A Visual Processing-Based Method for Detecting Building Settlement Misalignment Deviation
By constructing a global benchmark library and acquiring multi-view images, a high-precision three-dimensional digital model is generated, which solves the problem of easy deformation of fixed reference points in traditional detection methods and realizes high-precision detection of building settlement and displacement and full-dimensional deformation calculation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- 聊城大学东昌学院
- Filing Date
- 2026-04-08
- Publication Date
- 2026-06-30
Smart Images
Figure CN122306019A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of building deviation detection technology, and particularly relates to a method for detecting building settlement offset deviation based on visual processing. Background Technology
[0002] In the field of building engineering, the detection of building settlement and offset deviation is a core aspect of ensuring building structural safety and preventing engineering accidents. Various buildings in cities are prone to structural deformation in all dimensions due to factors such as geological structure, groundwater, and surrounding construction, placing high demands on the accuracy, comprehensive monitoring capabilities, and timeliness of detection methods. Meanwhile, traditional detection methods rely on periodic comparisons between fixed reference points on the building and surrounding areas within 50 to 100 meters. This method has drawbacks such as the permanent difference in the fixed reference points, the inability to measure large-area settlement, and the inability to measure deviation. Summary of the Invention
[0003] In view of the technical problems existing in the background art, the present invention proposes a method for detecting building settlement offset deviation based on visual processing.
[0004] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:
[0005] S1. Construction of the benchmark landmark library: Determine the target building to be detected and its corresponding detection coverage area, screen natural landmarks and man-made structures outside the detection coverage area that have long-term structural stability, complete the landmark stability classification and screening through AI visual feature analysis, and construct a global benchmark landmark library by obtaining the absolute geodetic coordinates of each landmark.
[0006] S2. Multi-view synchronous image acquisition: Using a mobile visual acquisition platform, multi-view synchronous image acquisition is performed on the entire structure of the target building and all landmarks in the benchmark landmark database to obtain a sequence of images.
[0007] S3. Construction of Global Absolute Coordinate System: Through feature matching algorithm, the feature points of the reference landmarks and the feature points of the target buildings in the sequence image are accurately matched and mismatched are eliminated. The absolute geodetic coordinates of the reference landmark library are used as global constraints to construct a unified global absolute coordinate system.
[0008] S4. Construction of 3D model of target building with absolute coordinates: Under the global absolute coordinate system, based on the sequence of images, a high-precision 3D digital model of the target building to be detected with absolute 3D coordinates is generated, and the absolute 3D coordinate dataset of the feature points of the target building is extracted.
[0009] S5. Multi-period detection and full-dimensional deviation calculation: Repeat steps S2 to S4 according to the preset detection cycle to complete the sequential image acquisition and 3D digital model construction for different detection cycles; use stable landmarks in the benchmark landmark library as rigid registration anchor points to complete the coordinate system registration of the multi-period 3D digital model; calculate the full-dimensional deformation deviation data of the target building through the difference calculation between the multi-period 3D digital model and the absolute 3D coordinate dataset; make compliance judgment on the full-dimensional deformation deviation data according to the preset deviation threshold of the relevant building deformation measurement specifications, output a standardized detection report, and provide graded warnings for abnormal deformations exceeding the threshold.
[0010] Preferably, the step S1 of constructing a global benchmark library specifically includes the following steps:
[0011] S11. Delineate the geological impact boundary of the target building, screen natural landmarks and man-made landmarks outside the detection coverage area, located in different geological structural units from the target building, without direct groundwater hydrological connection, and with a horizontal distance of not less than twice the height of the target building, to form a preliminary set of landmarks;
[0012] S12. For each landmark in the initial landmark set, extract three core indicators—temporal feature invariance, texture richness, and spatial distribution uniformity—through AI visual feature analysis, calculate the temporal variation variance of each landmark, and complete the quantitative scoring of landmark visual stability.
[0013] S13. Based on the visual stability quantitative score, landmarks are divided into Level 1 rigid stable landmarks, Level 2 sub-stable landmarks, and Level 3 unstable landmarks. Only Level 1 and Level 2 landmarks are retained in the benchmark landmark database. Among them, Level 1 rigid stable landmarks account for no less than 60% of the total number of landmarks and are evenly distributed in a ring around the target building.
[0014] Preferably, step S1 also includes a dynamic update mechanism for the benchmark landmark library: after each detection cycle is completed, the coordinate offset of each landmark in the benchmark landmark library is calculated using the current sequence image, and landmarks whose coordinate offset exceeds the preset stability threshold are downgraded and removed; at the same time, new landmarks that meet the stability requirements are added to complete the dynamic update of the benchmark landmark library. The updated landmark library must maintain the requirement of the number and spatial distribution uniformity of first-level rigid stable landmarks.
[0015] Preferably, step S12 includes the following steps:
[0016] S121. Set up an initial set of landmarks and collect multiple historical visual samples for each landmark. Perform preprocessing on each historical visual sample of each landmark in sequence, and then use an AI feature detection algorithm to filter feature points and extract a standardized SIFT feature point set for each landmark.
[0017] S122. Based on the standardized SIFT feature point set, the original value of the temporal feature invariance of the landmark is calculated by integrating the average feature point matching success rate and the average feature descriptor similarity.
[0018] S123. Calculate the texture richness value of the landmark by fusing the entropy value of the gray-level co-occurrence matrix and the average gradient magnitude;
[0019] S124. Calculate the spatial distribution uniformity of landmarks by combining relative position uniformity and global distribution uniformity;
[0020] S125. Calculate the single-period normalized index value for each time series image of the landmark, take the mean of the time series of each index, calculate the time series variation variance, and finally calculate the landmark visual stability quantitative score.
[0021] Preferably, step S2, multi-view synchronous image acquisition, specifically includes:
[0022] Multiple key structural parts of target buildings are set. During the acquisition process, the GNSS-IMU integrated navigation module on the platform is used to acquire the acquisition pose information in real time, ensuring that the image heading overlap between the reference landmark and the key structural parts of the target buildings is ≥85% and the lateral overlap is ≥70%, and each key structural part and the reference landmark are covered by a sequence of images with no less than 3 different shooting perspectives.
[0023] Preferably, the specific implementation of step S3, which uses a feature matching algorithm to accurately match and remove mismatches between reference landmark feature points and target building feature points in the image sequence, includes:
[0024] S31. Extract the SIFT feature point set of the reference landmark and the SIFT feature point set of the target building from the sequence image, and supplement the ORB feature point set of the reference landmark and the ORB feature point set of the target building. Normalize the SIFT descriptor and the ORB descriptor respectively, and weight them to obtain the fused feature descriptor, thereby obtaining the fused feature descriptor subset of the reference landmark and the fused feature descriptor subset of the target building.
[0025] S32. Using the K-nearest neighbor feature matching algorithm, calculate the Hamming distance similarity and Euclidean distance similarity between the fused feature description subset of the benchmark landmark and the fused feature description subset of the target building. Obtain the comprehensive matching similarity by weighted summation, and preset a similarity threshold. Select feature point matching pairs that are greater than or equal to the similarity threshold to form an initial matching pair set.
[0026] S33. Standardize the pixel coordinates of the initial matching pair set, and then perform iterative mismatch removal on the standardized initial matching pair set. In each iteration, randomly select four sets of non-repeating standardized matching pairs from the initial matching pair set as the minimum sample set, solve the transformation homography matrix between the reference landmark and the target building feature points, traverse all standardized matching pairs in the initial matching pair set, calculate the transformation error of each matching pair under the homography matrix, judge the magnitude of the transformation error and the interior point error threshold, if it is less than or equal to the threshold, it is determined as an interior point, count the total number of interior points in this iteration, if the total number of interior points is greater than the previous maximum total number of interior points, update the homography matrix and the maximum total number of interior points, after completing all iterations, use the optimal homography matrix as the benchmark, traverse the initial matching pair set again, remove all exterior points whose transformation error is greater than the interior point error threshold, and obtain the feature point matching pair set after mismatch removal.
[0027] Preferably, the specific implementation of generating a high-precision 3D digital model of the target building with absolute 3D coordinates based on the sequence images in step S4 is as follows:
[0028] Using the global absolute coordinate system as a reference, the absolute three-dimensional coordinates of all target building feature points and benchmark landmark feature points are directly assigned to the corresponding sparse point cloud to generate an initial sparse point cloud model with absolute coordinates; for the feature points of the target building, high-density point cloud reconstruction is performed to obtain a high-precision three-dimensional digital model.
[0029] Preferably, the specific implementation of step S5, which involves calculating the full-dimensional deformation deviation data of the target building through differential calculation between the multi-period three-dimensional digital model and the absolute three-dimensional coordinate dataset, is as follows:
[0030] S51. First, calculate the difference between the spatial coordinates of the feature points of the target building in the current cycle and the previous cycle to obtain the deformation deviation of the part in the X / Y / Z axes in the current cycle, which is recorded as the relative deformation deviation.
[0031] S52. Next, compare the spatial coordinates of the feature points of the target building in the current period with the spatial coordinates of the reference period, and record the deformation deviation as the absolute deformation deviation.
[0032] S53. Simultaneously, based on the relative deformation deviation and the absolute deformation deviation, the overall tilt change and settlement difference change are calculated.
[0033] Compared with existing technologies, the advantages and positive effects of this invention are as follows: This invention breaks through the limitations of traditional building deformation detection relying on near-field fixed reference points, constructing a global benchmark library with AI stability grading and dynamic updating mechanisms. This fundamentally solves industry pain points such as easily deformable reference points, benchmark failure, and the inability to detect large-area uniform settlement, ensuring the long-term rigidity and stability of the detection benchmark. Simultaneously, it integrates multi-feature matching and global constraints with absolute geodetic coordinates to achieve high-precision 3D modeling with absolute coordinates, eliminating visual reconstruction scale drift errors. Through multi-cycle model registration and differential calculation, it completes the calculation of all-dimensional deformations such as building settlement, horizontal offset, and overall tilt, significantly improving detection accuracy, coverage, and the timeliness of safety warnings. Attached Figure Description
[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0035] Figure 1 This is a schematic diagram of the structural process of a visual processing-based method for detecting building settlement offset deviation. Detailed Implementation
[0036] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0037] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.
[0038] In this embodiment, traditional detection methods rely on fixed reference points on buildings and periodically compare measurements with surrounding areas within a 50-100 meter radius. This method suffers from drawbacks such as the permanent inaccuracies of fixed reference points, inability to measure large-area settlement, and inability to measure deviations. This invention discloses a visual processing-based method for detecting building settlement offset deviations. Figure 1 .
[0039] First, a benchmark landmark library is constructed: the target buildings to be detected and their corresponding detection coverage areas are determined, natural landmarks and man-made structures outside the detection coverage areas that have long-term structural stability are screened, landmark stability is graded and screened through AI visual feature analysis, and the absolute geodetic coordinates of each landmark are obtained to construct a global benchmark landmark library.
[0040] Specifically, building a global benchmark library includes the following steps:
[0041] The geological impact boundary of the target building is delineated, and natural and man-made landmarks outside the detection coverage area, located in different geological structural units from the target building, without direct groundwater hydrological connection, and with a horizontal distance of not less than twice the height of the target building are screened to form a preliminary set of landmarks.
[0042] For each landmark in the initial set of landmarks, three core indicators—temporal feature invariance, texture richness, and spatial distribution uniformity—are extracted through AI visual feature analysis. The temporal variation variance of each landmark is calculated to complete the quantitative scoring of landmark visual stability.
[0043] Furthermore, specifically, this is achieved by setting the initial set of landmarks as follows: Where n is the number of initially selected landmarks; multiple historical visual samples are collected for each landmark. t represents the acquisition time; for each historical visual sample of each landmark, adaptive histogram equalization illumination correction, Gaussian filtering noise removal, and perspective transformation viewpoint distortion correction are performed sequentially. Then, feature points are selected using an AI feature detection algorithm to extract the standardized SIFT feature point set for each landmark. Based on standardized SIFT feature point sets Landmarks are calculated by integrating the average feature point matching success rate and the average feature descriptor similarity. Temporal characteristic invariance original value The calculation method is as follows: ,in, , , To calculate the average feature point matching success rate, ,in, The number of pairwise combinations of time series samples. For feature point set The number of correctly matched feature points after removing mismatches using the RANSAC algorithm. Landmarks In the The number of effective feature points in the period image The number of feature points contained therein; Average feature descriptor similarity: ,in, The average number of effective feature points. For reference period images The k-th feature descriptor, Let L be the Hamming distance, and L be the 128-dimensional dimension of the SIFT feature descriptor. To and The matched feature descriptors, where T is the total acquisition time; the average feature point matching success rate and average feature descriptor similarity are normalized and then weighted and summed to obtain the original value of time-series feature invariance. .
[0044] Calculate landmarks by fusing the entropy value of the gray-level co-occurrence matrix and the mean gradient magnitude. Texture richness value The calculation method is as follows: ,in, , The texture entropy value is calculated based on the gray-level co-occurrence matrix. The average gradient magnitude calculated based on the Sobel operator is... and Texture richness values are obtained by performing normalized weighted calculations separately.
[0045] Calculate landmarks by combining relative location uniformity and global distribution uniformity. Spatial distribution uniformity The calculation method is as follows: ,in, , To ensure the uniformity of the landmark's position relative to the target building, the horizontal distance between the landmark and the target building must be at least twice the height of the target building. The uniformity of the global spatial distribution of landmarks within the initial selection set; where, ,in, as a landmark Horizontal distance to the geometric center of the target building The maximum height of the building. The maximum horizontal distance of the initial set of landmarks; Based on spatial distribution entropy calculation, the calculation method is as follows: ,in, as a landmark and European distance, This represents the maximum European distance from the initial set of landmarks. (The remaining text appears to be incomplete and fragmented, possibly due to OCR errors.) and The spatial distribution uniformity is obtained by normalization and weighted summation. For landmarks For each time-series image, a single-period normalized index value is calculated. The mean of the time-series sequence of each index is taken, and the variance of the time-series variation is calculated. Finally, the quantitative score of landmark visual stability is calculated. ,in, These are the mean values of the time series of each indicator. This represents the combined time-series variance of the three indicators. The scoring results are limited to a range of 0 to 1; a value greater than 1 is assigned the value 1, and a value less than 0 is assigned the value 0.
[0046] Based on the visual stability quantitative score, landmarks are divided into Level 1 rigid stable landmarks, Level 2 sub-stable landmarks, and Level 3 unstable landmarks. Only Level 1 and Level 2 landmarks are retained in the benchmark landmark database. Among them, Level 1 rigid stable landmarks account for no less than 60% of the total number of landmarks and are evenly distributed in a ring around the target building.
[0047] In addition, the construction of the global benchmark landmark library also includes a dynamic update mechanism for the benchmark landmark library: after each detection cycle is completed, the coordinate offset of each landmark in the benchmark landmark library is calculated through the current sequence image, and landmarks whose coordinate offset exceeds the preset stability threshold are downgraded and removed; at the same time, new landmarks that meet the stability requirements are added to complete the dynamic update of the benchmark landmark library. The updated landmark library must maintain the requirement of the number and spatial distribution uniformity of first-level rigid stable landmarks.
[0048] Then, multi-view synchronous image acquisition is performed: a mobile visual acquisition platform is used to simultaneously acquire multi-view images of the entire structure of the target building and all landmarks in the reference landmark library, obtaining a sequence of images. The mobile visual acquisition platform includes, but is not limited to, drones. Further, multi-view synchronous image acquisition specifically includes: setting multiple key structural parts of the target building; during the acquisition process, the GNSS-IMU integrated navigation module on the platform acquires the acquisition pose information in real time, ensuring that the image heading overlap between the reference landmarks and the key structural parts of the target building is ≥85%, the lateral overlap is ≥70%, and each key structural part and the reference landmark are covered by a sequence of images from no less than three differentiated shooting perspectives.
[0049] Next, a global absolute coordinate system is constructed. A feature matching algorithm is used to accurately match and remove mismatches between the reference landmark feature points and target building feature points in the image sequence. Furthermore, the reference landmark SIFT feature point sets and target building SIFT feature point sets extracted from the image sequence are supplemented with reference landmark ORB feature point sets and target building ORB feature point sets. The SIFT descriptors and ORB descriptors are normalized and weighted to obtain fused feature descriptors, thus yielding the reference landmark fused feature descriptor subset and the target building fused feature descriptor subset.
[0050] The K-nearest neighbor feature matching algorithm is used to calculate the Hamming distance and Euclidean distance similarities between the baseline landmark fusion feature description subset and the target building fusion feature description subset. A weighted summation is then performed to obtain the comprehensive matching similarity. A preset similarity threshold is set, and feature point matching pairs with similarity values greater than or equal to the threshold are selected to form an initial matching pair set. The calculation process is as follows: Assuming the baseline landmark fusion feature description subset... A fusion descriptor with one reference landmark feature point (Corresponding to point A on the benchmark landmark), target building fusion feature description subset A fusion descriptor containing n target building feature points: At this point, first calculate and Hamming distance similarity Similarity to Euclidean distance The weighted fusion yields the overall similarity: ,calculate and Hamming distance similarity Similarity to Euclidean distance The overall similarity is obtained as follows: ; and so on. With all The overall similarity is used. The overall similarity scores are sorted, and the top two with the highest similarity are selected. If the highest similarity score is greater than or equal to a set similarity threshold, the two are considered an initial matching pair. All of them Repeat the above steps to obtain all the initial matching pairs.
[0051] The initial set of matched pairs is normalized by pixel coordinates. The coordinate normalization calculation method is as follows: , ,in, The mean of the pixel coordinates of all feature points in the initial matching pair set. The standard deviation of the pixel coordinates of all feature points in the initial matching pair set.
[0052] Iterative mismatch removal is performed on the standardized initial matching pair set. In each iteration, four non-repeating standardized matching pairs are randomly selected from the initial matching pair set as the minimum sample set. The transformation homography matrix between the reference landmark and the target building feature points is solved. All standardized matching pairs in the initial matching pair set are traversed, and the transformation error of each matching pair under the homography matrix is calculated. ,in, The target building feature points are predicted pixel coordinates after the homography matrix transformation of the baseline landmark feature points. The transformation error is compared with the inlier error threshold. If the transformation error is less than or equal to the threshold, it is determined as an inlier. The total number of inliers in this iteration is counted. If the total number of inliers is greater than the previous maximum number of inliers, the homography matrix and the maximum number of inliers are updated. After all iterations are completed, the initial matching pair set is retraced based on the optimal homography matrix. All outliers with transformation errors greater than the inlier error threshold are removed to obtain the feature point matching pair set after the mismatch removal.
[0053] Using the absolute geodetic coordinates of the benchmark landmark database as global constraints, a unified global absolute coordinate system is constructed. First, basic data preparation is completed, obtaining the set of matching pairs of fused feature points of the benchmark landmarks after feature matching and mismatch removal, the pixel coordinates of the matching pairs in each image sequence, and the pre-determined WGS84 absolute geodetic coordinates of each landmark in the benchmark landmark database, including planar coordinates and elevation data. Then, coordinate datum transformation is performed, converting the absolute geodetic coordinates of the benchmark landmarks into planar rectangular coordinates matching the target building area through Gaussian conformal projection. Combined with the corresponding elevation data, a unified three-dimensional rectangular coordinate system is formed for each landmark, serving as the rigid datum value for the global coordinates. Next, a solution model is constructed based on the bundle adjustment algorithm. Using the three-dimensional rectangular coordinates of the benchmark landmarks as fixed constraints, the interior and exterior orientation elements of each image sequence and the three-dimensional coordinates of the target building feature points are used as parameters to be solved. Based on the collinearity equation of the feature points, a reprojection error equation is constructed. Finally, the error equation is solved by iterative least squares method to minimize the global reprojection error of all feature points, complete the global optimization of the pose of all sequence images and the unified calculation of the three-dimensional coordinates of all feature points, and incorporate all feature points of the reference landmark and the target building into the same three-dimensional spatial coordinate system. This coordinate system is strictly bound to the absolute geodetic coordinates of the reference landmark library, thus completing the construction of the global absolute coordinate system.
[0054] Furthermore, relying on the global absolute coordinate system, a high-precision three-dimensional digital model of the target building to be detected with absolute three-dimensional coordinates is generated based on the sequence images, and the absolute three-dimensional coordinate dataset of the target building feature points is extracted.
[0055] Furthermore, the specific implementation of generating a high-precision 3D digital model of the target building with absolute 3D coordinates involves directly assigning the absolute 3D coordinates of all target building feature points and reference landmark feature points to the corresponding sparse point cloud, using the global absolute coordinate system as a reference, to generate an initial sparse point cloud model with absolute coordinates. During the generation process, the absolute 3D coordinates of the reference landmark and target building feature points are directly assigned to the corresponding 3D points in the sparse point cloud through a one-to-one correspondence between the feature points, ensuring that the coordinate system of the initial sparse point cloud is completely bound to the global absolute coordinate system, thus ensuring that the point cloud has a true geospatial reference. Subsequently, high-density point cloud reconstruction is performed on the target building area. Finally, the initial high-density point cloud is post-processed and optimized. A statistical filtering algorithm is used to remove outliers from the point cloud, and a bilateral filtering algorithm is used to smooth the remaining point cloud while preserving the edge structure features of the target building. The optimized high-density point cloud is then aligned with the initial sparse point cloud to ensure that the high-density point cloud fully inherits the coordinate reference of the global absolute coordinate system, ultimately generating a high-precision 3D digital model of the target building with absolute 3D coordinates.
[0056] Finally, multi-phase detection and full-dimensional deviation calculation are performed. The above steps are repeated according to the preset detection cycle to complete the sequential image acquisition and 3D digital model construction for different detection cycles. Stable landmarks in the benchmark landmark library are used as rigid registration anchor points to complete the coordinate system registration of the multi-cycle 3D digital model. Through the difference calculation between the multi-cycle 3D digital model and the absolute 3D coordinate dataset, the full-dimensional deformation deviation data of the target building is obtained. According to the preset deviation threshold of the relevant building deformation measurement specifications, the full-dimensional deformation deviation data is judged for compliance, a standardized detection report is output, and graded warnings are given for abnormal deformations exceeding the threshold. Furthermore, the specific implementation of obtaining the full-dimensional deformation deviation data of the target building through differential calculation of the multi-period 3D digital model and the absolute 3D coordinate dataset is as follows: First, the spatial coordinates of the feature points of the target building in the current period and the previous period are subtracted to obtain the deformation deviation of that part in the X / Y / Z axes in the current period, which is recorded as the relative deformation deviation. Then, the deformation deviation of the spatial coordinates of the feature points of the target building in the current period is compared with the spatial coordinates of the reference period, which is recorded as the absolute deformation deviation. At the same time, based on the relative deformation deviation and the absolute deformation deviation, the overall tilt change and settlement difference change are calculated. Specifically, the relative deformation deviation is calculated first. From the feature point set of the target building in each period's 3D digital model, a subset of key structural feature points is pre-defined. For the current detection period (referred to as period w) and the previous detection period (referred to as period w-1), the absolute 3D coordinates of each feature point in the key structural feature point subset in the two periods' models are extracted and recorded as (Xw, Yw, Zw) and (Xw-1, Yw-1, Zw-1), respectively. For each feature point, the differences in its three-axis coordinates are calculated to obtain the horizontal offset of the X-axis, the horizontal offset of the Y-axis, and the settlement of the Z-axis within the current period. The set of three-axis deviations for all feature points represents the relative deformation deviation for the current period. Next, the absolute deformation deviation is calculated. Using the absolute three-dimensional coordinates of the key structural feature points in the first detection period (denoted as the baseline period) as a reference, the coordinates of the corresponding feature points in the current period (period w) are extracted, and the differences in the three-axis coordinates are calculated, including the X-axis offset, the Y-axis offset, and the Z-axis settlement. The cumulative set of three-axis deviations for all feature points represents the absolute deformation deviation. Finally, the overall tilt change and settlement difference change are calculated. For the overall tilt, select the corresponding feature points at the top and bottom of the same vertical facade of the building, and calculate the horizontal offset vector of the top feature point relative to the bottom feature point in the current period and the reference period respectively. The overall tilt change of the facade is solved by the change of the angle between the two vectors. For the settlement difference, select the feature points on both sides of the adjacent column bases or adjacent settlement joints of the building, and calculate their cumulative settlement difference in the current period to obtain the local settlement difference. At the same time, select the feature points at the corners of the building's diagonal walls, and calculate their cumulative settlement difference to obtain the overall settlement difference.
[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A method for detecting building settlement offset deviation based on visual processing, characterized in that, Includes the following steps: S1. Construction of the benchmark landmark library: Determine the target building to be detected and its corresponding detection coverage area, screen natural landmarks and man-made structures outside the detection coverage area that have long-term structural stability, complete the landmark stability classification and screening through AI visual feature analysis, and construct a global benchmark landmark library by obtaining the absolute geodetic coordinates of each landmark. S2. Multi-view synchronous image acquisition: Using a mobile visual acquisition platform, multi-view synchronous image acquisition is performed on the entire structure of the target building and all landmarks in the benchmark landmark database to obtain a sequence of images. S3. Construction of Global Absolute Coordinate System: Through feature matching algorithm, the feature points of the reference landmarks and the feature points of the target buildings in the sequence image are accurately matched and mismatched are eliminated. The absolute geodetic coordinates of the reference landmark library are used as global constraints to construct a unified global absolute coordinate system. S4. Construction of 3D model of target building with absolute coordinates: Under the global absolute coordinate system, based on the sequence of images, a high-precision 3D digital model of the target building to be detected with absolute 3D coordinates is generated, and the absolute 3D coordinate dataset of the feature points of the target building is extracted. S5. Multi-period detection and full-dimensional deviation calculation: Repeat steps S2 to S4 according to the preset detection cycle to complete the sequential image acquisition and 3D digital model construction for different detection cycles; use stable landmarks in the benchmark landmark library as rigid registration anchor points to complete the coordinate system registration of the multi-period 3D digital model; calculate the full-dimensional deformation deviation data of the target building through the difference calculation between the multi-period 3D digital model and the absolute 3D coordinate dataset; make compliance judgment on the full-dimensional deformation deviation data according to the preset deviation threshold of the relevant building deformation measurement specifications, output a standardized detection report, and provide graded warnings for abnormal deformations exceeding the threshold.
2. The method for detecting building settlement offset deviation based on visual processing according to claim 1, wherein, The construction of the global benchmark library in step S1 specifically includes the following steps: S11. Delineate the geological impact boundary of the target building, screen natural landmarks and man-made landmarks outside the detection coverage area, located in different geological structural units from the target building, without direct groundwater hydrological connection, and with a horizontal distance of not less than twice the height of the target building, to form a preliminary set of landmarks; S12. For each landmark in the initial landmark set, extract three core indicators—temporal feature invariance, texture richness, and spatial distribution uniformity—through AI visual feature analysis, calculate the temporal variation variance of each landmark, and complete the quantitative scoring of landmark visual stability. S13. Based on the visual stability quantitative score, landmarks are divided into Level 1 rigid stable landmarks, Level 2 sub-stable landmarks, and Level 3 unstable landmarks. Only Level 1 and Level 2 landmarks are retained in the benchmark landmark database. Among them, Level 1 rigid stable landmarks account for no less than 60% of the total number of landmarks and are evenly distributed in a ring around the target building.
3. A visual processing based building settlement deviation detection method according to claim 2, wherein, Step S1 also includes a dynamic update mechanism for the benchmark landmark library: after each detection cycle is completed, the coordinate offset of each landmark in the benchmark landmark library is calculated using the current sequence image, and landmarks whose coordinate offset exceeds the preset stability threshold are downgraded and removed; at the same time, new landmarks that meet the stability requirements are added to complete the dynamic update of the benchmark landmark library. The updated landmark library must maintain the quantity and spatial distribution uniformity requirements of the first-level rigid stable landmarks.
4. The method for detecting building settlement offset deviation based on visual processing according to claim 2, characterized in that, The implementation of step S12 specifically includes the following steps: S121. Set up an initial set of landmarks and collect multiple historical visual samples for each landmark. Perform preprocessing on each historical visual sample of each landmark in sequence, and then use an AI feature detection algorithm to filter feature points and extract a standardized SIFT feature point set for each landmark. S122. Based on the standardized SIFT feature point set, the original value of the temporal feature invariance of the landmark is calculated by integrating the average feature point matching success rate and the average feature descriptor similarity. S123. Calculate the texture richness value of the landmark by fusing the entropy value of the gray-level co-occurrence matrix and the average gradient magnitude; S124. Calculate the spatial distribution uniformity of landmarks by combining relative position uniformity and global distribution uniformity; S125. Calculate the single-period normalized index value for each time series image of the landmark, take the mean of the time series of each index, calculate the time series variation variance, and finally calculate the landmark visual stability quantitative score.
5. The method for detecting building settlement offset deviation based on visual processing according to claim 1, wherein, Step S2, multi-view synchronous image acquisition, specifically includes: Multiple key structural parts of target buildings are set. During the acquisition process, the GNSS-IMU integrated navigation module on the platform is used to acquire the acquisition pose information in real time, ensuring that the image heading overlap between the reference landmark and the key structural parts of the target buildings is ≥85% and the lateral overlap is ≥70%, and each key structural part and the reference landmark are covered by a sequence of images with no less than 3 different shooting perspectives.
6. The method for detecting building settlement offset deviation based on visual processing according to claim 1, wherein, The specific implementation of step S3, which uses a feature matching algorithm to accurately match and remove mismatches between reference landmark feature points and target building feature points in the image sequence, includes: S31. Extract the SIFT feature point set of the reference landmark and the SIFT feature point set of the target building from the sequence image, and supplement the ORB feature point set of the reference landmark and the ORB feature point set of the target building. Normalize the SIFT descriptor and the ORB descriptor respectively, and weight them to obtain the fused feature descriptor, thereby obtaining the fused feature descriptor subset of the reference landmark and the fused feature descriptor subset of the target building. S32. Using the K-nearest neighbor feature matching algorithm, calculate the Hamming distance similarity and Euclidean distance similarity between the fused feature description subset of the benchmark landmark and the fused feature description subset of the target building. Obtain the comprehensive matching similarity by weighted summation, and preset a similarity threshold. Select feature point matching pairs that are greater than or equal to the similarity threshold to form an initial matching pair set. S33. Standardize the pixel coordinates of the initial matching pair set, and then perform iterative mismatch removal on the standardized initial matching pair set. In each iteration, randomly select four sets of non-repeating standardized matching pairs from the initial matching pair set as the minimum sample set, solve the transformation homography matrix between the reference landmark and the target building feature points, traverse all standardized matching pairs in the initial matching pair set, calculate the transformation error of each matching pair under the homography matrix, judge the magnitude of the transformation error and the interior point error threshold, if it is less than or equal to the threshold, it is determined as an interior point, count the total number of interior points in this iteration, if the total number of interior points is greater than the previous maximum total number of interior points, update the homography matrix and the maximum total number of interior points, after completing all iterations, use the optimal homography matrix as the benchmark, traverse the initial matching pair set again, remove all exterior points whose transformation error is greater than the interior point error threshold, and obtain the feature point matching pair set after mismatch removal.
7. The method for detecting building settlement offset deviation based on visual processing according to claim 1, wherein, The specific implementation of generating a high-precision 3D digital model of the target building with absolute 3D coordinates based on the sequence images in step S4 is as follows: Using the global absolute coordinate system as a reference, the absolute three-dimensional coordinates of all target building feature points and benchmark landmark feature points are directly assigned to the corresponding sparse point cloud to generate an initial sparse point cloud model with absolute coordinates; for the feature points of the target building, high-density point cloud reconstruction is performed to obtain a high-precision three-dimensional digital model.
8. The method for detecting building settlement offset deviation based on visual processing according to claim 1, characterized in that, The specific implementation of obtaining the full-dimensional deformation deviation data of the target building in step S5 by differential calculation between the multi-period three-dimensional digital model and the absolute three-dimensional coordinate dataset is as follows: S51. First, calculate the difference between the spatial coordinates of the feature points of the target building in the current cycle and the previous cycle to obtain the deformation deviation of the part in the X / Y / Z axes in the current cycle, which is recorded as the relative deformation deviation. S52. Next, compare the spatial coordinates of the feature points of the target building in the current period with the spatial coordinates of the reference period, and record the deformation deviation as the absolute deformation deviation. S53. Simultaneously, based on the relative deformation deviation and the absolute deformation deviation, the overall tilt change and settlement difference change are calculated.