Tour inspection defect duplicate removal method based on target detection and space coordinate conversion
By integrating POS data and lens pose into drone inspections, precise mapping and efficient deduplication of image defects to geographic coordinates are achieved, solving the problems of duplicate alarms and inaccurate positioning in drone inspections, and improving the accuracy and efficiency of defect management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-27
- Publication Date
- 2026-05-12
AI Technical Summary
The inspection of drones suffers from serious problems such as repeated alarms, lack of accurate pixel coordinate to geographic coordinate conversion, coarse spatial indexing and low deduplication efficiency. Especially in complex terrain and dynamic flight conditions, this leads to inaccurate defect location and ineffective removal of duplicate records.
By integrating UAV POS data, lens attitude, and camera parameters, image defects are accurately mapped to the WGS-84 geographic coordinate system. A circular buffer index with latitude and longitude as the primary key is used, combined with category consistency judgment to achieve automatic deduplication across viewpoints and flights. Efficient spatial indexing and query methods are adopted to support multiple sensors and multiple AI models.
It significantly reduces the rate of repeated alarms, improves the accuracy and efficiency of defect management, supports generalization and adaptability to various sensors and drones, and optimizes model performance through redundant recording.
Smart Images
Figure CN122024097A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the fields of drone inspection, computer vision and geographic information system (GIS) technology, and is mainly applied to intelligent defect identification and repeated alarm filtering in fields such as power inspection, pipeline inspection and bridge inspection. Background Technology
[0002] Drone inspections have become a frequent application in industries such as power, transportation, and energy. Drones use visible light cameras to capture inspection images and run target detection algorithms to automatically identify defects such as high-temperature points, insulator damage, and tower corrosion.
[0003] However, existing technologies have the following core problems: 1. Repeated alarms are serious If the same defect is repeatedly photographed from different angles, at different altitudes, or during multiple flight missions, the target detection model will repeatedly issue an alarm.
[0004] There is a lack of reliable spatial alignment mechanisms to determine whether two alarms originate from the same physical location.
[0005] 2. Lack of accurate pixel coordinate to geographic coordinate conversion Current technologies mostly rely on the latitude and longitude of the drone at the center of the image as the defect point, which can result in an error of several meters to tens of meters.
[0006] 3. Spatial indexing is coarse and does not support dynamic decision-making. The inaccurate latitude and longitude coordinates caused by the defect led to the failure of spatial deduplication.
[0007] Most inspection platforms use "pseudo-deduplication methods" such as string matching and filename comparison, which cannot adapt to large-scale inspection scenarios.
[0008] Therefore, a high-precision defect deduplication method that integrates visual inspection, POS data, camera models, and spatial databases is needed.
[0009] The current methods for deduplication of inspection defects using target detection and spatial coordinate transformation still face the following key technical challenges: 1. Challenges of accurate geographic mapping: The conversion process from image pixel coordinates to WGS-84 geographic coordinates is complex, involving the accumulation of multiple sources of errors such as camera distortion, UAV positioning and attitude errors, and coordinate model deviations. Furthermore, under complex terrain and dynamic flight conditions, traditional simple projection models can lead to positioning errors of more than meters, which seriously affects the accuracy of deduplication spatial reference.
[0010] 2. Bottlenecks in constructing spatially unique identifiers: The lack of uniformity in the spatiotemporal reference of multi-sensor (such as visible light and infrared) data, as well as the fluctuations in target detection confidence caused by environmental interference, make it challenging to stably bind defect categories, confidence levels, and geographical locations to form a "spatially unique key," making it difficult to guarantee the consistency of the identification of the same physical defect under different conditions.
[0011] 3. Performance bottleneck of massive spatial queries: When faced with millions of defect records accumulated over time, the efficiency of traditional spatial indexes decreases when the data is unevenly distributed (such as linear inspection paths), and the fixed threshold buffer query has a large computational overhead in high-concurrency real-time processing scenarios, making it difficult to meet the inspection system's requirement for efficient deduplication with minute-level response.
[0012] Complexity of intelligent decision-making for repeated alarms: The existing simple strategy of "discarding similar items and updating dissimilar items" cannot handle the semantic association between defects and the operation and maintenance priority. Furthermore, the fixed spatial deduplication threshold cannot adapt to multi-scale inspection scenarios and dynamically changing positioning errors, resulting in a rigid decision-making mechanism that lacks intelligent consideration of confidence decay and risk level. Summary of the Invention
[0013] To address the technical problems described in the background section, this invention primarily provides a method for deduplicating inspection defects based on target detection and spatial coordinate transformation. By fusing UAV POS data, lens attitude, and camera parameters, the center point of the defect bounding box in the image is accurately mapped to the WGS-84 geographic coordinate system. A circular buffer index is established using latitude and longitude as the spatial primary key. Combined with category consistency judgment, automatic deduplication across perspectives and flights is achieved, significantly reducing the false alarm rate and improving the quality of defect management.
[0014] 1. The specific technical solution is: a method for deduplicating inspection defects based on target detection and spatial coordinate transformation. (a) Multi-source data acquisition The visible light image of the object under test is acquired by the image acquisition unit mounted on the drone, and the POS (Position and Orientation System) data and the pitch and orientation angle data of the lens relative to the drone body are recorded at the same time as the image acquisition. The POS data includes at least the longitude, latitude and heading angle of the drone platform, and the lens angle data is used to characterize the shooting posture of the camera, providing a basis for subsequent coordinate transformation.
[0015] The UAV can be any one of a multi-rotor UAV, a fixed-wing UAV, or a hybrid-wing UAV, and can be flexibly selected according to the actual scene terrain, the object being measured, and the inspection range; the image acquisition unit can be at least one of a visible light camera, an infrared camera, or a multispectral camera, and obtain the camera intrinsic parameters (such as focal length, principal point coordinates) and distortion parameters (such as radial distortion coefficient, tangential distortion coefficient) through on-site calibration to ensure the imaging accuracy of the image data.
[0016] (b) Defect target detection A pre-trained target detection model is used to infer the visible light image to output the pixel-level bounding box and category confidence of the defective target. The target detection model is any one of YOLOv8, Faster R-CNN or SSD, and its training set contains defect samples in the same scene as the object being tested to ensure the model's ability to identify defects in the target scene.
[0017] (c) Spatial coordinate transformation Based on the POS data, lens angle, image acquisition unit intrinsic parameters and distortion parameters obtained in step (a), a mapping relationship between the image pixel coordinate system and the geographic coordinate system is established, and the pixel-level bounding box center coordinates of the defect target are converted into WGS-84 latitude and longitude to obtain the unique spatial location identifier of the defect. (d) Cache of records to be added to the database The defect category, confidence level, and WGS-84 latitude and longitude (Lat, Lon) are integrated into a single alarm record to be added to the warehouse and written to a temporary cache queue. (e) Spatial Index Construction A spatial index is constructed in the geographic database, with the latitude and longitude of the index defect as the primary key and a circular buffer index structure with a preset spatial threshold R as the radius. The geographic database adopts a storage scheme that supports efficient spatial indexing and querying, such as a relational database extended by PostGIS or MongoDB GeoJSON format. The spatial threshold R is dynamically adjusted according to the physical size of the object being measured and the inspection accuracy requirements, with a value range of 0.5 m to 25 m, to ensure that the spatial range of the same defect can be accurately covered, while avoiding the overlap of the spatial ranges of different defects.
[0018] (f) Multi-dimensional deduplication judgment For each alarm record to be added to the temporary cache queue, the following deduplication logic is executed: Perform a spatial query with the WGS-84 latitude and longitude (Lat, Lon) of the record as the center and R as the radius to determine whether there is an alarm record within the spatial range; If the query result is empty, the defect is determined to be a newly discovered defect, and it is written into the geographic database and a formal alarm record is formed. If the query result is not empty, then further compare the consistency of the categories of the alarm records to be added to the database with the existing alarm records: If the categories are the same, it is determined to be a duplicate alarm and discarded. At the same time, the discarded record is written to the redundant log table for subsequent model iteration optimization and false detection analysis. If the categories are different, update the category and confidence level of the alarm record corresponding to that spatial location, and refresh its timestamp.
[0019] Beneficial effects of this invention: 1. Visual inspection results are accurately projected onto geospatial data, significantly reducing misjudgments and duplicate recordings.
[0020] 2. Achieve efficient and large-scale deduplication of inspection data through spatial indexing and threshold filtering.
[0021] 3. Supports category conflict updates, improving the reliability and timeliness of defect records.
[0022] 4. It is compatible with multiple types of sensors, various drones, and multiple AI models, and has strong generalization capabilities.
[0023] 5. Redundant records are traceable, supporting model training iterations and continuous system optimization. Attached Figure Description
[0024] To further illustrate the technical solution of the present invention, the following is a brief description of the accompanying drawings: Figure 1 : Overall flowchart of the method of the present invention. Detailed Implementation
[0025] The present invention will be further described in detail below with reference to the embodiments.
[0026] Example 1: 1. Equipment selection and parameter settings Drone: Multi-rotor drone (DJI M300 RTK), equipped with an RTK positioning module, with a positioning accuracy of ±1cm +1ppm; Image acquisition unit: Visible light camera (Zenmuse P1), focal length 24mm, calibrated on-site using Zhang's calibration method, intrinsic parameter matrix K=[1000,0,500;0,1000,300;0,0,1], distortion parameter D=[-0.01,0.005,0,0]; The target detection model is YOLOv8. The training set contains 5,000 samples of common defects in power lines (such as broken insulators, broken conductor strands, bird nests, etc.). After data augmentation, the mAP (mean accuracy) reached 0.88. Spatial threshold R: The conductor diameter of the 110kV transmission line is 21.6mm, the insulator string length is 1.2m, the inspection accuracy requirement is ±5cm, and considering the drone's flight altitude of 100m and the camera focal length of 24mm, R is calculated to be 2.0m using the formula R=k×S×H / (P×f) (k=0.2, S=1.2m, H=100m, P=0.05m, f=24mm). Therefore, R is finally set to 2.0m.
[0027] The parameters in the formula R=k×S×H / (P×f) are expressed as follows: k: Proportional coefficient correction coefficient, used to balance the influence of various parameters on the spatial threshold, adapt to different inspection scenarios, set according to engineering practice experience, with a value range of 0.1 to 0.5. Larger values are used for complex scenarios (such as those with many obstructions), and smaller values are used for simple scenarios. S: The key dimensions of the measured object (such as component length and diameter) that are related to the physical dimensions of the measured object and the defects, reflecting the scale of the area where the defects are located. The core component dimensions where defects are likely to occur are selected, such as the length of the insulator string for power lines and the width of the box girder for bridges, to ensure that the threshold is adapted to the distribution range of defects. H: The vertical flight altitude of the UAV relative to the surface of the object being measured at the moment of image acquisition. It is determined by the flight path planning and needs to be combined with the inspection range and camera focal length to avoid the defect pixel ratio being too small due to excessive altitude. P: The inspection accuracy requirement is the preset defect location accuracy threshold, which is the upper limit of the allowable defect coordinate transformation error. It is set according to industry inspection standards, such as power line inspection accuracy ≥ ±5cm and bridge crack inspection accuracy ≥ ±1cm. f: The actual focal length used by the lens focal length image acquisition unit is the effective focal length determined after camera calibration. It is determined by the camera hardware parameters and inspection requirements. Short focal length is used for close-range inspections, and long focal length is used for large-area inspections.
[0028] 2. Implementation Steps (a) Data acquisition: The UAV flies along the preset route of the power line at an altitude of 100m and a speed of 8m / s. The visible light camera captures 10 frames per second and records the POS data (longitude Lon_UAV, latitude Lat_UAV, heading angle ψ), lens pitch angle α (-30°~30°) and heading angle β (0°~360°) of each frame simultaneously. (b) Defect detection: The acquired images are transmitted to the edge computing node, and the YOLOv8 model is used for inference to output the bounding box (x1,y1,x2,y2), category (e.g., "insulator damage") and confidence level (e.g., 0.92). (c) Coordinate transformation: The center coordinates (x0, y0) of the defect bounding box are distorted and transformed to the camera coordinate system through the intrinsic parameter matrix. Combined with POS data and lens angle, the WGS-84 latitude and longitude (Lat_defect, Lon_defect) are obtained through UTM projection transformation with a transformation error of ±3cm. (d) Cache storage: The defect category, confidence level, latitude and longitude, and collection timestamp are combined to form a record to be entered into the database and written into a Redis temporary cache queue; (e) Index construction: Using a PostgreSQL database extended with PostGIS, an R-tree spatial index was constructed with latitude and longitude as the primary key and a circular buffer radius of R=1m; (f) Deduplication determination: Perform a spatial query on the records in the cache queue. If there is already an alarm record of the "insulator damage" category within the latitude and longitude range of a certain record, it is determined to be a duplicate alarm. The record is discarded and written to the redundant log table. If there are records of different categories (such as "bird's nest"), the alarm record of that spatial location is updated to the category with higher confidence. (g) Model optimization: Monthly statistics of redundant log table data, extract 1000 duplicate detection samples and 500 false detection samples, manually label them and add them to the training set, and update the YOLOv8 model through data augmentation methods such as random flipping and brightness adjustment.
[0029] Example 2: Bridge Inspection Scenario 1. Equipment selection and parameter settings Unmanned Aerial Vehicle (UAV): Fixed-wing UAV (Wing Loong-2H), equipped with a PPP-RTK positioning module, with a positioning accuracy of ±2cm +1ppm; Image acquisition unit: multispectral camera (RedEdge-MX), focal length 16mm, after on-site calibration, intrinsic parameter matrix K=[800,0,400;0,800,300;0,0,1], distortion parameter D=[-0.008,0.003,0,0]; The target detection model is Faster R-CNN, with a training set containing 8,000 samples of bridge defects (such as cracks, concrete spalling, exposed rebar, etc.), achieving an mAP of 0.86. Spatial threshold R: The minimum detection accuracy for bridge crack width is 1mm. The width of the object being measured (bridge main beam) is 5m, so R is set to 3m.
[0030] 2. Implementation Results Bridge inspection using the method of this invention collected 100,000 frames of images and detected 320 defects. After deduplication, 298 valid defects were obtained, reducing the duplicate alarm rate from 35% in the prior art to 2.5%. After optimizing the redundant log table data model, the false detection rate decreased from 12% to 4%, the inspection data processing efficiency was improved by 60%, and the defect review time for maintenance personnel was shortened by 50%.
[0031] Additional note: The POS data described in this invention can be obtained by fusing the GPS / BeiDou positioning module and the IMU inertial measurement unit carried by the UAV, ensuring the synchronization and accuracy of position and attitude data; The dynamic adjustment of the spatial threshold R can be automatically achieved through the background algorithm, and is updated in real time according to the preset parameters of the object under test (such as physical size and inspection accuracy) and real-time flight data (such as altitude and distance). Spatial indexes in geodatabases can be dynamically optimized based on the data volume. When the data volume exceeds 100,000 records, a partitioned indexing strategy is adopted to improve query efficiency. Model iteration and optimization can be achieved using an automated training platform, which can trigger training tasks regularly (e.g., weekly or monthly) without human intervention, thus enabling continuous model upgrades.
[0032] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.
Claims
1. A method for deduplicating inspection defects based on target detection and spatial coordinate transformation: (a) Acquire visible light images of the object under test through the image acquisition unit mounted on the UAV, and simultaneously record the POS (Position and Orientation System) data and the pitch and orientation angle data of the lens relative to the UAV body at the moment of image acquisition. The POS data includes at least the longitude, latitude and heading angle of the UAV platform. (b) The visible light image is inferred using a pre-trained target detection model to output the pixel-level bounding box and category confidence of the defective target; (c) Based on the POS data, lens angle, image acquisition unit intrinsic parameters and distortion parameters, establish the mapping relationship between the image pixel coordinate system and the geographic coordinate system, and convert the pixel-level bounding box center coordinates of the defect target into WGS-84 latitude and longitude to obtain the unique spatial location identifier of the defect. (d) The category, confidence level, and WGS-84 latitude and longitude (Lat, Lon) of the defect are written as an alarm record to be put into storage and written into a temporary cache queue; (e) Construct a circular buffer index in the geographic database with latitude and longitude as the primary key and a preset spatial threshold R as the radius; (f) For each alarm record to be added to the database in the temporary cache queue, perform a spatial query with (Lat, Lon) as the center and R as the radius. If the query result is empty, the defect is determined to be a newly discovered defect, and it is written into the geographic database to form a formal alarm record. If the query result is not empty, the category consistency between the alarm record to be added to the database and the existing alarm record is further compared. If the categories are the same, it is determined to be a duplicate alarm and discarded. If the categories are different, the category and confidence of the alarm record corresponding to the spatial location are updated, and its timestamp is refreshed.
2. The method according to claim 1, characterized in that, The spatial threshold R is dynamically adjusted according to the physical size of the object being measured and the inspection accuracy requirements, and its value ranges from 0.5 m to 25 m.
3. The method according to claim 1 or 2, characterized in that, The target detection model is any one of YOLOv8, Faster R-CNN or SSD, and its training set contains defect samples in the same scene as the object being tested.
4. The method according to any one of claims 1-3, characterized in that, The image acquisition unit is at least one of a visible light camera, an infrared camera, or a multispectral camera, and its intrinsic parameters and distortion parameters are obtained through on-site calibration.
5. The method according to any one of claims 1-4, characterized in that, The geographic database is stored in a PostGIS extended relational database or MongoDB GeoJSON format to support efficient spatial indexing and querying.
6. The method according to any one of claims 1-5, characterized in that, The drone includes various configurations and can be any one of multi-rotor drones, fixed-wing drones, or hybrid-wing drones.
7. The method according to any one of claims 1-6, characterized in that, In step (f), if a duplicate alarm is determined and discarded, the discarded record is further written into a redundant log table for subsequent model iteration optimization and false detection analysis.