Illegal construction identification method based on aerial photography registration and difference target detection
By combining high-precision image acquisition, accurate registration, and multi-dimensional difference detection with construction site feature recognition, the problem of large errors and high false alarm rates in illegal construction identification using UAV aerial photography has been solved, achieving automated and accurate identification of illegal construction and law enforcement support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGSU XINWANG VIDEO SOFTWARE TECH CO LTD
- Filing Date
- 2026-01-29
- Publication Date
- 2026-04-24
AI Technical Summary
Traditional regulatory methods suffer from poor timeliness, high cost, low accuracy, and slow data updates in identifying illegal construction. Furthermore, drone aerial photography is difficult to accurately register, resulting in large errors and a high false alarm rate. It also lacks the ability to identify characteristic targets on construction sites and cannot automatically link to spatial databases of planning approvals and construction permits.
The method employs high-precision temporal aerial photography acquisition, accurate image registration under POS constraints, multi-dimensional difference extraction and construction patch screening, accurate detection of construction site feature targets and determination of illegal construction. It combines SIFT+RANSAC algorithm and YOLOv8-seg instance segmentation network to achieve image registration and difference target detection, and performs compliance verification through spatial database.
It significantly improves image registration accuracy, reduces false alarm rate, and achieves full automation from image acquisition to alarm push, providing accurate law enforcement basis and applicable to the supervision of illegal construction in urban, rural and remote areas.
Smart Images
Figure CN121921692A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of UAV aerial surveying, image registration, target detection, and spatial database processing, and particularly to a method for identifying illegal constructions based on aerial image registration and differential target detection. Background Technology
[0002] With urbanization and land scarcity, illegal construction (unregistered construction, construction without approval, and construction exceeding permitted area) has gradually become a significant challenge for urban management. Traditional supervision relies on manual inspections, which has several drawbacks. Poor timeliness: Illegal construction is often temporary and concealed, making manual inspections slow; High cost: The monitoring area is large, and manual inspection requires a large amount of manpower; Low accuracy: It is difficult for humans to maintain objectivity, and the workload cannot cover all areas; Slow data updates: Regulatory authorities have difficulty obtaining timely information about new construction projects.
[0003] Drone aerial photography has the advantages of high resolution, high timeliness, low cost and wide coverage, but there are significant differences in heading, lighting and attitude between flights, and direct comparison by pixel will produce a lot of errors.
[0004] Therefore, it is necessary to provide a new method for identifying illegal constructions based on aerial image registration and differential target detection to solve the above-mentioned technical problems. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention provides a method based on image registration and difference detection to automatically identify newly added construction sites and illegal construction activities within a region.
[0006] This invention achieves automated detection and spatial storage of illegal constructions through rigorous registration of flight images, combined with difference detection and construction site target detection.
[0007] This invention aims to solve the following technical problems existing in the current illegal construction identification based on aerial image registration and difference target detection: Aerial images at different times are difficult to align accurately: During the flight of a drone, there are positional shifts, attitude jitters, and lens pitch changes. In addition, the lighting conditions vary significantly in different seasons or times of day, resulting in geometric distortion and radiation distortion between images. Pixel-level alignment cannot be achieved through simple affine transformations, which affects the accuracy of subsequent change detection.
[0008] Large errors in extracting newly added change areas: Traditional image difference methods are sensitive to non-constructive changes such as shadow movement, water surface reflection, and vegetation growth, which easily generate a large number of false change areas; lacking semantic understanding capabilities, it is difficult to distinguish between real construction sites and natural disturbances, resulting in a high false alarm rate.
[0009] The ability to identify the nature of changed areas is insufficient: relying solely on patch change detection can only locate "changes", but cannot determine whether it is a construction site; there is a lack of intelligent identification mechanisms for typical construction site features such as tower cranes, fences, construction vehicles, and exposed soil, making it difficult to provide effective basis for law enforcement judgments.
[0010] The determination of illegal construction lacks data support: the spatial database of planning approvals and construction permits cannot be automatically linked, and the spatial overlay analysis of changing areas and legal project scope cannot be performed, resulting in the inability to distinguish between compliant construction and unauthorized construction, thus restricting the ability of automated regulatory decision-making.
[0011] The illegal construction identification method based on aerial image registration and difference target detection provided by this invention includes the following steps: S1: High-precision time-series aerial photography acquisition; S2: Precise image registration under POS constraints; S3: Multi-dimensional difference extraction and construction drawing patch selection; S4: Precise detection of feature targets at construction sites; S5: Illegal Construction Determination and Compliance Verification.
[0012] Preferably, in step S1, the high-precision time-series aerial image acquisition includes the following steps: S11: Route Planning: Adopt a preset route strategy with a heading overlap of ≥80% and a lateral overlap of ≥60%, using the formula:
[0013] Where L is the effective swath width of the image heading. The flight step length is the heading step length.
[0014] Where W is the effective swath width of the image in the side direction, and Su is the spacing between the side lines; Strictly control the degree of overlap; S12: Data Acquisition: The UAV periodically photographs the target area along a preset flight path to acquire visible light remote sensing images. Simultaneously, it records the POS data (longitude, latitude, heading angle), lens pitch and azimuth angles, and timestamp for each image, forming a time-series aerial image sequence. ; Where n is the number of images; POS data includes longitude, latitude, and heading angle; S13: Preprocessing Optimization: Noise reduction and radiometric correction are performed on the acquired images, using the formula:
[0015] in,( (x,y) represents the original pixel value, D(x,y) represents dark current noise, and R(x,y) represents the relative radiometric correction coefficient to eliminate interference from the shooting environment.
[0016] Preferably, in step S2, the precise image registration under POS constraints includes the following steps: S21: Intelligent Image Pairing: Calculates spatial distance between images based on POS data.
[0017] Where i,j are image indices, used for filtering. Two images of the same spatial location from different periods were used as reference images. With the image to be detected ; S22: Hybrid registration algorithm execution: The registration process is optimized by combining the SIFT+RANSAC algorithm with POS data constraints; S23: Feature point extraction: for and Extracting SIFT feature point sets ; S24: Initial Matching and Purification: Through Filtering matching pairs based on homography matrix (satisfy ), outliers are eliminated iteratively using the RANSAC algorithm; S25: POS Constraint Optimization: Correct the homography matrix H by combining POS data, and finally apply the registration error formula:
[0018] Where (K is the number of interior points, (xk′, yk′) are the registered coordinates) control the error to ensure Pixel.
[0019] Preferably, in step S3, the multi-dimensional difference extraction and construction map patch screening includes the following steps: S31: Construction of Multi-channel Difference Intensity Map: Integrating multi-dimensional difference features to avoid the limitations of single features. S32: ; S33: Gray-scale difference: (Normalized to [0,1]); S34: Texture Feature Difference (T represents the contrast feature of the gray-level co-occurrence matrix); S35: Weighted Fusion Generate a difference intensity map D to enhance the real change characteristics; S36: Precise Extraction of Difference Pixels: Employing Adaptive Threshold Segmentation, Threshold... (mean, σ is the standard deviation), extract the set of differing pixels. ; S37: Differential Patch Generation: (for...) Perform connected component analysis (connectivity condition is pixels) All belong to ), generate a set of connected components Calculate the minimum bounding rectangle of each connected component. As a candidate region.
[0020] Preferably, in step S4, the accurate detection of construction site feature targets includes the following steps: S41: Detection Model Selection: A pre-trained YOLOv8-seg instance segmentation network is adopted. The training samples include positive samples such as construction machinery, foundation pits, material piles, temporary prefabricated houses, and construction workers from the perspective of UAVs, as well as negative samples such as bare land, vegetation, and hardened ground, to improve the model's specificity in recognizing construction site features. S42: Candidate Region Reasoning: The candidate regions generated in step three... Clip and input the model, and set the confidence threshold.
[0021] S43: If there is a construction site feature target with a confidence level ≥ 0.7, then mark the candidate area as a suspected illegal construction area and record the target category c and confidence level conf to effectively distinguish between construction changes and natural changes.
[0022] Preferably, in step S5, the determination of illegal construction and compliance verification includes the following steps: S51: Police Record Generation: Calculate the area of suspected areas. (r is the image ground sampling distance GSD), generate a map containing the latitude and longitude of the patch center. The illegal construction alarm record Reck includes area Ak, target category c, confidence level conf, drone shooting height h, image resolution r, and detection timestamp t. S52: Spatial Database Storage and Compliance Verification: Writes the Reck to a PostGIS extended spatial database, which supports ST_Buffer and ST_Contains spatial operations. S53: Buffer Generation: For the spatial extent P of the registered construction permit, via... {ST_Buffer}(P,5m) Generate permission buffer B; S54: Spatial Containment Determination: via {ST_Contains}(B,({Lon}_c,{Lat}_c)) Perform spatial overlay analysis. If the result is True, mark it as legal construction and cancel the alarm; otherwise, keep the alarm status and push it to the monitoring platform to achieve accurate definition of legal and illegal construction.
[0023] Compared with related technologies, the illegal construction identification method based on aerial image registration and differential target detection provided by this invention has the following beneficial effects: Significantly improved registration accuracy: By combining the SIFT+RANSAC algorithm with POS data constraints, the registration error is controlled within 0.3 pixels, solving the registration problem caused by differences in image pose, position, and illumination, and laying a precise spatial foundation for subsequent processes; Enhanced effectiveness of difference extraction: Multi-dimensional difference fusion and adaptive threshold segmentation effectively suppress noise and lighting interference, accurately capture real changes such as early small-scale construction site traces, and avoid the target being obscured by interference information; The false alarm rate is significantly reduced: the YOLOv8-seg instance segmentation model is used to screen construction feature targets, and spatial compliance verification is performed in combination with the planning permit database. The interference from natural changes and legal construction is filtered in two ways, which significantly reduces the false alarm rate. Strong automation and law enforcement support capabilities: It achieves full-process automation from image acquisition to alarm push, replacing traditional manual patrols and visual interpretation, and improving recognition efficiency; the alarm record contains complete spatial, attribute and image parameter information, providing law enforcement officers with accurate and reliable law enforcement evidence; Wide adaptability: Through designs such as flight path overlap control, dynamic threshold adjustment, and multi-model compatibility, it is adaptable to illegal construction supervision scenarios of different scales and terrains, including cities, towns, and remote areas, and has strong versatility. Attached Figure Description
[0024] Figure 1 The flowchart of the illegal construction identification method based on aerial image registration and differential target detection provided by the present invention is shown. Detailed Implementation
[0025] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0026] In the specific implementation process, such as Figure 1 As shown, an illegal construction identification method based on aerial image registration and difference target detection is proposed. By constructing a time-series aerial image sequence, combining POS-assisted and SIFT-RANSAC algorithms to achieve sub-pixel-level image registration, generating multi-channel difference maps and extracting change patches, then using an instance segmentation model to identify typical construction site features, and finally determining whether it constitutes illegal construction through spatial overlay analysis with the construction permit database, a closed-loop identification mechanism is formed.
[0027] The present invention provides a method for identifying illegal constructions based on aerial image registration and difference target detection, comprising the following steps: Step 1: High-precision time-series aerial photography acquisition Route planning: A preset route strategy with a heading overlap of ≥80% and a lateral overlap of ≥60% is adopted, using the formula:
[0028] L represents the effective swath width of the image heading. The flight step length is the heading step length.
[0029] W represents the effective lateral swath width of the image, and Su represents the lateral flight path spacing. Strictly control the overlap to ensure full coverage of the target area image, laying the foundation for the integrity of difference detection.
[0030] Data Acquisition: The target area is periodically photographed by a drone along a preset flight path to acquire visible light remote sensing images. The POS data (longitude, latitude, and heading angle), lens pitch and azimuth angles, and timestamps of each image are recorded simultaneously to form a time-series aerial image sequence. (n is the number of images), providing multi-dimensional auxiliary data for subsequent registration; Preprocessing optimization: Noise reduction and radiometric correction are performed on the acquired images, using the formula:
[0031] ( (x,y) represents the original pixel value, D(x,y) represents dark current noise, and R(x,y) represents the relative radiometric correction coefficient. This eliminates interference from the shooting environment and improves image quality.
[0032] Step 2: Precise Image Registration under POS Constraints Intelligent image pairing: Calculating spatial distance between images based on POS data.
[0033] (i,j are image indices), filter Two images of the same spatial location from different periods were used as reference images. (Early) and the image to be detected (Recently) Improved matching accuracy; Hybrid registration algorithm execution: The registration process is optimized by combining the SIFT+RANSAC algorithm with POS data constraints. Feature point extraction: and Extracting SIFT feature point set: (For feature descriptors); Initial matching and purification: via:
[0034] Filtering matching pairs based on homography matrix:
[0035] (satisfy ), outliers are eliminated iteratively using the RANSAC algorithm; POS constraint optimization: The homography matrix H is corrected using POS data, and finally, the registration error formula is applied.
[0036] (K is the number of interior points, (xk′, yk′) are the registered coordinates) Control error to ensure Pixels provide a precise spatial reference for subsequent difference extraction.
[0037] Step 3: Multi-dimensional difference extraction and construction drawing feature selection Construction of multi-channel difference intensity maps: Integrating multi-dimensional difference features to avoid the limitations of single features.
[0038] Grayscale difference: (Normalized to [0,1]); Texture feature difference (T represents the contrast feature of the gray-level co-occurrence matrix); Weighted fusion: Generate a difference intensity map D to enhance the real change characteristics; Accurate extraction of difference pixels: Adaptive threshold segmentation is used, and the threshold value is... (mean, σ is the standard deviation), extract the set of differing pixels. Reduce noise interference; Differential patch generation: for Perform connected component analysis (connectivity condition is pixels) All belong to ), generate a set of connected components Calculate the minimum bounding rectangle of each connected component. As a candidate region.
[0039] Step 4: Precise Detection of Feature Targets on the Construction Site Detection model selection: A pre-trained YOLOv8-seg instance segmentation network is adopted. The training samples include positive samples such as construction machinery, foundation pits, material piles, temporary prefabricated houses, and construction workers from the perspective of UAVs, as well as negative samples such as bare land, vegetation, and hardened ground, to improve the model's specificity in recognizing construction site features. Candidate region reasoning: The candidate regions generated in step three... Clip and input the model, and set the confidence threshold.
[0040] If a construction site feature target with a confidence level ≥ 0.7 exists, then the candidate area is marked as a suspected illegal construction area, and the target category is recorded. c With confidence level conf Effectively distinguish between construction changes and natural changes.
[0041] Step 5: Determination of Illegal Construction and Compliance Verification Alarm record generation: Calculate the area of suspected areas:
[0042] (r is the image ground sampling distance GSD), generate a map containing the latitude and longitude of the patch center. ,area Ak Target Category c Confidence level conf Drone shooting altitude h Image resolution r and detection timestamp t Illegal construction alarm records Reck This provides a complete basis for law enforcement evidence collection; Spatial database storage and compliance verification: Reck Write to the PostGIS extended spatial database, which supports ST_Buffer and ST_Contains spatial operations: Buffer zone generation: For the spatial extent P of the registered construction permit, via: {ST_Buffer}(P,5m) Generate permission buffer B; Spatial containment determination: through {ST_Contains}(B,({Lon}_c,{Lat}_c)) Perform spatial overlay analysis. If the result is True, mark it as legal construction and cancel the alarm; otherwise, keep the alarm status and push it to the monitoring platform to achieve accurate definition of legal and illegal construction.
[0043] The illegal construction identification method based on aerial image registration and difference target detection of the present invention: Significantly improved registration accuracy: By combining the SIFT+RANSAC algorithm with POS data constraints, the registration error is controlled within 0.3 pixels, solving the registration problem caused by differences in image pose, position, and illumination, and laying a precise spatial foundation for subsequent processes; Enhanced effectiveness of difference extraction: Multi-dimensional difference fusion and adaptive threshold segmentation effectively suppress noise and lighting interference, accurately capture real changes such as early small-scale construction site traces, and avoid the target being obscured by interference information; The false alarm rate is significantly reduced: the YOLOv8-seg instance segmentation model is used to screen construction feature targets, and spatial compliance verification is performed in combination with the planning permit database. The interference from natural changes and legal construction is filtered in two ways, which significantly reduces the false alarm rate. Strong automation and law enforcement support capabilities: It achieves full-process automation from image acquisition to alarm push, replacing traditional manual patrols and visual interpretation, and improving recognition efficiency; the alarm record contains complete spatial, attribute and image parameter information, providing law enforcement officers with accurate and reliable law enforcement evidence; Wide adaptability: Through designs such as flight path overlap control, dynamic threshold adjustment, and multi-model compatibility, it is adaptable to illegal construction supervision scenarios of different scales and terrains, including cities, towns, and remote areas, and has strong versatility.
[0044] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for identifying illegal constructions based on aerial image registration and difference target detection, characterized in that, Includes the following steps: S1: High-precision time-series aerial photography acquisition; S2: Precise image registration under POS constraints; S3: Multi-dimensional difference extraction and construction drawing patch selection; S4: Precise detection of characteristic targets at construction sites; S5: Illegal Construction Determination and Compliance Verification.
2. The illegal construction identification method based on aerial image registration and difference target detection according to claim 1, characterized in that: In step S1, the high-precision time-series aerial image acquisition includes the following steps: S11: Route Planning: Adopt a preset route strategy with a heading overlap of ≥80% and a lateral overlap of ≥60%, using the formula: Where L is the effective swath width of the image heading. The flight step length is the heading step length. Where W is the effective swath width of the image in the side direction, and Su is the spacing between the side lines; Strictly control the degree of overlap; S12: Data Acquisition: The UAV periodically photographs the target area along a preset flight path to acquire visible light remote sensing images. Simultaneously, it records the POS data (longitude, latitude, heading angle), lens pitch and azimuth angles, and timestamp for each image, forming a time-series aerial image sequence. ; Where n is the number of images; POS data includes longitude, latitude, and heading angle; S13: Preprocessing Optimization: Noise reduction and radiometric correction are performed on the acquired images, using the formula: in,( (x,y) represents the original pixel value, D(x,y) represents dark current noise, and R(x,y) represents the relative radiometric correction coefficient to eliminate interference from the shooting environment.
3. The illegal construction identification method based on aerial image registration and difference target detection according to claim 1, characterized in that: In step S2, accurate image registration under POS constraints includes the following steps: S21: Intelligent Image Pairing: Calculates spatial distance between images based on POS data. Where i,j are image indices, used for filtering. Two images of the same spatial location from different periods were used as reference images. With the image to be detected ; S22: Hybrid registration algorithm execution: The registration process is optimized by combining the SIFT+RANSAC algorithm with POS data constraints; S23: Feature point extraction: for and Extracting SIFT feature point sets ; S24: Initial Matching and Purification: Through Filtering matching pairs based on homography matrix (satisfy ), outliers are eliminated iteratively using the RANSAC algorithm; S25: POS Constraint Optimization: Correct the homography matrix H by combining POS data, and finally apply the registration error formula: Where (K is the number of interior points, (xk′, yk′) are the registered coordinates) control the error to ensure Pixel.
4. The illegal construction identification method based on aerial image registration and difference target detection according to claim 1, characterized in that: In step S3, the multi-dimensional difference extraction and construction map patch screening include the following steps: S31: Construction of Multi-channel Difference Intensity Map: Integrating multi-dimensional difference features to avoid the limitations of single features. S32： ； S33: Gray-scale difference: (Normalized to [0,1]); S34: Texture Feature Difference (T represents the contrast feature of the gray-level co-occurrence matrix); S35: Weighted Fusion Generate a difference intensity map D to enhance the real change characteristics; S36: Precise Extraction of Difference Pixels: Employing Adaptive Threshold Segmentation, Threshold... (mean, σ is the standard deviation), extract the set of differing pixels. ; S37: Differential Patch Generation: For Perform connected component analysis (connectivity condition is pixels) All belong to ), generate a set of connected components Calculate the minimum bounding rectangle of each connected component. As a candidate region.
5. The method for identifying illegal constructions based on aerial image registration and difference target detection according to claim 1, characterized in that: In step S4, the accurate detection of construction site feature targets includes the following steps: S41: Detection Model Selection: A pre-trained YOLOv8-seg instance segmentation network is adopted. The training samples include positive samples such as construction machinery, foundation pits, material piles, temporary prefabricated houses, and construction workers from the perspective of UAVs, as well as negative samples such as bare land, vegetation, and hardened ground, to improve the model's specificity in recognizing construction site features. S42: Candidate Region Reasoning: The candidate regions generated in step three... Clip and input the model, and set the confidence threshold. S43: If there is a construction site feature target with a confidence level ≥ 0.7, then mark the candidate area as a suspected illegal construction area and record the target category c and confidence level conf to effectively distinguish between construction changes and natural changes.
6. The method for identifying illegal constructions based on aerial image registration and difference target detection according to claim 1, characterized in that: In step S5, the determination of illegal construction and compliance verification includes the following steps: S51: Police Record Generation: Calculate the area of suspected areas. (r is the image ground sampling distance GSD), generate a map containing the latitude and longitude of the patch center. The illegal construction alarm record Reck includes area Ak, target category c, confidence level conf, drone shooting height h, image resolution r, and detection timestamp t. S52: Spatial Database Storage and Compliance Verification: Writes the Reck to a PostGIS extended spatial database, which supports ST_Buffer and ST_Contains spatial operations. S53: Buffer Generation: For the spatial extent P of the registered construction permit, via... {ST_Buffer}(P,5m) Generate permission buffer B; S54: Spatial Containment Determination: via {ST_Contains}(B,({Lon}_c,{Lat}_c)) Perform spatial overlay analysis. If the result is True, mark it as legal construction and cancel the alarm; otherwise, keep the alarm status and push it to the monitoring platform to achieve accurate definition of legal and illegal construction.