images. High-resolution image acquisition and coverage guarantee: The image resolution acquired by the camera equipment is no less than 4000×3000 pixels (4K level), ensuring that details such as the pillar texture and mesh of the fence are clearly distinguishable in the image; at the same time, through preset shooting angle and horizontal distance control, the coverage of each image meets the following requirements: the top and bottom edges of the fence pillars are included, and the left and right edges cover the complete mesh area between two adjacent pillars, avoiding the omission of fence components or image edge cropping problems, forming a continuous high-resolution sequence of images; Step S2-3: Synchronously collect and record data from the UAV positioning and attitude determination system: Positioning and attitude determination data acquisition timing: At the same instant that the airborne camera completes the acquisition of a single image (with a time deviation of no more than 10 milliseconds), the flight control system triggers the positioning and attitude determination system (composed of a GNSS module and an inertial measurement unit) to acquire data, ensuring that the acquired data is completely synchronized with the corresponding image in time, and avoiding subsequent data correlation deviations due to time differences; Positioning and attitude determination data acquisition: The positioning and attitude determination system acquires the following data: latitude and longitude (accuracy to 6 decimal places, corresponding to a positioning error of approximately 0.1 meters on the ground), altitude (accuracy to 0.1 meters), pitch angle (forward and backward tilt angle of the UAV fuselage, accuracy to 0.1 degrees), roll angle (left and right tilt angle of the UAV fuselage, accuracy to 0.1 degrees), and yaw angle (angle of the UAV's nose pointing, accuracy to 0.1 degrees). This data is transmitted in real time to the flight control system's storage module through an internal data interface. Data integrity verification: The flight control system performs integrity verification on each acquired positioning and attitude data, checking for missing data or outliers (such as latitude and longitude exceeding the reasonable range or angle values ​​exceeding ±90 degrees). If data anomalies are found, the flight control system immediately re-triggers acquisition until complete and normal data is obtained, ensuring that each image corresponds to a set of valid positioning and attitude data. Step S2-4: Construct a structured dataset by storing the spatiotemporal association between images and localization / pose determination data: Spatiotemporal correlation of data: Each high-resolution image is assigned a unique image identifier (such as an encoding containing the acquisition time, drone number, and shooting point number), and the same identifier is assigned to the corresponding acquired positioning and attitude data; the image data and positioning and attitude data are associated one-to-one through the identifiers to achieve precise matching of the two in time (the same acquisition moment) and space (the same shooting position). Data classification and storage: A hierarchical storage architecture is used to store the associated data; high-resolution image files (such as JPEG and TIFF formats) are stored on the UAV's onboard high-capacity storage device (such as SSD solid-state drive) or transmitted to the ground storage server in real time via wireless link; positioning and attitude data (stored in the form of structured tables, including fields such as image identifier, latitude and longitude, altitude, pitch angle, roll angle, yaw angle, and acquisition time) are stored in a relational database (such as MySQL) for easy subsequent data query and retrieval; Structured dataset construction: The stored image and positioning / attitude data are organized and sorted according to the acquisition time and shooting point location to form a structured dataset containing three-dimensional information of "image-position-attitude". The dataset also needs to include metadata information (such as acquisition date, drone model, camera equipment parameters, and flight path name) to ensure that subsequent steps (such as intelligent recognition in S3 and three-dimensional calibration in S4) can quickly call up the required data and improve overall processing efficiency.

[0017] In this embodiment, step S3 uses a pre-trained deep learning target recognition model to intelligently process the acquired high-resolution sequence images of the fence, simultaneously detecting tilt anomalies in the fence posts and damage anomalies in the mesh panels. It accurately extracts key information such as anomaly type and degree, providing clear and reliable anomaly data support for the subsequent three-dimensional geo-calibration of the anomaly results in step S4, thus achieving automated and high-precision identification of fence anomalies. The detailed steps are as follows: Step S3-1: Locate the isolation fence posts using a pre-trained deep learning object detection model: Pre-trained model selection and loading: Select a pre-trained deep learning model suitable for small object detection. This model needs to be trained on a large-scale isolation fence pillar sample dataset. The samples need to cover pillar images under different lighting conditions, different weather conditions, and different terrain environments to ensure that the model has a stable pillar recognition capability. Load the trained model weight file into the inference framework of the image processing server to complete the model initialization and ensure that the model can run normally. Sequence Image Input and Inference: The high-resolution sequence images from the structured dataset constructed in step S2 are input one by one into the initialized target detection model according to the image acquisition order; the model extracts feature information from the images step by step through neural network layers such as convolution and pooling, performs target prediction on the feature maps, and finally outputs the bounding box coordinates of all isolation fence posts in each image. The coordinates must clearly reflect the position range of the posts in the image, including the image coordinate values ​​of the upper left and lower right corners; Column location confirmation: The bounding box results output by the model are filtered to remove bounding boxes with a confidence level below 0.8 to avoid misidentifying interfering objects in the background as columns; bounding boxes with a confidence level that meets the requirements are retained, and the corresponding column locations are marked in the original image to ensure that each marked position accurately corresponds to the actual fence column, thereby achieving precise positioning of the column in the image. Step S3-2: Cropping the column region sub-image based on bounding box coordinates: Determining the cropping range: Based on the coordinates of the single column bounding box output in step S3-1, in order to avoid missing the details of the column edge during the cropping process, the bounding box is extended outward by 5 to 10 pixels in each of the four directions of top, bottom, left and right. The extension range needs to be flexibly adjusted according to the image resolution. The core principle is not to introduce too much irrelevant background, but to retain only the column and a small amount of surrounding environment to reduce interference in subsequent processing. Sub-image cropping execution: Using cropping functions from a professional image processing library, sub-images corresponding to individual columns are separated from the original high-resolution image according to the defined cropping range; each sub-image contains only one isolation fence column, ensuring that subsequent calculations of the column's tilt angle are not affected by other columns or complex backgrounds; Sub-image naming and association: Assign an identifier to each cropped column sub-image that is consistent with the original image, and add the column number after the identifier, such as adding col1, col2, etc. after the identifier of the original image; through this naming rule, it is ensured that the sub-image can be associated with the original image and the positioning and orientation data corresponding to the original image. If it is necessary to trace abnormal information later, the original data can be quickly located. Step S3-3: Perform edge detection and contour extraction on the column region sub-image: Sub-image preprocessing: First, the sub-image of the column region is converted to grayscale to eliminate the interference of color channels on subsequent detection. Then, Gaussian filtering is used to smooth the grayscale image. The convolution kernel size is set to 5×5. The filtering removes noise in the image, such as bright and dark spots caused by uneven lighting and small particle interference during image acquisition, thereby improving the accuracy of subsequent edge detection. Edge detection execution: The Canny edge detection algorithm is used for edge extraction. Reasonable high and low thresholds are set, usually the high threshold is set to 200 and the low threshold is set to 100. The algorithm calculates the gray-level gradient of the image pixels, identifies the regions where the pixel gray-level changes abruptly, and generates a binary edge image containing only the edge of the pillar, clearly presenting the outline of the pillar, laying the foundation for subsequent contour extraction. Main axis identification: Use contour extraction algorithm to extract all contours of the column from the edge image, calculate the perimeter and area of ​​each contour, and filter out the contour with the largest area and the longest perimeter. This contour is the main contour of the column, and other small interfering contours are excluded. The selected main contour is fitted with a straight line by the least squares method. The fitted straight line is the main axis of the column, which can accurately reflect the overall direction of the column in the image. Step S3-4: Calculate the tilt angle of the column's main axis relative to the direction of gravity: Gravity direction definition: In the image coordinate system, the gravity direction is consistent with the vertical direction of the image, that is, along the positive y-axis of the image coordinate system, which is the direction from the bottom of the image to the top of the image. This direction is used as the reference direction for calculating the tilt angle. Angle calculation method: Construct direction vectors for the main axis of the column and the direction of gravity, respectively, and calculate the angle between the two direction vectors by vector dot product; during the calculation process, it is necessary to ensure the accuracy of the vector direction to avoid errors in angle calculation due to confusion of directions; Tilting angle regression: The calculated included angle is numerically processed and retained to one decimal place. The final value is the tilt angle of the column relative to the direction of gravity. If the angle is 0 degrees, it means that the column is in a vertical state. If the angle is greater than 0 degrees, it means that the column is tilted to the side that deviates from the direction of gravity. The larger the angle value, the more serious the tilt of the column. Step S3-5: Compare the tilt angle with the preset threshold to mark abnormal tilt of the column: Preset threshold determination: Based on the maintenance standards and industry specifications of highway guardrails, and with reference to relevant documents such as the "Design Specifications for Highway Traffic Safety Facilities", a preset threshold for the tilt angle of the posts is established. Under normal circumstances, this threshold is set to 5 degrees. When the tilt angle of the posts exceeds 5 degrees, the stability of the posts will decrease significantly, which may affect the overall protective function of the guardrail and should be judged as abnormal. Anomaly detection and marking: The tilt angle of a single column calculated in step S3-4 is compared with the preset 5-degree threshold one by one; if the tilt angle is less than or equal to 5 degrees, the column is determined to be in normal condition; if the tilt angle is greater than 5 degrees, the column is determined to have a tilt anomaly, and the column is marked with a red border in the column sub-image and the original image. At the same time, the anomaly information is recorded, including the image identifier, column number, tilt angle value, etc. Temporary storage of abnormal information: All marked column tilt abnormal information is organized according to the acquisition order of the original images and stored in a temporary database table; the fields of the database table must include the abnormality type, i.e., column tilt, associated image identifier, the bounding box coordinates of the column in the original image, and the tilt angle, to ensure that the abnormal data can be quickly queried and traced in subsequent steps. Step S3-6: Locate the isolation fence patch region using a pre-trained deep learning object detection model: Mesh detection model loading: Select a pre-trained deep learning model suitable for large-area region detection. This model needs to be trained on a dataset of isolation fence mesh samples with different degrees of damage and different installation angles to ensure that the model can accurately identify the mesh region; Load the trained model weight file into the inference framework to complete the model initialization and ensure that the model can stably process sequential images; Sequence Image Mesh Detection: The high-resolution sequence images from step S2 are input one by one into the initialized mesh detection model according to the acquisition order; the model outputs the bounding box coordinates of all isolation fence mesh regions in each image through feature extraction and target prediction. The coordinates must clearly reflect the position range of the mesh region in the image, and each bounding box must completely surround a single mesh, usually the mesh region between two adjacent posts. Mesh region filtering: The mesh bounding box results output by the model are filtered. First, bounding boxes with a confidence score below 0.75 are removed to avoid misidentifying similar mesh structures in the background as meshes. Second, the overlap between the bounding box and the adjacent column bounding boxes is checked. Under normal circumstances, there is a reasonable overlap between the mesh and the column, with the overlapping area accounting for 5% to 10% of the area of ​​the mesh bounding box. The overlap check further ensures the accuracy of the mesh region positioning. Step S3-7: Extract sub-images of mesh regions based on bounding box coordinates: Extraction range determination: Based on the coordinates of the single mesh bounding box output in steps S3-6, considering the importance of the connection between the mesh and the column for damage detection, the left and right sides of the bounding box are extended by 3 to 5 pixels towards the adjacent column to ensure that the extracted sub-image can completely preserve the connection between the mesh and the column and avoid loss of connection area information due to cropping. Sub-image extraction of mesh: Using the cropping function of the image processing library, according to the determined extraction range, the sub-images corresponding to each mesh are separated from the original high-resolution image; each sub-image contains only one complete isolation fence mesh and includes the connection part between the mesh and the post, ensuring that subsequent semantic segmentation can fully cover the mesh area and not miss any potential damage points; Sub-image association storage: Assign an identifier to each extracted mesh sub-image that is consistent with the original image, and add the mesh sequence number after the identifier, such as adding mesh1, mesh2, etc. after the identifier of the original image; store the mesh sub-image with the identifier in a dedicated image folder, and establish the association relationship between the sub-image and the original image, as well as the corresponding positioning and orientation data of the original image, to facilitate the subsequent tracking of abnormal information; Step S3-8: Generate binary masks for damaged areas of the mesh using a pre-trained pixel-level semantic segmentation model: Semantic segmentation model selection and loading: Select a pre-trained pixel-level semantic segmentation model suitable for fine segmentation tasks. This model needs to be trained on a sample dataset containing various scenarios such as complete mesh, mesh holes and damage, and mesh tearing to ensure that the model can accurately distinguish between complete and damaged areas of the mesh; Load the trained model weight file into the inference framework to complete model initialization; Mesh sub-image segmentation inference: The mesh region sub-images extracted in steps S3-7 are input sequentially into the initialized semantic segmentation model; the model extracts the feature information of the mesh sub-images through the encoder, and then restores the spatial resolution of the image through the decoder, classifying each pixel in the image. The classification result contains only two categories: complete mesh pixels and damaged mesh pixels. Binary mask generation: Based on the pixel classification results output by the model, a binary mask image of the damaged area of ​​the mesh is generated; in the binary mask image, pixels of the intact mesh correspond to black pixels with a pixel value of 0; pixels of the damaged mesh correspond to white pixels with a pixel value of 255; the shape, location and range of the damaged area of ​​the mesh can be intuitively presented through the binary mask, providing a clear basis for subsequent calculation of the damaged area; Step S3-9: Calculate the actual physical area of ​​the damaged area of ​​the wire mesh: Calculation of pixel area of ​​damaged region: Use the pixel counting function of the image processing library to count the pixels of the binary mask image generated in step S3-8, and count the total number of white pixels. This number is the pixel area of ​​the damaged area of ​​the mesh. During the counting process, it is necessary to ensure that no white pixels are missed to avoid inaccurate subsequent area calculation due to counting errors. Actual physical area conversion: The area conversion is carried out by combining the camera parameters and shooting distance during the drone shooting; the camera parameters include the camera sensor size and image resolution, and the physical size of a single pixel can be calculated by the sensor size and resolution; the shooting distance is obtained from the positioning and attitude data recorded in step S2; at the same time, combined with the inherent parameters of the camera, namely the focal length, the pixel area of ​​the damaged area is converted into the actual physical area through the comprehensive calculation of the above parameters. Area result processing: The calculated actual physical area is numerically processed and retained to two decimal places to ensure that the area accuracy meets the maintenance requirements of the fence. It can accurately distinguish between damaged areas of different sizes. For example, it can clearly distinguish the difference between 0.1 square meters and 0.2 square meters of damage, providing accurate data for subsequent anomaly judgment. Step S3-10: Compare the damaged area with the preset threshold to mark abnormal mesh damage. Determining the preset threshold for damaged area: Referring to the maintenance specifications of highway guardrails and considering the impact of mesh damage on the protective function of the guardrail, a preset threshold for the damaged area of ​​the mesh is established. Under normal circumstances, this threshold is set at 0.5 square meters. When the damaged area of ​​the mesh exceeds 0.5 square meters, the mesh cannot effectively prevent pedestrians and small animals from entering the highway, and the protective function of the guardrail will be greatly reduced, which needs to be judged as abnormal. Mesh Damage Anomaly Judgment: The actual physical area of ​​damage to a single mesh panel calculated in step S3-9 is compared one by one with the preset threshold of 0.5 square meters. If the damaged area is less than or equal to 0.5 square meters, the mesh panel is determined to be in normal condition, and such minor damage will not have a significant impact on the overall protective function of the fence. If the damaged area is greater than 0.5 square meters, the mesh panel is determined to have an anomaly. The mesh panel is marked with a yellow border in the mesh panel sub-image and the original image, and the anomaly information, including image identifier, mesh panel number, and damaged area value, is recorded. Anomaly Information Summary: All marked mesh damage anomaly information is organized according to the acquisition order of the original images, merged with the column tilt anomaly information obtained in step S3-5, and stored in a unified anomaly information database table; the fields of the database table must include the anomaly type, i.e., mesh damage, associated image identifier, bounding box coordinates of the mesh in the original image, and damage area, to ensure that complete and accurate anomaly data is provided for the three-dimensional geocalibration of the anomaly results in the subsequent step S4.

[0018] In this embodiment, step S4 involves fusing the UAV positioning and attitude data recorded in step S2 with the fence anomaly identification results from step S3, and using a spatial forward intersection algorithm to calculate the three-dimensional geographic coordinates of the anomaly point. This accurately pinpoints the anomaly to the geographic coordinate system, ultimately generating a fence anomaly report containing the anomaly type, location coordinates, and severity. This provides precise spatial location information for subsequent maintenance of highway fences, enabling the location and traceability of anomalies. The detailed steps are as follows: Step S4-1: Extract anomaly information and the coordinates of the anomaly points in the image: Anomaly information extraction: Filter and extract all abnormal data one by one from the identification results of step S3, clarify the type of each anomaly, that is, distinguish between column tilting anomaly and mesh damage anomaly; at the same time, record the associated information corresponding to each anomaly, including the original image identifier where the anomaly is located, to ensure that the corresponding image data can be traced in the future. Anomaly point coordinate determination: For different types of anomalies, determine the coordinates of the anomaly points in the image; for column tilting anomalies, use the column bounding box located in step S3 as a reference to calculate the coordinates of the geometric midpoint of the bottom edge of the bounding box. This midpoint is the image coordinate of the column tilting anomaly point, which can accurately represent the actual bottom position of the column in the image; for mesh damage anomalies, use the binary mask of the mesh damage area generated in step S3 as a basis to calculate the coordinates of the geometric center point of the damage area. This center point is the image coordinate of the mesh damage anomaly point, which can reflect the core position of the damage area; Coordinate recording and association: The image coordinates of the determined anomaly points are bound and recorded with the corresponding anomaly type and the image identifier of the anomaly, forming a basic information table of anomaly points. This ensures that each anomaly point can be accurately associated with the original image and anomaly type, avoiding information confusion. Step S4-2: Obtain the corresponding camera position and pose data based on the image identifier: Image identifier matching: Using the image identifier of the outlier recorded in step S4-1 as an index, a matching query is performed in the structured dataset constructed in step S2; this structured dataset stores the identifiers of all images and their corresponding localization and orientation system data, and the relevant data entries of the target image can be quickly located through identifier matching. Camera position data acquisition: Extract camera position data from the matched data entries, including the latitude, longitude and altitude of the UAV at the moment the image was acquired; these data directly reflect the spatial position of the UAV when the image was acquired, and are the basic reference for subsequent calculation of the three-dimensional coordinates of anomalies; Camera attitude data acquisition: Simultaneously extract camera attitude data from the data entries, including the UAV's pitch, roll, and yaw angles; these data describe the camera's shooting angle and orientation when acquiring images, which is crucial for accurately calculating the spatial location of anomalies, and it is necessary to ensure that the data is extracted completely without omission; Data verification: The extracted camera position and attitude data are verified a second time with the corresponding anomaly point basic information to confirm that the data and the image where the anomaly point is located are completely matched, so as to avoid the deviation of subsequent 3D coordinate calculation due to data matching errors. Step S4-3: Find other images and corresponding data containing the same anomaly: Anomaly coverage determination: Based on the camera position data (latitude, longitude, and altitude) and image acquisition parameters (such as shooting angle and field of view) of the image where the anomaly is located obtained in step S4-2, and combined with the linear distribution characteristics of the highway guardrail, determine which adjacent images may cover the anomaly; usually, images captured by adjacent acquisition points will have a certain overlapping area, and the anomaly is likely to appear in the adjacent images. Target image screening: In the high-resolution sequence of images in step S2, adjacent images that may cover the anomaly point are screened according to the image acquisition order. Generally, at least two images that meet the conditions need to be screened. During the screening process, the geographical coverage of the images can be combined, and the latitude and longitude of the camera can be compared to determine whether the image overlaps with the area where the anomaly point is located. Anomaly point coordinate extraction: For each selected target image, refer to the method in step S4-1 to find the location coordinates of the same anomaly point in the image; if the anomaly point in the target image has been identified in step S3, its coordinates are directly extracted; if it has not been identified individually, the coordinates are located in the image and calculated based on the characteristics of the anomaly point (such as the outline of the broken mesh in the shape of a column). Corresponding data acquisition: For each target image, repeat step S4-2 to acquire its corresponding camera position data (latitude, longitude, and altitude) and attitude data (pitch angle, roll angle, and yaw angle), and bind and record them with the coordinates of the outlier points in the image to form multiple sets of "image-outlier point coordinates-camera data" correspondences; Step S4-4: Calculate the 3D geographic coordinates of the anomaly points using the spatial forward intersection algorithm. Algorithm data preparation: Organize the multiple sets of camera position data, camera attitude data and corresponding outlier coordinates in the images obtained in step S4-3 to ensure that the format of each set of data is consistent and meets the input requirements of the spatial forward intersection algorithm; usually, it is necessary to convert the camera position to coordinates in the geographic coordinate system, convert the camera attitude data to the angle parameters required by the algorithm, and convert the outlier image coordinates to standardized coordinates in the image coordinate system. Algorithm execution process: The spatial forward intersection algorithm is started, and multiple sets of prepared data are input into the algorithm; the algorithm constructs projection rays from different camera perspectives and uses the intersection relationship of multiple sets of projection rays to calculate the three-dimensional coordinates of the anomaly in the real geographic space; this process takes into account the differences in camera position and attitude to ensure that the calculation results can accurately reflect the actual geographic location of the anomaly. Coordinate Result Verification: The reasonableness of the 3D geographic coordinates calculated by the algorithm is verified. The verification methods include comparing the digital elevation model data of the area where the anomaly point is located to determine whether the calculated altitude matches the terrain of the area; comparing the coordinates calculated from adjacent images of the anomaly point to determine whether the results are within a reasonable error range (usually the longitude and latitude errors do not exceed 0.5 meters, and the altitude error does not exceed 0.3 meters). If the verification passes, the coordinate results are retained; if they fail, the data is re-examined or more target image data is added before recalculation. Step S4-5: Associate the 3D coordinates with anomaly information and calibrate them to the geographic coordinate system: Anomaly information is associated with 3D coordinates: The 3D geographic coordinates (longitude, latitude, and altitude) of the anomaly points verified in step S4-4 are bound to the anomaly type (pillar tilt or mesh damage) extracted in step S4-1 and the anomaly severity information recorded in step S3; the anomaly severity is determined according to the judgment result of step S3. For example, the larger the tilt angle of the pillar, the higher the severity (e.g., a tilt of 5 to 10 degrees is a slight tilt, and a tilt of more than 10 degrees is a severe tilt), and the larger the damaged area of ​​the mesh, the higher the severity (e.g., a damage of 0.5 to 1 square meter is a slight damage, and a damage of more than 1 square meter is a severe damage). Geographic coordinate system calibration: Open professional geographic information system software (such as ArcGIS) and import the associated anomaly information into the software; load the geographic coordinate system layer of the target highway (such as the National Geodetic Coordinate System 2000) in the software, and accurately mark the location of each anomaly point in the layer according to the three-dimensional geographic coordinates of the anomaly points; when marking, different marking symbols should be selected according to the anomaly type, such as using red triangles to mark column tilting anomalies and yellow circles to mark mesh damage anomalies, which is convenient for intuitive differentiation; Calibration result check: Check whether the calibration position of each anomaly point is accurate in the geographic information system software; by overlaying the digital elevation model layer and the fence GIS geographic information layer in step S2, it can be determined whether the anomaly point is located within the actual location range of the fence, ensuring that the calibration result is completely consistent with the real geographic scene and there is no positional offset; Step S4-6: Summarize and generate an isolation fence anomaly report: Anomaly Information Summary: Summarize and organize all anomaly point information calibrated to the geographic coordinate system in steps S4-5; the summary content includes the unique number of each anomaly, anomaly type (pillar tilt or mesh damage), three-dimensional geographic coordinates (longitude, latitude, altitude), anomaly severity (mild, moderate, severe), and associated original image identifier; during the summary process, the anomaly points should be arranged according to their geographical location (e.g., from the starting point to the end point of the highway) to facilitate maintenance personnel to check in order; Report Content Organization: The summarized anomaly information should be organized into a structured fence anomaly report; the report should begin by stating the time of report generation, the target highway section, and the drone data collection parameters (such as flight altitude and camera angle); the main body of the report should present detailed information on each anomaly in tabular form, with columns including anomaly number, anomaly type, longitude, latitude, altitude, severity, and associated image identifiers; the report can end with anomaly distribution statistics charts (such as a percentage chart of anomalies of different severity levels and a schematic diagram of anomaly distribution on the highway section) to provide data support for maintenance decisions; Report Verification and Output: The generated fence anomaly report is fully verified to check for missing information (such as missing coordinates or severity), data errors (such as coordinates not matching the actual location), and formatting issues. After verification, the report is exported to a commonly used format (such as PDF or Excel) to ensure that the report can be easily viewed and used by highway management maintenance personnel, providing clear and accurate guidance for subsequent fence repair work.

[0019] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for identifying damage to highway guardrails using drone aerial photography, characterized in that: Includes the following steps: S1. Data Acquisition Flight Route Planning: Based on the GIS geographic information and digital elevation model of the highway guardrail, a terrain-following flight route adapted to the linear distribution of the guardrail is generated. The planning parameters of the terrain-following flight route include: the horizontal distance between the UAV and the guardrail, the photography angle, the flight altitude and the equidistant shooting points. The flight altitude is dynamically adjusted based on the digital elevation model to maintain a constant altitude relative to the ground. S2. Automated Image Data Acquisition: Control the UAV to fly autonomously along the terrain-following flight path, and acquire images of the fence at preset equidistant shooting points and set shooting angles, obtaining high-resolution sequence images including fence pillars and mesh; at the same time, record the UAV's positioning and attitude determination system data at each shooting point, including latitude and longitude, altitude, pitch angle, roll angle and yaw angle. S3. Intelligent Recognition and Anomaly Detection of Isolation Fence Components: Input the sequence of images into a pre-trained deep learning target recognition model, and simultaneously perform the following recognition and calculations: Target detection is performed on the fence posts in the image to locate their positions. The tilt angle relative to the direction of gravity is calculated based on the image coordinate system regression. If the tilt angle exceeds a preset threshold, it is marked as an abnormal tilt of the post. This process includes the following sub-steps: S3-1. Using a pre-trained deep learning target detection model, target detection of isolation fence pillars is performed on the input sequence images, and the bounding box coordinates of each pillar in the image are output to locate the position of the pillar. S3-2. Based on the bounding box coordinates, crop out the column region sub-image from the original image as the input for subsequent tilt angle calculation; S3-3. Perform edge detection and contour extraction on the sub-image of the column region to identify the main axis of the column; S3-4. In the image coordinate system, calculate the angle between the main axis and the direction of gravity, where the direction of gravity is determined based on the vertical direction of the image, thereby regressing the tilt angle of the column relative to the direction of gravity. S3-5. Compare the calculated tilt angle with the preset threshold. If the tilt angle exceeds the threshold, mark the column as tilting abnormally and record the abnormal information. Target detection is performed on the mesh regions in the image, and pixel-level semantic segmentation is performed within the detected mesh regions to identify mesh damage areas. If the area of ​​the damaged area exceeds a preset threshold, it is marked as mesh damage anomaly. This includes the following sub-steps: S3-6. Using a pre-trained deep learning target detection model, target detection is performed on the input sequence images in the isolation fence region, and the bounding box coordinates of each mesh region in the image are output. S3-7. Based on the bounding box coordinates, extract the mesh region sub-image from the original image as the input for subsequent semantic segmentation; S3-8. Using a pre-trained pixel-level semantic segmentation model, the sub-image of the mesh region is segmented, and the pixels are classified into complete meshes and damaged meshes, thereby generating a binary mask of the damaged mesh region. S3-9. Based on the binary mask, calculate the pixel area of ​​the damaged area of ​​the mesh, and convert the pixel area into the actual physical area by combining the camera parameters and shooting distance. S3-10. Compare the calculated actual physical area with the preset threshold. If the damaged area exceeds the threshold, mark the mesh as abnormal and record the abnormal information. S4. 3D Geographic Calibration of Abnormal Results: The positioning and attitude determination system data recorded in step S2 is integrated with the identification results in step S3. The 3D geographic coordinates of the column tilting abnormality and the mesh damage abnormality are calculated by the spatial forward intersection algorithm. The identified abnormal status is calibrated to the geographic coordinate system to generate an isolation fence abnormality report with geographic coordinates. The report includes the abnormality type, location coordinates and severity. 2.The method for identifying damage of a highway guardrail by a UAV aerial shot according to claim 1, characterized in that: In step S1, based on the GIS geographic information and digital elevation model of the highway guardrail, a terrain-following flight path adapted to the linear distribution of the guardrail is generated, including the following sub-steps: S1-1. Obtain GIS geographic information of highway guardrails and extract linear distribution centerline data of guardrails; S1-2. Based on the linear distribution centerline data, determine the preset lateral horizontal distance between the UAV and the isolation fence, thereby generating a preliminary flight path parallel to the centerline of the isolation fence; S1-3. Based on the preset photography angle and reference flight altitude, and combined with digital elevation model data, perform elevation correction on the initial flight path and dynamically adjust the flight altitude to maintain a constant altitude of the UAV relative to the ground. S1-4. On the corrected flight path, equidistant shooting points are generated based on preset equidistant intervals, thus forming the final terrain-following flight path. 3.The method of claim 1, wherein: Step S2 specifically includes the following sub-steps: S2-1. Load the terrain-following flight path data generated in step S1, control the UAV to fly autonomously along the path, and monitor the UAV's position information in real time to ensure that it follows the predetermined path. S2-2. When the UAV flies to the equidistant shooting point on the terrain-following flight path, the onboard camera equipment is automatically triggered based on the position monitoring results to capture images of the fence at a preset shooting angle, thereby obtaining a series of high-resolution sequential images to ensure that each image covers the fence posts and mesh area. S2-3. At the moment of each image acquisition, synchronously acquire and record the UAV's positioning and attitude determination system data, including latitude and longitude, altitude, pitch angle, roll angle and yaw angle, to form attitude and position information corresponding to each image; S2-4. The acquired high-resolution sequence images are spatiotemporally correlated and stored with the corresponding positioning and attitude determination system data to construct a structured dataset, providing input for subsequent intelligent identification and anomaly detection of the isolation fence components. 4.The method of claim 1, wherein the method further comprises: By fusing the positioning and attitude determination system data recorded in step S2 with the identification results in step S3, the three-dimensional geographic coordinates of the column tilt anomaly and the mesh damage anomaly are calculated using the spatial forward intersection algorithm. The identified anomaly states are then mapped to the geographic coordinate system to generate an isolation fence anomaly report with geographic coordinates. The report includes the anomaly type, location coordinates, and severity, and includes the following sub-steps: S4-1. Extract abnormal information from the identification results of step S3, including the abnormality type and the position coordinates of the abnormal point in the image; wherein, for the column tilting abnormality, the position coordinates of the abnormal point are the coordinates of the center point of the bottom of the column in the image; for the mesh damage abnormality, the position coordinates of the abnormal point are the coordinates of the center point of the damaged area in the image. S4-2. Based on the image identifier of the abnormal point in step S3, obtain the camera position and attitude data of the corresponding image from the positioning and attitude system data recorded in step S2, including latitude and longitude, altitude, pitch angle, roll angle and yaw angle. S4-3. For each anomaly, based on the image sequence and geographic information, find at least two other images containing the anomaly from the sequence images collected in step S2, and obtain the location coordinates of the anomaly in these images and its corresponding camera position and attitude data. S4-4. Using the multiple camera position and attitude data obtained in step S4-3 and the position coordinates of the anomaly points in the corresponding images, calculate the three-dimensional geographic coordinates of the anomaly points using the spatial forward intersection algorithm. S4-5. Associate the three-dimensional geographic coordinates calculated in step S4-4 with the anomaly type and severity in step S3, and label the anomaly status to the geographic coordinate system. S4-6. Summarize all calibrated abnormal states and generate an isolation fence anomaly report. The report includes the anomaly type, three-dimensional geographic coordinates, and severity.

Images