An unmanned aerial vehicle city illegal building patrol evidence collection method based on multi-modal fusion

Abstract

This invention discloses a method for urban illegal construction inspection and evidence collection using unmanned aerial vehicles (UAVs) based on multimodal fusion, comprising: Step 1: collecting historical compliant spatial data and performing 3D reconstruction to obtain a historical legal 3D reference volume; Step 2: collecting current inspection images and pose data and performing spatial registration and differencing to obtain a 3D residual volume; Step 3: performing multi-view consistency screening on the 3D residual volume to obtain a stable residual voxel set; Step 4: extracting component structure sequences and inputting them into an improved OneFlow framework to obtain structure generation probability and structure offset probability; Step 5: calculating anomaly scores and constructing a thermal activity time fingerprint; Step 6: generating a regulatory conflict body and calculating the confidence level for illegal construction evidence collection; Step 7: sorting and filtering abnormal areas and generating an evidence collection package. This invention achieves high-precision identification and evidence collection of illegal construction.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and machine learning technology, and in particular to a method for using drones to patrol and collect evidence of illegal construction in cities based on multimodal fusion. Background Technology

[0002] With the acceleration of urbanization and the widespread application of drone technology in urban management, using drones to inspect and collect evidence of illegal constructions in cities has become an important technical means. Existing technologies usually use visible light images or simple multi-source data to identify and compare target areas, and determine whether there are abnormalities in building structures through target detection or change detection methods. However, there are still many shortcomings in practical applications.

[0003] Traditional drone-based inspection methods often rely on single images or simple overlays of multimodal data for analysis, lacking a unified three-dimensional spatial benchmark modeling mechanism. This makes it difficult to establish precise alignment relationships between data acquired at different times, thus affecting the accurate determination of changes in illegal structures. Existing change detection methods are mostly based on two-dimensional pixel differences or local feature matching, which are insufficient to effectively reflect the true structural changes of buildings in three-dimensional space and have inadequate ability to identify complex forms of illegal structures such as roof additions and three-dimensional extensions. In terms of anomaly detection, existing methods are usually based on single features or simple thresholds, lacking the ability to model the laws of structural generation and the process of structural evolution. This results in a weak ability to distinguish complex structural changes, making it susceptible to interference from occlusion, changes in lighting, and temporary storage, leading to misjudgments. Furthermore, most methods fail to correlate thermal infrared time-series information with spatial structural changes, making it difficult to identify continuously used illegal structures and reducing the reliability of evidence collection results. In terms of regulatory constraint analysis, existing technologies mostly rely on simple spatial superposition judgments, lacking a comprehensive computational mechanism for multiple types of spatial constraints. Simultaneously, existing evidence collection methods typically lack a unified evaluation system based on probability distributions, making it difficult to quantify, fuse, and rank different abnormal factors, affecting the stability and consistency of illegal structure identification results.

[0004] Therefore, how to provide a method for inspecting and collecting evidence of illegal construction in cities using drones based on multimodal fusion is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a method for urban illegal construction inspection and evidence collection using drones based on multimodal fusion. This invention constructs a three-dimensional spatial representation system that integrates multimodal data, and achieves accurate extraction of structural changes through three-dimensional residual volume and multi-view consistency screening. It introduces an improved OneFlow framework to perform dual-path modeling of structural generation and offset, thereby improving the ability to identify complex illegal constructions. By combining thermal infrared time fingerprinting and regulatory conflict analysis, it achieves unified quantitative evaluation and ranking screening of multiple factors, and finally generates highly reliable evidence collection results for illegal constructions.

[0006] A method for detecting and collecting evidence of illegal construction in cities using unmanned aerial vehicles (UAVs) based on multimodal fusion, according to an embodiment of the present invention, includes the following steps: Step 1: Collect historical compliant spatial data of the target area and perform 3D reconstruction to obtain the historical legal 3D reference volume of the building; Step 2: Collect the visible light image, oblique image, thermal infrared image, RTK positioning data and IMU attitude data of the current inspection, perform spatial registration on the visible light image and oblique image, and perform difference with the historical legal three-dimensional reference volume to obtain the three-dimensional residual volume; Step 3: Perform multi-view consistency screening on the three-dimensional residual volume to obtain a stable set of residual voxels; Step 4: Project the stable residual voxel set onto the roof plane and perform component segmentation, extract the component structure sequence, input the component structure sequence into the improved OneFlow framework, perform generation path mapping and offset path mapping on the component structure sequence, and obtain the component structure generation probability and structure offset probability of the roof. Step 5: Calculate the roof component anomaly score based on the structure generation probability and the structure offset probability, and normalize the multi-period thermal infrared images to extract the temperature change sequence of the region corresponding to the stable residual voxel set, thereby obtaining the thermal activity time fingerprint; Step 6: Map the stable residual voxel set to historical compliance space data to obtain the regulatory conflict body, and calculate the confidence level of illegal construction evidence collection based on the roof component anomaly score, thermal activity time fingerprint and regulatory conflict body; Step 7: Sort the confidence scores of the illegal construction evidence collection and divide them into quantile intervals. Generate illegal construction evidence collection packages for the regions located in abnormal quantile intervals.

[0007] Optionally, step one specifically includes: Collect historical compliant spatial data of the target area, including historical orthophotos, historical oblique images, building approval outline data, building setback data, height restriction surface data, and setback line data; A two-dimensional surface reference plane is established based on the historical orthophoto, and the geographic coordinates of each pixel in the historical orthophoto are mapped to a unified plane coordinate system to form a set of surface grid points. Based on the historical oblique images, feature points of the building facade and roof edge are extracted. Spatial back projection is performed on the feature points of the building facade and roof edge in combination with image imaging parameters and exterior orientation elements to generate a multi-view three-dimensional point set. Spatial alignment is performed between the multi-view 3D point set and the surface grid point set, and coordinate averaging is performed on points from different viewpoints to form a 3D point cloud of building history. Based on the building approval outline data, spatial clipping is performed on the historical 3D point cloud of the building, deleting points located outside the approval outline and retaining points located inside the approval outline to generate a valid building point cloud; Based on the height restriction surface data, obtain the height restriction value of the corresponding spatial location, compare the height value of each point in the legal building point cloud with the corresponding height restriction value, and delete points whose height value is greater than the corresponding height restriction value; Based on the building setback line data and the setback line data, the distance from each point in the legal building point cloud to the setback line is calculated. Points whose distance is less than the predetermined setback distance and are located within the building setback line are marked as legal points, forming a legal three-dimensional reference body of the building's history.

[0008] Optionally, step two specifically involves: Collect visible light images, oblique images, thermal infrared images, RTK positioning data and IMU attitude data of the current inspection, and align the RTK positioning data and IMU attitude data in time to form a pose data sequence; Based on the pose data sequence, the exterior orientation elements corresponding to each frame of visible light image and oblique image are obtained, and the conversion relationship between image coordinates and geographic coordinates is established by combining the camera intrinsic parameters. Based on the exterior orientation elements and camera intrinsic parameters, spatial back projection is performed on the pixels in the visible light image and oblique image to convert the pixel coordinates into spatial point coordinates, forming the current inspection three-dimensional point set. The current patrol 3D point set is mapped to a unified spatial coordinate system, and coordinate merging is performed on points whose coordinate difference is less than the neighborhood threshold within the spatial neighborhood to form the current patrol point cloud. The current patrol point cloud is spatially aligned with the historical legal 3D reference volume, and points at the same spatial location are matched point by point to calculate the spatial coordinate difference between corresponding points. Each point is marked according to the spatial coordinate difference. Points with a planar difference greater than a first difference threshold are marked as planar change points, and points with a height difference greater than a second difference threshold are marked as height change points, thus forming a three-dimensional residual volume.

[0009] Optionally, step three specifically includes: Obtain the spatial coordinates of each residual voxel in the 3D residual volume, and obtain the exterior orientation elements of each frame image according to the pose data sequence; Based on the exterior orientation elements and camera intrinsic parameters, multi-view projection is performed on each residual voxel to map the spatial coordinates of each residual voxel to multiple image planes, thereby obtaining the corresponding projected pixel coordinates. Based on the projected pixel coordinates, the grayscale values ​​and texture features of the corresponding regions in each image are extracted to form a feature vector; Cosine similarity is calculated for the feature vectors of the same residual voxel in different images to obtain multi-view feature similarity values; Residual voxels with feature similarity values ​​greater than the similarity threshold are marked as consistent voxels, while residual voxels with feature similarity values ​​less than the similarity threshold are removed. Clustering is performed on residual voxels labeled as uniform voxels according to spatial connectivity, and the set of residual voxels in the same connected region is defined as the stable set of residual voxels.

[0010] Optionally, step four specifically involves: The stable residual voxel set is projected onto the roof plane and component segmentation is performed to obtain multiple roof candidate components; Extract the component structure sequence for each candidate roof component, the component structure sequence including the component center coordinate sequence, component height sequence, component area sequence, component boundary direction sequence, and component adjacency sequence; The component structure sequence is input into the improved OneFlow framework, which includes a structure sequence encoding module, a structure generation stream, and an offset evolution stream. The structural sequence encoding module normalizes and concatenates the component center coordinate sequence, component height sequence, component area sequence, component boundary direction sequence, and component adjacency sequence to form a component structure representation vector; The structure generation flow includes multiple layers of first reversible affine coupling units. Each layer of first reversible affine coupling unit divides the component structure representation vector into a first sub-vector and a second sub-vector. The first sub-vector is input into a first parameter network, and a first scaling vector and a first translation vector are output. A potential path vector is generated based on the first scaling vector, the first translation vector, and the second sub-vector. The structure generation probability is calculated based on the standard normal density of the potential path vector and the Jacobian determinant of the first reversible affine coupling unit. The offset evolution flow includes a structural differential coding unit and a multi-layer second reversible affine coupling unit. The structural differential coding unit performs same-dimensional difference between the component structure representation vector and the legal component representation vector of the corresponding roof area in the historical legal three-dimensional reference volume to form a structural offset representation vector. Each layer of the second reversible affine coupling unit divides the structure offset representation vector into a third sub-vector and a fourth sub-vector. The third sub-vector is input into the second parameter network, and the output is a second scaling vector and a second translation vector. The offset path potential vector is calculated based on the second scaling vector, the second translation vector and the fourth sub-vector. The structure offset probability is calculated based on the standard normal density of the offset path potential vector and the Jacobian determinant of the second reversible affine coupling unit. Output the structure generation probability and the structure offset probability.

[0011] Optionally, step five specifically includes: The structure generation probability and structure offset probability are obtained, and the two types of probabilities corresponding to the same roof candidate component are matched accordingly. Calculate the difference between the structural offset probability and the structural generation probability, and divide the difference by the sum of the structural generation probability and a constant term to obtain the component anomaly value; Normalize the component anomaly values ​​of all candidate roof components, subtract the minimum value from each component anomaly value and divide by the difference between the maximum and minimum values ​​to obtain the roof component anomaly score. Multiple thermal infrared images are acquired and arranged in chronological order to form a thermal infrared image sequence. Temperature normalization processing is performed on the thermal infrared image sequence to map the temperature values ​​of the images at each time point to a unified temperature range. Based on the spatial coordinates of the stable residual voxel set, the corresponding region is mapped onto the thermal infrared images of each period, and the pixel temperature value of the corresponding region is extracted. The pixel temperature values ​​are arranged in chronological order to form a temperature change sequence; The temperature difference sequence between adjacent time points is calculated based on the temperature change sequence, and a linear fit is performed on the temperature change sequence to obtain the temperature change slope, which together with the average temperature and the temperature difference amplitude constitutes the thermal activity time fingerprint.

[0012] Optionally, step six specifically includes: Obtain the spatial coordinates of a stable set of residual voxels, and map the spatial coordinates to the corresponding spatial areas of building approval outlines, building red lines, height restriction surfaces, and setback lines in historical compliant spatial data; For each residual voxel, determine the positional relationship between its spatial coordinates and the building approval outline, and mark residual voxels located outside the approval outline as outline conflict voxels. Obtain the height value of the residual voxel and compare it with the height limit value of the height limit surface at the corresponding spatial location. Mark the residual voxel with the height value greater than the height limit value as the height limit conflict voxel. Calculate the Euclidean distance from the residual voxel to the setback line, and mark the residual voxels whose distance is less than the setback distance as setback conflict voxels; Contour conflict voxels, height limit conflict voxels, and boundary conflict voxels are merged to form a regulatory conflict voxel; Based on the spatial coordinates of the stable residual voxel set, the roof component anomaly score and thermal activity time fingerprint are correlated with the corresponding area of ​​the regulatory conflict body; The abnormal scores of roof components, thermal activity time fingerprints, and regulatory conflict bodies were normalized, and the normalized values ​​were summed to obtain the confidence level of evidence collection for illegal construction.

[0013] Optionally, step seven specifically includes: Obtain the confidence level values ​​for illegal construction evidence collection corresponding to each spatial area, and sort the spatial areas according to the value to form a sequence arranged in ascending order of confidence level; Based on the sorted sequence, the quantile position is calculated according to the proportion of the sorted sequence number of each spatial region in the sequence to the total number, and the sequence is divided into multiple quantile intervals; Spatial regions with quantile positions higher than a set proportion are marked as abnormal regions; For each abnormal region, extract the corresponding spatial coordinate range, three-dimensional residual voxel set, and roof component anomaly score; Extract the temperature change sequence of the abnormal region in multiple thermal infrared images and the corresponding thermal activity time fingerprint; Extract the regulatory conflict information corresponding to the abnormal area, including the spatial distribution of contour conflict voxels, height limit conflict voxels, and boundary conflict voxels; The spatial coordinate range, three-dimensional residual voxel set, roof component anomaly score, thermal activity time fingerprint, and regulatory conflict information are encapsulated according to data fields to form an illegal construction evidence package.

[0014] The beneficial effects of this invention are: This invention constructs a multimodal 3D spatial representation system integrating visible light imagery, oblique imagery, thermal infrared imagery, and pose data. It combines unified spatial registration and differential modeling of historical compliance spatial data with current inspection data. Addressing the shortcomings of existing UAV inspection methods, such as the lack of precise 3D benchmarks, difficulty in quantifying structural changes, and challenges in the collaborative utilization of multi-source information, this invention proposes a spatial anomaly extraction mechanism based on 3D residuals and multi-view consistency screening. This significantly improves the positioning accuracy and stability for complex building structural changes. In the structural analysis stage, an improved OneFlow framework is introduced. Through a dual-path modeling mechanism of structural sequence encoding, structural generation flow, and migration evolution flow, the generation rules and migration process of roof components are jointly characterized, enabling the identification of anomalies. The probabilistic representation and fine differentiation of structures effectively improve the ability to identify concealed illegal constructions and complex component combinations. In the multimodal information fusion stage, the time series information of thermal infrared images is correlated with the spatial residual region, and thermal activity time fingerprints are constructed through temperature change sequences to enhance the ability to distinguish continuously used structures. In the comprehensive judgment stage, a unified quantitative assessment of spatial violations, structural anomalies, and behavioral characteristics is achieved through the construction of regulatory conflict bodies and a multi-factor normalization fusion mechanism. Abnormal areas are screened through quantile interval sorting to improve the stability and consistency of illegal construction judgment. Finally, an illegal construction evidence package containing spatial range, structural information, thermal behavioral characteristics, and regulatory conflict information is formed, enabling high-precision identification and traceable evidence collection of urban illegal constructions. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the overall process of a method for inspecting and collecting evidence of illegal construction in cities using drones based on multimodal fusion, as proposed in this invention. Figure 2 This is a schematic diagram of the improved OneFlow framework structure in the UAV urban illegal construction inspection and evidence collection method based on multimodal fusion proposed in this invention. Detailed Implementation

[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0017] refer to Figures 1-2 A method for detecting and collecting evidence of illegal construction in cities using drones based on multimodal fusion includes the following steps: Step 1: Collect historical compliant spatial data of the target area and perform 3D reconstruction to obtain the historical legal 3D reference volume of the building; Step 2: Collect the visible light image, oblique image, thermal infrared image, RTK positioning data and IMU attitude data of the current inspection, perform spatial registration on the visible light image and oblique image, and perform difference with the historical legal 3D reference volume to obtain the 3D residual volume; Step 3: Perform multi-view consistency screening on the 3D residual volume to obtain a stable set of residual voxels; Step 4: Project the stable residual voxel set onto the roof plane and segment the components, extract the component structure sequence, input the component structure sequence into the improved OneFlow framework, perform generation path mapping and offset path mapping on the component structure sequence, and obtain the component structure generation probability and structure offset probability of the roof. Step 5: Calculate the anomaly score of the roof components based on the structure generation probability and the structure offset probability, and normalize the multi-period thermal infrared images to extract the temperature change sequence of the region corresponding to the stable residual voxel set, thereby obtaining the thermal activity time fingerprint. Step 6: Map the stable residual voxel set to historical compliance spatial data to obtain the regulatory conflict body, and calculate the confidence level of illegal construction evidence collection based on the roof component anomaly score, thermal activity time fingerprint and regulatory conflict body; Step 7: Sort the confidence levels of illegal construction evidence collection and divide them into quantile intervals. Generate illegal construction evidence collection packages for areas located in abnormal quantile intervals.

[0018] In this embodiment, step one specifically includes: Collect historical compliant spatial data for the target area, including historical orthophotos, historical oblique images, building approval outline data, building setback data, height restriction surface data, and setback line data; A two-dimensional surface reference plane is established based on historical orthophotos, and the geographic coordinates of each pixel in the historical orthophotos are mapped to a unified plane coordinate system to form a set of surface grid points. Based on historical oblique images, feature points of building facade and roof edge are extracted. Spatial back projection is performed on the feature points of building facade and roof edge in combination with image imaging parameters and exterior orientation elements to generate multi-view three-dimensional point sets. Spatial alignment is performed between the multi-view 3D point set and the surface grid point set, and coordinate averaging is performed on points from different viewpoints to form a 3D point cloud of building history. Based on the building approval outline data, spatial clipping is performed on the historical 3D point cloud of the building, deleting points outside the approval outline and retaining points inside the approval outline to generate a valid building point cloud; Based on the height restriction surface data, obtain the height restriction value of the corresponding spatial location, compare the height value of each point in the legal building point cloud with the corresponding height restriction value, and delete the points whose height value is greater than the corresponding height restriction value; Based on the building setback line data and the setback line data, the distance from each point in the legal building point cloud to the setback line is calculated. Points whose distance is less than the predetermined setback distance and are located within the building setback line are marked as legal points, forming a legal three-dimensional reference body of the building's history.

[0019] In this implementation, the image imaging parameters include a focal length of 35 mm, a pixel size of 0.004 mm, and principal point coordinates of (0, 0). The exterior orientation elements include the camera's spatial position coordinates and attitude angle, with the attitude angle ranging from -30 degrees to +30 degrees. When performing spatial alignment between the multi-view 3D point set and the surface grid point set, the spatial neighborhood range is 0.5 meters to 1.5 meters. Within this range, the coordinate components of points from different viewpoints are summed separately and divided by the number of points to obtain the average coordinates. The height limit value corresponding to the height limit surface is in the range of 20 meters to 100 meters, and the height value of each point is compared with the height limit value of the corresponding location. The distance from the point to the setback line is calculated using Euclidean distance, and the distance value is the square root of the sum of the squares of the differences between the point coordinates and the coordinates of the nearest point on the setback line. The setback distance is 3 meters to 10 meters.

[0020] In this embodiment, step two specifically includes: Collect visible light images, oblique images, thermal infrared images, RTK positioning data and IMU attitude data of the current inspection, and align the RTK positioning data and IMU attitude data in time to form a pose data sequence; Based on the pose data sequence, the exterior orientation elements corresponding to each frame of visible light image and oblique image are obtained, and the transformation relationship between image coordinates and geographic coordinates is established by combining the camera intrinsic parameters. Based on the exterior orientation elements and camera intrinsic parameters, spatial back projection is performed on the pixels in the visible light image and oblique image to convert the pixel coordinates into spatial point coordinates, forming the current inspection 3D point set; Map the current patrol 3D point set to a unified spatial coordinate system, and merge the coordinates of points whose coordinate difference is less than the neighborhood threshold within the spatial neighborhood to form the current patrol point cloud. Align the current patrol point cloud with the historical legal 3D reference volume in execution space, perform point-by-point matching for points at the same spatial location, and calculate the spatial coordinate difference between corresponding points; Each point is marked based on the spatial coordinate difference. Points with a planar difference greater than the first difference threshold are marked as planar change points, and points with a height difference greater than the second difference threshold are marked as height change points, thus forming a three-dimensional residual volume.

[0021] In this implementation, the time alignment of the pose data sequence uses a timestamp difference of no more than 0.01 seconds as the matching condition; the camera intrinsic parameters include a focal length of 35 mm and a pixel size of 0.004 mm, and the exterior orientation elements include the camera spatial coordinates and attitude angles, with the attitude angles ranging from -30 degrees to +30 degrees; during the spatial back projection process, the pixel coordinates are combined with the camera intrinsic parameters and exterior orientation elements to calculate the corresponding spatial point coordinates; the spatial neighborhood range is 0.5 meters to 1.0 meters, and when the Euclidean distance between two points is less than this range, coordinate merging is performed by summing the corresponding point coordinates and dividing by the number of points; the planar direction difference threshold is 0.3 meters to 0.8 meters, and the height direction difference threshold is 0.5 meters to 1.5 meters, and the points are marked according to the difference thresholds.

[0022] In this embodiment, step three specifically includes: Obtain the spatial coordinates of each residual voxel in the 3D residual volume, and obtain the exterior orientation elements of each frame image according to the pose data sequence; Multi-view projection is performed on each residual voxel based on the exterior orientation elements and camera intrinsic parameters, and the spatial coordinates of each residual voxel are mapped to multiple image planes to obtain the corresponding projected pixel coordinates. Based on the projected pixel coordinates, the grayscale values ​​and texture features of the corresponding regions in each image are extracted to form a feature vector; Cosine similarity is calculated for the feature vectors of the same residual voxel in different images to obtain multi-view feature similarity values; Residual voxels with feature similarity values ​​greater than the similarity threshold are marked as consistent voxels, while residual voxels with feature similarity values ​​less than the similarity threshold are removed. Clustering is performed on residual voxels labeled as uniform voxels according to spatial connectivity, and the set of residual voxels in the same connected region is defined as the stable set of residual voxels.

[0023] In this implementation, the camera's intrinsic parameters include a focal length of 35 mm and a pixel size of 0.004 mm, while the exterior orientation elements include the camera's spatial coordinates and attitude angle. During multi-view projection, the residual voxel spatial coordinates are combined with the camera's intrinsic parameters and exterior orientation elements to calculate the corresponding projected pixel coordinates. The feature vector consists of grayscale values ​​and texture direction gradients, where the texture direction gradient is calculated using the grayscale difference between adjacent pixels. The feature similarity is calculated using cosine similarity, by dividing the dot product of feature vectors from different viewpoints by the product of their magnitudes. The similarity threshold is set between 0.7 and 0.9. Spatial connectivity is determined using a 26-neighborhood approach, where adjacent voxels are considered connected when the spatial distance between them is less than 0.5 meters. During clustering, regions with more than 50 connected voxels are retained.

[0024] In this embodiment, step four specifically includes: The stable residual voxel set is projected onto the roof plane and component segmentation is performed to obtain multiple roof candidate components; Extract the component structure sequence for each candidate roof component. The component structure sequence includes the component center coordinate sequence, component height sequence, component area sequence, component boundary direction sequence, and component adjacency sequence. The component structure sequence is input into the improved OneFlow framework, which includes a structure sequence encoding module, a structure generation stream, and an offset evolution stream. The structural sequence encoding module performs normalization and splicing on the component center coordinate sequence, component height sequence, component area sequence, component boundary direction sequence, and component adjacency sequence to form a component structural representation vector; The structure generation flow includes multiple layers of first reversible affine coupling units. Each layer of first reversible affine coupling unit divides the component structure representation vector into a first sub-vector and a second sub-vector. The first sub-vector is input into a first parameter network, and the output is a first scaling vector and a first translation vector. The potential vector of the generation path is calculated based on the first scaling vector, the first translation vector, and the second sub-vector. The structure generation probability is calculated based on the standard normal density of the potential vector of the generation path and the Jacobian determinant of the first reversible affine coupling unit. The offset evolution flow includes a structural differential coding unit and a multi-layer second reversible affine coupling unit. The structural differential coding unit performs the same-dimensional difference between the component structural representation vector and the legal component representation vector of the corresponding roof region in the historical legal three-dimensional reference volume to form a structural offset representation vector. Each layer of the second reversible affine coupling unit divides the structure offset representation vector into a third sub-vector and a fourth sub-vector. The third sub-vector is input into the second parameter network, and the output is a second scaling vector and a second translation vector. The offset path potential vector is calculated based on the second scaling vector, the second translation vector and the fourth sub-vector. The structure offset probability is calculated based on the standard normal density of the offset path potential vector and the Jacobian determinant of the second reversible affine coupling unit. Output the structure generation probability and the structure offset probability.

[0025] In this implementation, the component center coordinate sequence is represented by three-dimensional coordinates, the height sequence ranges from 0.5 to 50 meters, the area sequence ranges from 1 to 200 square meters, the boundary direction is represented by angles from 0 to 360 degrees, and the adjacency sequence is represented by a binary matrix. Normalization is performed using a linear scaling method with a minimum value of 0 and a maximum value of 1. Both the structure generation flow and the offset evolution flow contain 4 to 8 layers of reversible affine coupling units, with each layer dividing the input vector into two parts of equal length. The first parameter network and the second parameter network both adopt a three-layer fully connected structure, with 128, 64, and 32 hidden layer nodes, respectively. The dimensions of the latent vectors for the generation path and the latent vectors for the offset path are both 32. The Jacobian determinant is calculated using the product of the scaling vector elements of each layer. The difference in each dimension in the structural differential coding unit is controlled within the range of -5 to +5. Both the improved OneFlow framework and the traditional OneFlow framework belong to the probabilistic flow model based on invertible transformation. They both perform multi-layer invertible mapping on the input data to map the high-dimensional feature space to the latent space, and calculate the sample probability density based on the standard distribution in the latent space. Structurally, they both use affine coupling units as the basic computational units, and achieve invertible transformation by performing scaling and translation operations on sub-vectors. They also calculate the relationship between the probability densities before and after the transformation using the Jacobian determinant, thereby completing the modeling of the input data distribution. The improved OneFlow framework in this implementation introduces a structure sequence encoding module, a structure generation flow, and an offset evolution flow on the basis of the traditional structure. The structure sequence encoding module uniformly encodes the component center coordinates, height, area, boundary direction, and adjacency relationship. The structure generation flow performs a layer-by-layer reversible mapping on the encoded structure sequence to obtain the latent representation of the generation path. The offset evolution flow adds a structure differential encoding unit at the input end, performs the same-dimensional difference between the current structure representation and the historical legal structure representation to obtain the structure offset representation, and performs offset path mapping through an independent reversible affine coupling unit to achieve separate modeling of the generation path and the offset path. Through the above improvements, the improved OneFlow framework can simultaneously characterize the structural generation rules and offset evolution features of roof components, enabling the model to not only identify whether the structure conforms to the existing distribution, but also to determine the direction and magnitude of the structure's change relative to its historical legal state. Compared to the traditional OneFlow method that relies on only a single probability distribution, the dual-flow structure in this implementation improves the ability to distinguish complex roof structure changes and reduces misjudgments caused by obstruction, noise, or temporary stacking, thereby improving the stability and accuracy of illegal construction identification.

[0026] In this embodiment, step five specifically includes: Obtain the structure generation probability and structure offset probability, and perform corresponding matching of the two types of probabilities for the same roof candidate component; Calculate the difference between the structural offset probability and the structural generation probability, and divide the difference by the sum of the structural generation probability and the constant term to obtain the component anomaly value; Normalize the component anomaly values ​​of all candidate roof components, subtract the minimum value from each component anomaly value and divide by the difference between the maximum and minimum values ​​to obtain the roof component anomaly score. Multiple thermal infrared images were acquired and arranged in chronological order to form a thermal infrared image sequence. Temperature normalization processing was performed on the thermal infrared image sequence to map the temperature values ​​of the images at each time point to a unified temperature range. Based on the spatial coordinates of the stable residual voxel set, the corresponding region is mapped onto the thermal infrared images of each period, and the pixel temperature value of the corresponding region is extracted. The pixel temperature values ​​are arranged in chronological order to form a temperature change sequence; The temperature difference sequence between adjacent time points is calculated based on the temperature change sequence, and a linear fit is performed on the temperature change sequence to obtain the temperature change slope. Together with the average temperature and the amplitude of the temperature difference, they constitute the thermal activity time fingerprint.

[0027] In this implementation, both the probability of structure generation and the probability of structure displacement are between 0 and 1, and the constant term is between 0.01 and 0.05. The component anomaly value is obtained by dividing the difference between the probability of structure displacement and the probability of structure generation by the sum of the probability of structure generation and the constant term. The normalization process adopts a linear mapping method with a minimum value of 0 and a maximum value of 1. The temperature range of the thermal infrared image is between -20 and 80 degrees Celsius, and the normalization interval is set to 0 to 1. The time interval of the temperature change sequence is between 1 and 5 minutes. The temperature difference sequence is obtained by subtracting the temperature values ​​at adjacent time points. The slope of the linear fitting is calculated using the least squares method. The average temperature is obtained by summing the temperature values ​​at each time point and dividing by the number of sampling points. The amplitude of the temperature difference is the difference between the maximum and minimum values ​​in the difference sequence.

[0028] In this embodiment, step six specifically includes: Obtain the spatial coordinates of a stable set of residual voxels, and map the spatial coordinates to the corresponding spatial areas of building approval outlines, building red lines, height restriction surfaces, and setback lines in historical compliant spatial data; For each residual voxel, determine the positional relationship between its spatial coordinates and the building approval outline, and mark residual voxels located outside the approval outline as outline conflict voxels. Obtain the height value of the residual voxel and compare it with the height limit value of the height limit surface at the corresponding spatial location. Mark the residual voxel with the height value greater than the height limit value as the height limit conflict voxel. Calculate the Euclidean distance from the residual voxel to the setback line, and mark the residual voxels whose distance is less than the setback distance as setback conflict voxels; Contour conflict voxels, height limit conflict voxels, and boundary conflict voxels are merged to form a regulatory conflict voxel; Based on the spatial coordinates of the stable residual voxel set, the roof component anomaly score and thermal activity time fingerprint are correlated with the corresponding area of ​​the regulatory conflict body; The abnormal scores of roof components, thermal activity time fingerprints, and regulatory conflict bodies were normalized, and the normalized values ​​were summed to obtain the confidence level of evidence collection for illegal construction.

[0029] In this implementation, a unified geographic coordinate system is used for spatial coordinate mapping. The building approval outline, building red line, height restriction surface, and setback line are all converted to the same coordinate reference. The height restriction value ranges from 20 to 100 meters. The setback distance ranges from 3 to 10 meters. The distance from the residual voxel to the setback line is calculated using three-dimensional Euclidean distance. The roof component anomaly score ranges from 0 to 1. The average temperature in the thermal activity time fingerprint is a normalized value ranging from -20 to 80 degrees Celsius, and the temperature difference ranges from 0 to 30 degrees Celsius. The normalization process uses a linear mapping with a minimum value of 0 and a maximum value of 1. The confidence level for obtaining evidence of illegal construction is obtained by adding the corresponding values ​​of the normalized roof component anomaly score, thermal activity time fingerprint, and regulatory conflict body, with a total value ranging from 0 to 3.

[0030] In this embodiment, step seven specifically includes: Obtain the confidence level values ​​for illegal construction evidence collection corresponding to each spatial area, and sort the spatial areas according to the value to form a sequence arranged in ascending order of confidence level; Based on the sorted sequence, the quantile position is calculated according to the proportion of the sorted sequence number of each spatial region in the sequence to the total number, and the sequence is divided into multiple quantile intervals; Spatial regions with quantile positions higher than a set proportion are marked as abnormal regions; For each abnormal region, extract the corresponding spatial coordinate range, three-dimensional residual voxel set, and roof component anomaly score; Extract the temperature change sequence of the abnormal region in multiple thermal infrared images and the corresponding thermal activity time fingerprint; Extract the regulatory conflict information corresponding to the abnormal area, including the spatial distribution of contour conflict voxels, height limit conflict voxels, and boundary conflict voxels; The spatial coordinate range, three-dimensional residual voxel set, roof component anomaly score, thermal activity time fingerprint, and regulatory conflict information are encapsulated according to data fields to form an illegal construction evidence package.

[0031] In this implementation, the confidence level for illegal construction evidence collection ranges from 0 to 3; during the sorting process, values ​​are arranged from smallest to largest; quantile positions are calculated by dividing the sorting number by the total number of spatial regions; quantile intervals are divided into 4 to 10 equal intervals; the quantile ratio corresponding to abnormal regions is between 0.8 and 0.95, meaning regions with a sorting position ratio greater than this range are marked as abnormal regions; the spatial coordinate range is represented using a minimum bounding box; the three-dimensional residual voxel set is represented using a grid with voxel sizes of 0.2 meters to 0.5 meters; the thermal activity time fingerprint includes average temperature, temperature difference amplitude, and temperature change slope; the number of various conflicting voxels in the regulatory conflict body is counted and recorded; the illegal construction evidence collection package is encapsulated according to the field structure, including coordinate fields, residual fields, abnormal score fields, thermal fingerprint fields, and conflict fields.

[0032] Example 1: To verify the feasibility of this invention in a real-world scenario, it was applied to the task of inspecting and collecting evidence of illegal constructions in a typical old urban area of ​​a coastal city. This area has a high building density and complex roof structures, with numerous historical additions, temporary sheds, and concealed illegal structures. Traditional manual inspections rely on ground observation and simple image comparison, resulting in low inspection efficiency, high misjudgment rates, and insufficient evidence. Furthermore, due to severe obstruction between buildings, significant variations in lighting conditions, and the high degree of similarity between some illegal structures and legal buildings, existing methods based on single images or simple difference methods are insufficient for accurately identifying illegal structures, especially in scenarios involving roof extensions and mezzanine additions.

[0033] In this embodiment, a three-dimensional reference volume is first established for the target area using historical compliant spatial data to ensure that subsequent inspection data has a unified spatial reference. The UAV inspects the target area according to a predetermined route, collecting visible light, oblique, and thermal infrared images, while simultaneously recording RTK positioning data and IMU attitude data. Through spatial registration and 3D reconstruction, the current inspection data is converted into a 3D point cloud format and differentially processed with the historical 3D reference volume to generate a 3D residual volume. Compared to traditional two-dimensional differential methods, this process can directly reflect changes in building height and spatial structure, effectively solving the problem that traditional methods cannot identify three-dimensional expansions.

[0034] Subsequently, a multi-view consistency screening was performed on the 3D residual voxels, removing noise points that appeared only in a single viewpoint and retaining only residual voxels with consistent projection characteristics across multiple views, thus forming a stable set of residual voxels. This process significantly reduced false detections caused by tree occlusion, lighting variations, and temporary objects. In the roof structure analysis phase, the stable residual voxels were projected onto the roof plane and component segmentation was performed to extract the component structure sequence. A dual-path modeling of the structure sequence was then performed using the improved OneFlow framework, calculating the structure generation probability and the structure offset probability separately, thereby quantifying whether the current structure conforms to the generation patterns of historical legitimate structures and the degree of offset.

[0035] In terms of thermal infrared data processing, multiple thermal infrared images are normalized, and temperature change sequences corresponding to residual regions are extracted. A thermal activity temporal fingerprint is constructed using average temperature, temperature fluctuation amplitude, and change slope. This fingerprint reflects whether the structure has been continuously used, effectively distinguishing between temporary stockpiles and actually used illegal structures. Furthermore, residual voxels are mapped to approved outlines, height restriction surfaces, and setback lines to generate a regulatory conflict body. This body is then combined with structural anomaly scores and the thermal activity temporal fingerprint for comprehensive calculation to obtain the confidence level for illegal construction evidence collection. Finally, by ranking the confidence levels and dividing them into quantile intervals, high-risk areas are selected, and an illegal construction evidence collection package containing spatial location, structural information, thermal behavior, and regulatory conflict information is generated, enabling traceable evidence collection.

[0036] To verify the effectiveness of the present invention, a comparative experiment was conducted between the present method and the traditional method. 1,000 roof structures were selected as samples, including 420 real illegal structures. The experimental results are shown in Table 1.

[0037] Table 1 Performance Comparison of Multimodal Illegal Construction Detection Methods

[0038] As shown in Table 1, this invention significantly outperforms traditional two-dimensional difference methods and single-modal deep learning methods in all core indicators, demonstrating strong comprehensive performance advantages. In terms of detection accuracy, this invention achieves 93.6%, an improvement of over 20 percentage points compared to traditional methods, indicating that the processing method based on three-dimensional residual volume and multimodal fusion can more comprehensively capture changes in building structures. The false positive rate and false negative rate are reduced to 5.3% and 6.8% respectively, indicating that multi-view consistency screening and structural modeling effectively suppress interference caused by factors such as occlusion and lighting changes.

[0039] In terms of structural recognition capabilities, this invention achieves a recognition rate exceeding 90% for both 3D structures and roof additions, significantly outperforming comparative methods. This demonstrates that the improved OneFlow framework accurately characterizes structural generation and offset features. Simultaneously, the thermal behavior recognition accuracy reaches 89.7%, resolving the issue of traditional methods' inability to utilize thermal infrared information. Regarding regulatory conflict determination, this invention achieves a 94.3% accuracy rate, indicating that multi-source data fusion and spatial mapping mechanisms improve determination accuracy.

[0040] Furthermore, the stability of the evidence confidence level is significantly improved, with a standard deviation of only 0.07, indicating stronger consistency of the results. In terms of processing efficiency, this invention maintains a high processing speed while improving accuracy, making it suitable for practical engineering applications. Overall, this invention demonstrates significant advantages in accuracy, stability, and practicality.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for unmanned aerial vehicle (UAV) patrol and evidence collection for illegal construction in cities based on multimodal fusion, characterized in that, Includes the following steps: Step 1: Collect historical compliant spatial data of the target area and perform 3D reconstruction to obtain the historical legal 3D reference volume of the building; Step 2: Collect the visible light image, oblique image, thermal infrared image, RTK positioning data and IMU attitude data of the current inspection, perform spatial registration on the visible light image and oblique image, and perform difference with the historical legal three-dimensional reference volume to obtain the three-dimensional residual volume; Step 3: Perform multi-view consistency screening on the three-dimensional residual volume to obtain a stable set of residual voxels; Step 4: Project the stable residual voxel set onto the roof plane and perform component segmentation, extract the component structure sequence, input the component structure sequence into the improved OneFlow framework, perform generation path mapping and offset path mapping on the component structure sequence, and obtain the component structure generation probability and structure offset probability of the roof. Step 5: Calculate the roof component anomaly score based on the structure generation probability and the structure offset probability, and normalize the multi-period thermal infrared images to extract the temperature change sequence of the region corresponding to the stable residual voxel set, thereby obtaining the thermal activity time fingerprint; Step 6: Map the stable residual voxel set to historical compliance space data to obtain the regulatory conflict body, and calculate the confidence level of illegal construction evidence collection based on the roof component anomaly score, thermal activity time fingerprint and regulatory conflict body; Step 7: Sort the confidence scores of the illegal construction evidence collection and divide them into quantile intervals. Generate illegal construction evidence collection packages for the regions located in abnormal quantile intervals.

2. The method for urban illegal construction inspection and evidence collection based on multimodal fusion using unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Step one specifically involves: Collect historical compliant spatial data of the target area, including historical orthophotos, historical oblique images, building approval outline data, building setback data, height restriction surface data, and setback line data; A two-dimensional surface reference plane is established based on the historical orthophoto, and the geographic coordinates of each pixel in the historical orthophoto are mapped to a unified plane coordinate system to form a set of surface grid points. Based on the historical oblique images, feature points of the building facade and roof edge are extracted. Spatial back projection is performed on the feature points of the building facade and roof edge in combination with image imaging parameters and exterior orientation elements to generate a multi-view three-dimensional point set. Spatial alignment is performed between the multi-view 3D point set and the surface grid point set, and coordinate averaging is performed on points from different viewpoints to form a 3D point cloud of building history. Based on the building approval outline data, spatial clipping is performed on the historical 3D point cloud of the building, deleting points located outside the approval outline and retaining points located inside the approval outline to generate a valid building point cloud; Based on the height restriction surface data, obtain the height restriction value of the corresponding spatial location, compare the height value of each point in the legal building point cloud with the corresponding height restriction value, and delete points whose height value is greater than the corresponding height restriction value; Based on the building setback line data and the setback line data, the distance from each point in the legal building point cloud to the setback line is calculated. Points whose distance is less than the predetermined setback distance and are located within the building setback line are marked as legal points, forming a legal three-dimensional reference body of the building's history.

3. The method for urban illegal construction inspection and evidence collection based on multimodal fusion using unmanned aerial vehicles (UAVs) according to claim 1, characterized in that, Step two specifically involves: Collect visible light images, oblique images, thermal infrared images, RTK positioning data and IMU attitude data of the current inspection, and align the RTK positioning data and IMU attitude data in time to form a pose data sequence; Based on the pose data sequence, the exterior orientation elements corresponding to each frame of visible light image and oblique image are obtained, and the conversion relationship between image coordinates and geographic coordinates is established by combining the camera intrinsic parameters. Based on the exterior orientation elements and camera intrinsic parameters, spatial back projection is performed on the pixels in the visible light image and oblique image to convert the pixel coordinates into spatial point coordinates, forming the current inspection three-dimensional point set. The current patrol 3D point set is mapped to a unified spatial coordinate system, and coordinate merging is performed on points whose coordinate difference is less than the neighborhood threshold within the spatial neighborhood to form the current patrol point cloud. The current patrol point cloud is spatially aligned with the historical legal 3D reference volume, and points at the same spatial location are matched point by point to calculate the spatial coordinate difference between corresponding points. Each point is marked according to the spatial coordinate difference. Points with a planar difference greater than a first difference threshold are marked as planar change points, and points with a height difference greater than a second difference threshold are marked as height change points, thus forming a three-dimensional residual volume.

4. The method for unmanned aerial vehicle (UAV) urban illegal construction inspection and evidence collection based on multimodal fusion according to claim 1, characterized in that, Step three specifically involves: Obtain the spatial coordinates of each residual voxel in the 3D residual volume, and obtain the exterior orientation elements of each frame image according to the pose data sequence; Based on the exterior orientation elements and camera intrinsic parameters, multi-view projection is performed on each residual voxel to map the spatial coordinates of each residual voxel to multiple image planes, thereby obtaining the corresponding projected pixel coordinates. Based on the projected pixel coordinates, the grayscale values ​​and texture features of the corresponding regions in each image are extracted to form a feature vector; Cosine similarity is calculated for the feature vectors of the same residual voxel in different images to obtain multi-view feature similarity values; Residual voxels with feature similarity values ​​greater than the similarity threshold are marked as consistent voxels, while residual voxels with feature similarity values ​​less than the similarity threshold are removed. Clustering is performed on residual voxels labeled as uniform voxels according to spatial connectivity, and the set of residual voxels in the same connected region is defined as the stable set of residual voxels.

5. The method for unmanned aerial vehicle (UAV) urban illegal construction inspection and evidence collection based on multimodal fusion according to claim 1, characterized in that, Step four specifically involves: The stable residual voxel set is projected onto the roof plane and component segmentation is performed to obtain multiple roof candidate components; Extract the component structure sequence for each candidate roof component, the component structure sequence including the component center coordinate sequence, component height sequence, component area sequence, component boundary direction sequence, and component adjacency sequence; The component structure sequence is input into the improved OneFlow framework, which includes a structure sequence encoding module, a structure generation stream, and an offset evolution stream. The structural sequence encoding module performs normalization and splicing on the component center coordinate sequence, component height sequence, component area sequence, component boundary direction sequence, and component adjacency sequence to form a component structure representation vector; The structure generation flow includes multiple layers of first reversible affine coupling units. Each layer of first reversible affine coupling unit divides the component structure representation vector into a first sub-vector and a second sub-vector. The first sub-vector is input into a first parameter network, and a first scaling vector and a first translation vector are output. A potential path vector is generated based on the first scaling vector, the first translation vector, and the second sub-vector. The structure generation probability is calculated based on the standard normal density of the potential path vector and the Jacobian determinant of the first reversible affine coupling unit. The offset evolution flow includes a structural differential coding unit and a multi-layer second reversible affine coupling unit. The structural differential coding unit performs same-dimensional difference between the component structure representation vector and the legal component representation vector of the corresponding roof area in the historical legal three-dimensional reference volume to form a structural offset representation vector. Each layer of the second reversible affine coupling unit divides the structure offset representation vector into a third sub-vector and a fourth sub-vector. The third sub-vector is input into the second parameter network, and the output is a second scaling vector and a second translation vector. The offset path potential vector is calculated based on the second scaling vector, the second translation vector and the fourth sub-vector. The structure offset probability is calculated based on the standard normal density of the offset path potential vector and the Jacobian determinant of the second reversible affine coupling unit. Output the probability of structure generation and the probability of structure offset.

6. The method for unmanned aerial vehicle (UAV) urban illegal construction inspection and evidence collection based on multimodal fusion according to claim 1, characterized in that, Step five specifically involves: The structure generation probability and structure offset probability are obtained, and the two types of probabilities corresponding to the same roof candidate component are matched accordingly. Calculate the difference between the structural offset probability and the structural generation probability, and divide the difference by the sum of the structural generation probability and a constant term to obtain the component anomaly value; Normalize the component anomaly values ​​of all candidate roof components, subtract the minimum value from each component anomaly value and divide by the difference between the maximum and minimum values ​​to obtain the roof component anomaly score. Multiple thermal infrared images are acquired and arranged in chronological order to form a thermal infrared image sequence. Temperature normalization processing is performed on the thermal infrared image sequence to map the temperature values ​​of the images at each time point to a unified temperature range. Based on the spatial coordinates of the stable residual voxel set, the corresponding region is mapped onto the thermal infrared images of each period, and the pixel temperature value of the corresponding region is extracted. The pixel temperature values ​​are arranged in chronological order to form a temperature change sequence; The temperature difference sequence between adjacent time points is calculated based on the temperature change sequence, and a linear fit is performed on the temperature change sequence to obtain the temperature change slope, which together with the average temperature and the temperature difference amplitude constitutes the thermal activity time fingerprint.

7. The method for unmanned aerial vehicle (UAV) urban illegal construction inspection and evidence collection based on multimodal fusion according to claim 1, characterized in that, Step six specifically involves: Obtain the spatial coordinates of a stable set of residual voxels, and map the spatial coordinates to the corresponding spatial areas of building approval outlines, building red lines, height restriction surfaces, and setback lines in historical compliant spatial data; For each residual voxel, determine the positional relationship between its spatial coordinates and the building approval outline, and mark residual voxels located outside the approval outline as outline conflict voxels. Obtain the height value of the residual voxel and compare it with the height limit value of the height limit surface at the corresponding spatial location. Mark the residual voxel with the height value greater than the height limit value as the height limit conflict voxel. Calculate the Euclidean distance from the residual voxel to the setback line, and mark the residual voxels whose distance is less than the setback distance as setback conflict voxels; Contour conflict voxels, height limit conflict voxels, and boundary conflict voxels are merged to form a regulatory conflict voxel; Based on the spatial coordinates of the stable residual voxel set, the roof component anomaly score and thermal activity time fingerprint are correlated with the corresponding area of ​​the regulatory conflict body; The abnormal scores of roof components, thermal activity time fingerprints, and regulatory conflict bodies were normalized, and the normalized values ​​were summed to obtain the confidence level of evidence collection for illegal construction.

8. A method for urban illegal construction inspection and evidence collection using unmanned aerial vehicles (UAVs) based on multimodal fusion as described in claim 1, characterized in that, Step seven specifically involves: Obtain the confidence level values ​​for illegal construction evidence collection corresponding to each spatial area, and sort the spatial areas according to the value to form a sequence arranged in ascending order of confidence level; Based on the sorted sequence, the quantile position is calculated according to the proportion of the sorted sequence number of each spatial region in the sequence to the total number, and the sequence is divided into multiple quantile intervals; Spatial regions with quantile positions higher than a set proportion are marked as abnormal regions; For each abnormal region, extract the corresponding spatial coordinate range, three-dimensional residual voxel set, and roof component anomaly score; Extract the temperature change sequence of the abnormal region in multiple thermal infrared images and the corresponding thermal activity time fingerprint; Extract the regulatory conflict information corresponding to the abnormal area, including the spatial distribution of contour conflict voxels, height limit conflict voxels, and boundary conflict voxels; The spatial coordinate range, three-dimensional residual voxel set, roof component anomaly score, thermal activity time fingerprint, and regulatory conflict information are encapsulated according to data fields to form an illegal construction evidence package.

Images