Search

CN-122024097-A - Inspection defect duplication elimination method based on target detection and space coordinate conversion

CN122024097ACN 122024097 ACN122024097 ACN 122024097ACN-122024097-A

Abstract

The invention discloses a patrol defect deduplication method based on target detection and space coordinate conversion, and relates to the field of unmanned aerial vehicle patrol and artificial intelligence. The method comprises the steps of collecting images of a measured object, synchronizing POS data and lens attitude data through an unmanned aerial vehicle, outputting defect information through a pre-trained target detection model, mapping coordinates of defective pixels into WGS-84 longitude and latitude as unique space identification by combining camera parameters, constructing a circular buffer index of a dynamic space threshold R, realizing duplication elimination through space query, category judgment and confidence comparison, and repeatedly recording and writing into a redundant log table for model iterative optimization. The invention solves the problems of repeated alarm, insufficient positioning precision and single deduplication logic in the prior art, has the advantages of accurate deduplication, strong adaptability and high efficiency, realizes closed-loop optimization, and is suitable for unmanned aerial vehicle inspection of large-scale infrastructures such as power lines, bridges and the like.

Inventors

  • LI YONGJUN
  • Chen Shutu
  • XU MIAO
  • ZHANG DAN
  • WANG LIN

Assignees

  • 江苏欣网视讯软件技术有限公司
  • 南京欣网飞联无人机科技有限公司

Dates

Publication Date
20260512
Application Date
20260127

Claims (7)

  1. 1. A patrol defect deduplication method based on target detection and space coordinate conversion comprises the following steps: (a) The method comprises the steps that a visible light image of a measured object is obtained through an image acquisition unit carried on an unmanned aerial vehicle, POS (Position and Orientation System) data at the image acquisition time and pitching and direction angle data of a lens relative to an unmanned aerial vehicle body are synchronously recorded, and POS data at least comprise longitude, latitude and course angle of a platform of the unmanned aerial vehicle; (b) Reasoning the visible light image by adopting a pre-trained target detection model to output a pixel-level bounding box and a category confidence of a defect target; (c) Based on the POS data, the lens angle, the internal parameters of the image acquisition unit and the distortion parameters, establishing a mapping relation between an image pixel coordinate system and a geographic coordinate system, and converting a pixel-level bounding box center coordinate of the defect target into WGS-84 longitude and latitude to obtain a unique spatial position identifier of the defect; (d) Taking the category, the confidence coefficient and the WGS-84 longitude and latitude (Lat, lon) of the defect as an alarm record to be put in storage, and writing the alarm record into a temporary cache queue; (e) Constructing a circular buffer area index taking longitude and latitude as a main key and taking a preset space threshold value R as a radius in a geographic database; (f) And executing space inquiry on each alarm record to be put in storage in the temporary buffer queue by taking (Lat, lon) as a circle center and taking R as a radius, if the inquiry result is empty, judging the defect as a newly found defect, writing the defect into the geographic database and forming a formal alarm record, if the inquiry result is not empty, further comparing the consistency of the alarm record to be put in storage with the category of the existing alarm record, if the categories are the same, judging that the alarm is repeated and discarding the alarm record, and if the categories are different, updating the category and the confidence coefficient of the alarm record corresponding to the space position and refreshing the timestamp of the alarm record.
  2. 2. The method according to claim 1, wherein the spatial threshold R is dynamically adjusted according to the physical size of the measured object and the inspection accuracy requirement, and the range of the spatial threshold R is 0.5 m-25 m.
  3. 3. The method according to claim 1 or 2, wherein the object detection model is any one of YOLOv, fast R-CNN or SSD, and the training set contains defect samples in the same scene as the object under test.
  4. 4. A method according to any of claims 1-3, wherein the image acquisition unit is at least one of a visible light camera, an infrared camera or a multispectral camera, and its internal parameters and distortion parameters are obtained by field calibration.
  5. 5. The method of any of claims 1-4, wherein the geographic database is stored in PostGIS extended relational database or MongoDB GeoJSON format to support efficient spatial indexing and querying.
  6. 6. The method of any of claims 1-5, wherein the drone comprises a plurality of configurations, which may be any of a multi-rotor drone, a fixed-wing drone, a mixed-wing drone.
  7. 7. The method of any one of claims 1-6, wherein in step (f), if it is determined that the alarm is repeated and discarded, the discarded record is further written into a redundant log table for subsequent model iterative optimization and false positive analysis.

Description

Inspection defect duplication elimination method based on target detection and space coordinate conversion Technical Field The invention belongs to the technical field of unmanned aerial vehicle inspection, computer vision and Geographic Information System (GIS), and is mainly applied to intelligent defect identification and repeated alarm filtration in the fields of electric power inspection, pipeline inspection, bridge inspection and the like. Background Unmanned aerial vehicle inspection has become the high frequency application scene in industries such as electric power, traffic, energy. The unmanned aerial vehicle collects inspection images through the visible light camera and runs a target detection algorithm, so that defects such as high temperature points, insulator damage and pole tower corrosion can be automatically identified. However, the prior art has the following core problems: 1. repeated alarm is serious  The same defect is repeatedly shot in different angles, different heights or multiple flight tasks, and the target detection model can repeatedly alarm.  Lack a reliable spatial alignment mechanism to determine if two alarms come from the same physical location. 2. Lack of accurate conversion of pixel coordinates to geographic coordinates In the prior art, the longitude and latitude of the unmanned aerial vehicle at the center point of the image are used as defect points, and the error can reach several meters to tens of meters. 3. Coarse spatial index, not supporting dynamic decisions  Defects cause spatial deduplication failure due to inaccurate longitude and latitude falling points. Most inspection platforms adopt a character string matching and file name comparison peer-to-peer pseudo-deduplication mode, and cannot adapt to a large-scale inspection scene. Therefore, a high-precision defect deduplication method that fuses visual inspection, POS data, camera models, and spatial databases is needed. The following key technical problems still exist in the current inspection defect deduplication of target detection and space coordinate conversion: 1. the accurate geographic mapping has the challenges that the conversion process from the image pixel coordinates to the WGS-84 geographic coordinates is complex, the accumulation of multisource errors such as camera distortion, unmanned plane positioning and attitude errors, coordinate model errors and the like is involved, and under the condition of complex terrain and dynamic flight, the traditional simple projection model can cause positioning errors above the meter level, and the spatial reference accuracy of de-duplication is seriously affected. 2. The bottleneck of space unique identification construction is that space-time references of multi-sensor (such as visible light and infrared) data are not uniform, and target detection confidence is subject to fluctuation caused by environmental interference, so that the stable binding of defect types, confidence and geographic positions to form a space unique key is challenging, and the identification consistency of the same physical defect under different conditions is difficult to ensure. 3. The performance bottleneck of massive space inquiry is that in the face of millions of defect records accumulated in history, the efficiency of traditional space indexes is reduced when data is unevenly distributed (such as linear inspection paths), and the calculation cost of buffer area inquiry with fixed threshold is high in a high concurrency real-time processing scene, so that the high-efficiency deduplication requirement of an inspection system on minute-level response is difficult to meet. The complexity of repeated alarm intelligent decision-making is that the existing simple strategy of 'similar discarding and heterogeneous updating' cannot process semantic association and operation and maintenance priority among defects, and a fixed space deduplication threshold cannot adapt to positioning errors of a multi-scale inspection scene and dynamic changes, so that a decision mechanism is stiff, and intelligent consideration of confidence level attenuation and risk level is lacked. Disclosure of Invention In order to solve the technical problems in the background art, the invention mainly provides a patrol defect deduplication method based on target detection and space coordinate conversion, which is used for precisely mapping a center point of a defect bounding box in an image to a WGS-84 geographic coordinate system by fusing unmanned aerial vehicle POS data, lens gestures and camera parameters, establishing a circular buffer zone index by taking longitude and latitude as a space main key, realizing automatic deduplication of cross-view angles and cross-voyage times in combination with category consistency judgment, obviously reducing false alarm rate and improving defect management quality. 1. The inspection defect duplication elimination method based on target detection and space coordinate conve