Building engineering construction period deformation early warning method and system based on unmanned aerial vehicle oblique photography

Abstract

The application provides a construction period deformation early warning method and system based on unmanned aerial vehicle oblique photography, relates to the technical field of engineering surveying and building construction safety monitoring, and the method comprises the following steps: acquiring a multi-view image sequence of a foundation pit support structure collected by an unmanned aerial vehicle-mounted oblique photography camera along a preset flight route and airborne positioning and pose data in multiple monitoring periods during the construction period; performing three-dimensional point cloud reconstruction on the multi-view image sequence and the airborne positioning and pose data in the same monitoring period to obtain dense three-dimensional point clouds of the surface of the foundation pit support structure in each monitoring period. The application realizes accurate deformation early warning and spatial location classification early warning under construction dynamic interference by extracting stable structure point clusters in the point clouds of adjacent monitoring periods as self-reference.

Description

Technical Field

[0001] This invention relates to the field of engineering surveying and construction safety monitoring technology, and in particular to a method and system for early warning of deformation during the construction period of building engineering based on UAV oblique photography. Background Technology

[0002] With its advantages of non-contact acquisition and high-efficiency reconstruction, UAV oblique photography technology has been widely used in the field of deformation monitoring during the construction phase of building engineering. In scenarios where deep foundation pit excavation and support structure construction are carried out simultaneously, UAVs collect multi-angle overlapping images of retaining walls, capping beams, and anchor cable systems along a preset flight path at fixed intervals. Through 3D reconstruction, multi-temporal real-world image clouds are generated to track the spatial deformation evolution of the support system under the coupling effect of earth pressure release and construction load.

[0003] Existing monitoring methods mainly rely on global registration of multi-temporal point clouds and displacement difference calculation in absolute coordinate systems. In the monitoring of deep foundation pit construction of high-rise buildings, tilted images are usually acquired at fixed intervals. After reconstructing the point cloud, adjacent data are aligned through global rigid transformation, and the three-dimensional coordinate difference of each point is directly calculated to determine whether the support structure exceeds the limit. However, this existing technology has exposed technical defects in the actual dynamic construction environment: the multi-temporal point cloud calculation process lacks dynamic interference removal and self-baseline stabilization mechanisms, resulting in severe distortion of the deformation vector field and delayed early warning response. During the construction period, dynamic non-structural elements such as scaffolding erection and dismantling, temporary material loading, and heavy machinery movement will frequently change with the construction period, directly introducing a large amount of pseudo-differential noise when matching point clouds of adjacent periods. At the same time, traditional global registration relies heavily on absolute spatial coordinates or manually set static control points, which are easily rendered ineffective in complex construction sites due to the obstruction, damage, or deterioration of the line of sight of the reference points. This makes it impossible for the system to effectively separate dynamic construction interference from actual structural deformation. The calculated displacement field exhibits high-frequency oscillations and local jumps, and deformation trend identification relies heavily on manual experience filtering. When the support structure actually undergoes gradual lateral displacement, traditional methods fail to trigger accurate early warnings due to noise masking and inaccurate benchmarks, or frequently generate false alarms and missed alarms, failing to meet the high-efficiency and high-reliability safety management requirements during the construction period. Summary of the Invention

[0004] This invention provides a method and system for early warning of deformation during the construction period of building engineering based on UAV oblique photography. By extracting stable structural point clusters in the point cloud of adjacent monitoring periods as self-reference, and calculating the distance change along the surface normal direction to construct a three-dimensional deformation vector field, it can realize accurate deformation early warning and spatial location-based hierarchical early warning under construction dynamic interference.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for early warning of deformation during the construction period of building engineering based on UAV oblique photography, the method comprising: Acquire multi-view image sequences and airborne positioning and attitude data of the foundation pit support structure collected by UAVs equipped with oblique photography cameras along a preset flight path during multiple monitoring periods of the construction period; reconstruct three-dimensional point clouds from the multi-view image sequences and airborne positioning and attitude data of the same monitoring period to obtain dense three-dimensional point clouds of the foundation pit support structure surface in each monitoring period. Based on the dense 3D point clouds of each monitoring period, extract the point clouds of the previous period and the point clouds of the next period of adjacent monitoring periods. Using the coordinate system of the previous period point cloud as the reference, calculate the rigid transformation matrix between the previous period point cloud and the next period point cloud. Perform coordinate mapping on the next period point cloud through the rigid transformation matrix to obtain the registered point cloud pair. Spatial neighborhood matching is performed on the registered point cloud pairs to obtain a set of neighborhood point pairs; the geometric difference value of each set of neighborhood point pairs is calculated based on the set of neighborhood point pairs to obtain a set of geometric difference values; based on the set of geometric difference values, the corresponding neighborhood point pairs with difference values ​​lower than the preset stability threshold are selected to obtain a set of stable point clusters of the support structure. Based on the spatial position of each point cluster in the set of stable point clusters of the support structure on the surface of the foundation pit support structure, the distance change of the registered point cloud pair along the surface normal direction at that position is calculated to obtain the three-dimensional deformation vector field. The change in distance at each spatial location in the three-dimensional deformation vector field is taken as the deformation displacement value at that spatial location, and compared with a preset warning threshold to generate the deformation warning level at that spatial location, thus completing the deformation warning during the construction period.

[0006] Secondly, a construction deformation early warning system based on UAV oblique photography includes: The acquisition module is used to acquire multi-view image sequences and airborne positioning and attitude data of the foundation pit support structure collected by UAVs equipped with oblique photography cameras along preset routes during multiple monitoring periods in the construction period; and to reconstruct three-dimensional point clouds from the multi-view image sequences and airborne positioning and attitude data of the same monitoring period to obtain dense three-dimensional point clouds of the foundation pit support structure surface in each monitoring period. The processing module is used to extract the point cloud of the previous period and the point cloud of the next period from the dense 3D point cloud of each monitoring period. Using the coordinate system of the previous period point cloud as the reference, it calculates the rigid transformation matrix between the previous period point cloud and the next period point cloud. The rigid transformation matrix is ​​used to perform coordinate mapping on the next period point cloud to obtain the registered point cloud pair. The matching module is used to perform spatial neighborhood matching on the registered point cloud pairs to obtain a set of neighborhood point pairs; calculate the geometric difference value of each set of neighborhood point pairs based on the set of neighborhood point pairs to obtain a set of geometric difference values; and filter the corresponding neighborhood point pairs whose difference values ​​are lower than the preset stability threshold based on the set of geometric difference values ​​to obtain a set of stable point clusters of the support structure. The calculation module is used to calculate the distance change of the registered point cloud pair along the surface normal direction of the surface at the location of each point cluster in the set of stable point clusters of the support structure, based on the spatial position of each point cluster on the surface of the foundation pit support structure, and to obtain the three-dimensional deformation vector field. The evaluation module is used to take the distance change of each spatial location in the three-dimensional deformation vector field as the deformation displacement value of that spatial location, compare it with the preset warning threshold, generate the deformation warning level of that spatial location, and complete the deformation warning during the construction period.

[0007] Thirdly, a computing device includes: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0008] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0009] The above-described solution of the present invention has at least the following beneficial effects: This invention employs rigid registration of point clouds from adjacent monitoring periods, extracts stable point clusters of the support structure as self-references through spatial neighborhood matching and geometric difference value screening, and constructs a three-dimensional deformation vector field by calculating the distance change along the surface normal direction based on the spatial position of the stable point clusters. This field is then compared with the warning threshold to generate a spatially positioned deformation warning level. This overcomes the technical problems of traditional methods, such as the difficulty in separating dynamic construction interference from actual structural deformation, reliance on absolute coordinates or static control points leading to deformation field distortion and warning lag. It achieves the technical effects of effective suppression of dynamic noise, spatial continuity and reliability of the deformation vector field, accurate positioning of warning levels, and improved timeliness of safety response during construction. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating a method for early warning of deformation during the construction period of a building project based on oblique photography by an unmanned aerial vehicle (UAV), as provided in an embodiment of the present invention.

[0011] Figure 2 This is a schematic diagram of a construction deformation early warning system based on UAV oblique photography, provided by an embodiment of the present invention. Detailed Implementation

[0012] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0013] like Figure 1 As shown in the figure, an embodiment of the present invention proposes a method for early warning of deformation during the construction period of building engineering based on UAV oblique photography. The method includes the following steps: Acquire multi-view image sequences and airborne positioning and attitude data of the foundation pit support structure collected by UAVs equipped with oblique photography cameras along a preset flight path during multiple monitoring periods of the construction period; reconstruct three-dimensional point clouds from the multi-view image sequences and airborne positioning and attitude data of the same monitoring period to obtain dense three-dimensional point clouds of the foundation pit support structure surface in each monitoring period. Based on the dense 3D point clouds of each monitoring period, extract the point clouds of the previous period and the point clouds of the next period of adjacent monitoring periods. Using the coordinate system of the previous period point cloud as the reference, calculate the rigid transformation matrix between the previous period point cloud and the next period point cloud. Perform coordinate mapping on the next period point cloud through the rigid transformation matrix to obtain the registered point cloud pair. Spatial neighborhood matching is performed on the registered point cloud pairs to obtain a set of neighborhood point pairs; the geometric difference value of each set of neighborhood point pairs is calculated based on the set of neighborhood point pairs to obtain a set of geometric difference values; the corresponding neighborhood point pairs with difference values ​​lower than the preset stability threshold are selected based on the set of geometric difference values ​​to obtain a set of stable point clusters of the support structure. Based on the spatial position of each point cluster in the set of stable point clusters of the support structure on the surface of the foundation pit support structure, the distance change of the registered point cloud pair along the surface normal direction at that position is calculated to obtain the three-dimensional deformation vector field. The change in distance at each spatial location in the three-dimensional deformation vector field is taken as the deformation displacement value at that spatial location, and compared with a preset warning threshold to generate the deformation warning level at that spatial location, thus completing the deformation warning during the construction period.

[0014] In this embodiment of the invention, after rigid registration of point clouds in adjacent monitoring periods, stable point clusters of the support structure are extracted as self-references through spatial neighborhood matching and geometric difference value screening. Based on the spatial position of the stable point clusters, the distance change along the surface normal direction is calculated to construct a three-dimensional deformation vector field. Then, it is compared with the warning threshold to generate a spatially positioned deformation warning level. This overcomes the technical problems of traditional methods, such as the difficulty in separating dynamic construction interference from actual structural deformation, reliance on absolute coordinates or static control points leading to deformation field distortion and warning lag. It achieves the technical effects of effective suppression of dynamic noise, spatial continuity and reliability of the deformation vector field, accurate positioning of warning levels, and improved timeliness of safety response during construction.

[0015] In a preferred embodiment of the present invention, a multi-view image sequence of the foundation pit support structure and airborne positioning and attitude data are acquired by an UAV equipped with an oblique photography camera along a preset flight path during multiple monitoring periods of the construction period; the multi-view image sequence and airborne positioning and attitude data of the same monitoring period are reconstructed into a three-dimensional point cloud to obtain a dense three-dimensional point cloud of the surface of the foundation pit support structure for each monitoring period, including: The project aims to acquire multi-view image sequences of the foundation pit support structure and corresponding airborne positioning and attitude data collected by a drone equipped with an oblique photography camera along a preset flight path during multiple monitoring periods within the construction phase. Specifically, this includes: determining the monitoring frequency based on the scale, shape, and construction progress of the foundation pit support structure during the construction phase, according to the excavation depth and support structure type; setting one monitoring period per day during the excavation phase and one monitoring period every three days after the foundation pit support is completed; and adding an additional emergency monitoring period after severe rainstorms or strong winds. Before monitoring, the drone and its onboard equipment undergo comprehensive testing. A multi-rotor drone equipped with a five-lens oblique photography camera is selected, with the camera lens focal length set to 24mm and the image resolution adjusted to 5472×3648 pixels, including cracks in the retaining wall, displacement traces in the capping beam, and connection points of the anchor cable system. Simultaneously, the airborne positioning and attitude system is tested. This system integrates a GNSS positioning module and an IMU inertial measurement module. The GNSS positioning accuracy is calibrated to the centimeter level, and the IMU inertial measurement module is used to collect camera attitude data, with a sampling frequency set to 10Hz, to determine the accuracy of the positioning and attitude data acquisition. A pre-planned circular flight path is used, centered on the foundation pit support structure. The flight altitude is determined based on the side length of the foundation pit, generally controlled between 50-80 meters. The flight path ensures complete coverage of the entire foundation pit support structure area, without any omissions or obstructions. The lateral overlap of the flight paths is set at 70%, and the forward overlap is set at 80%. The spacing between adjacent flight paths is calculated based on the flight altitude and camera focal length. The calculation process is as follows: Flight path spacing Where D is the distance between adjacent flight paths; H is the UAV's flight altitude; and θ is the camera's field of view. To ensure lateral overlap, the UAV was controlled to fly at a constant speed along a preset route during each monitoring period, with a flight speed maintained at 5 m / s. During flight, the oblique photography camera simultaneously acquired images from the orthogonal view and four oblique viewpoints, acquiring one image every 2 seconds to form a multi-view image sequence of the foundation pit support structure for that monitoring period. The images were saved in JPG format, and the acquisition timestamp of each image was recorded for subsequent association with airborne positioning and attitude data. The airborne positioning and attitude system was activated synchronously with the camera acquisition, acquiring 3D positioning and attitude data in real time at the exposure time of each image, i.e., airborne positioning and attitude data. The 3D positioning data consisted of the engineering coordinates (X, Y, Z) at the camera exposure time, where X was the east coordinate, Y was the north coordinate, and Z was the celestial coordinate. The attitude data included heading angle, roll angle, and pitch angle, used to describe the camera's horizontal rotation angle, longitudinal tilt angle, and lateral tilt angle, respectively. The accuracy of the attitude data was controlled within 0.1°. During the data collection process, a dedicated person was assigned to monitor the drone's flight status in real time, avoiding scaffolding, heavy machinery, and temporary material storage areas within the construction site. If any obstructions were encountered, the drone's flight altitude or flight path was adjusted promptly. After each monitoring period was completed, the multi-view image sequence and airborne positioning and attitude data were exported synchronously and stored using a unified engineering coordinate system. The above data collection, verification, and storage operations were repeated to sequentially obtain the multi-view image sequence and airborne positioning and attitude data corresponding to all monitoring periods during the construction period.

[0016] A multi-view image sequence and airborne positioning and attitude determination data corresponding to a monitoring period are selected. Feature points are extracted from the multi-view images within this monitoring period, and corresponding feature points are matched between adjacent images to obtain a set of feature point matching pairs for that monitoring period. Specifically, this includes: selecting any monitoring period as the current processing period, retrieving the multi-view image sequence and airborne positioning and attitude determination data corresponding to that monitoring period from all stored data, preprocessing the multi-view image sequence to eliminate image noise and distortion, and using a Gaussian filtering algorithm to reduce noise in each image during preprocessing while preserving the surface details of the foundation pit support structure to the greatest extent. In this preprocessing, the Gaussian filter kernel size is set to 3×3, and the filtering formula is: ; in, The grayscale values ​​of the pixels in the filtered image. , This represents the coordinates of the current pixel relative to the center of the filter kernel, with values ​​of -1, 0, and 1, corresponding to the 9 pixel positions of the 3×3 filter kernel. The standard deviation is Gaussian, set to 1.0 in this case. The smaller the value, the gentler the filtering effect. The weighted average gray value of each pixel is calculated using this formula to complete single-pixel noise reduction. After noise reduction, camera intrinsic calibration data is used to correct image distortion, eliminating image deformation caused by lens distortion. After preprocessing, scale-invariant feature point extraction is performed on each image in the multi-view image sequence. The Gaussian difference pyramid algorithm is used to detect feature points in the image. This algorithm constructs Gaussian difference images at different scales and detects local extrema as feature points. The specific operation process is as follows: constructing a Gaussian pyramid... The images are scaled at different scales with a scaling factor of 0.8 and a pyramid of 6 layers, with each layer being 0.8 times the size of the previous layer. Difference operations are performed between Gaussian images at adjacent scales to obtain Difference-of-Gaussian (DOB) images. Local extrema are detected in the DOB images, i.e., points where the value is the maximum or minimum at their own location, the corresponding location at adjacent scales, and the location of adjacent pixels. These local extrema are used as initial feature points. After extracting the initial feature points, they are filtered to remove pseudo-feature points with abnormal edge response values. The edge response value calculation process is as follows: ,in This is the edge response value. The determinant of the Hessian matrix at the location of the feature point. The trace of the Hessian matrix, The scaling factor is set to 0.04, and the edge response value threshold is set to 0.01. Feature points with edge response values ​​greater than the threshold are retained, while pseudo-feature points with edge response values ​​less than or equal to the threshold, such as the edges of retaining walls, the corners of capping beams, and the fixing points of anchor cables, are removed. After feature point extraction, corresponding feature points are matched between adjacent images in the multi-view image sequence. Corresponding feature points are the projections of the same spatial point onto different images. The matching uses a fast approximate nearest neighbor matching algorithm. The specific matching process is as follows: calculate the SIFT descriptor of all feature points in each image. The SIFT descriptor is a 128-dimensional vector used to describe the local texture features of the feature points. Use the feature point descriptor of the first image in the adjacent images as the query set and the feature point descriptor of the second image as the training set. Calculate the Euclidean distance between each feature point in the query set and all feature points in the training set using the fast approximate nearest neighbor algorithm. The Euclidean distance calculation formula is: ,in The Euclidean distance is the descriptor of two feature points in adjacent images. , Let be any two components of the feature point descriptor in the first image. , For the corresponding components of the feature point descriptors in the second image, a Euclidean distance matching threshold is set. Based on the image resolution and feature point extraction accuracy, the matching threshold is set to 200. For each feature point in the query set, the feature point with the smallest Euclidean distance in the training set that is less than the matching threshold is retained as a matching candidate point. If the ratio of the smallest Euclidean distance to the second smallest Euclidean distance is greater than 0.8, it is judged as a mismatch and is removed. The calculation process for this ratio is as follows: ,in The ratio, This is the minimum Euclidean distance. The second smallest Euclidean distance is used to further filter out mismatched feature point pairs. All valid matching feature point pairs that pass the filter are integrated to obtain the feature point matching pair set for the current monitoring period. Each feature point matching pair contains the pixel coordinates and descriptor information of two corresponding feature points in adjacent images.

[0017] Based on the feature point matching pair set and airborne positioning and attitude data for the monitoring period, the camera spatial position and attitude at each image exposure time are calculated to obtain the camera pose parameter set for the monitoring period. Specifically, this includes: calling the feature point matching pair set and airborne positioning and attitude data for the current monitoring period, combining the photogrammetric collinearity equation to calculate the camera spatial position and attitude at each image exposure time, constructing the camera pose parameter set, and determining the specific form of the photogrammetric collinearity equation, which is: ; ; in , These are the pixel coordinates of the feature points in the image. For camera focal length, , , , , , , , , For the elements of the rotation matrix, The three-dimensional spatial coordinates of the feature points Given the camera's 3D spatial coordinates, first calculate the reprojection error of the feature points. The reprojection error is the deviation between the actual pixel coordinates of the feature points and the pixel coordinates calculated using the collinearity equation. The formula for calculating this error is: ,in This represents the reprojection error of a single set of feature points. , These are the actual pixel coordinates of the feature points on the image. , The pixel coordinates of feature points are calculated by substituting the camera pose parameters into the collinearity equation. Combined with the reprojection error, a bundle adjustment optimization algorithm is used to solve the camera pose parameters. Its core principle is to construct an error equation and minimize the sum of reprojection errors for all feature points, thereby achieving the optimal solution for the camera's spatial position and rotation matrix. The specific implementation process is as follows: The reprojection error threshold is set to 1.5 pixels. First, the 3D positioning data from the airborne positioning and attitude determination data is used as the camera's spatial position... The initial value is taken from the attitude data in the airborne positioning and attitude determination data as the rotation matrix. R The initial values ​​are then substituted into the photogrammetric collinearity equation to calculate the reprojection error for each set of feature point matching pairs. Construct the objective function for the sum of reprojection errors. The objective function is: ; in, This is the sum of the reprojection errors for all feature points. This represents the total number of feature point matching pairs. For the first Reprojection error of a group of feature points , For the first The actual pixel coordinates of the group of feature points , For the first The pixel coordinates of the group of feature points are calculated using the initial camera pose parameters. The core of bundle adjustment optimization is to iteratively adjust the spatial position of the camera. With elements of the rotation matrix Gradually reduce the objective function The value of the iterative optimization is to make the sum of the reprojection errors equal to the sum of the values ​​... To minimize this error, the number of iterations was set to 50 in this optimization process. After each iteration, the reprojection error and the total error of all feature points were recalculated. If the difference between the total reprojection error of two adjacent iterations was less than 0.01, then... ,in, For the first The sum of errors from each iteration. For the first The iteration stops when the sum of errors from each iteration is reached. The resulting camera spatial position and rotation matrix is ​​the optimal solution, which minimizes reprojection errors. Based on the optimal rotation matrix, the camera's attitude parameters, namely yaw angle, roll angle, and pitch angle, are calculated. The specific calculation process is as follows: Yaw angle... Roll angle Pitch angle ,in The heading angle, measured in degrees, describes the angle of rotation of the camera around the yaw axis. The roll angle, measured in degrees, is used to describe the angle of tilt of the camera about the east axis. The pitch angle, measured in degrees, describes the camera's tilt angle around the north axis. This calculation process is performed on each image in the multi-view image sequence to obtain the camera's three-dimensional spatial coordinates at each image's exposure time. With attitude parameters The camera spatial position and attitude parameters of all images are integrated in the order of image acquisition timestamps to obtain the set of camera pose parameters for the current monitoring period.

[0018] Based on the camera pose parameter set and feature point matching pairs for the monitoring period, sparse 3D reconstruction is performed to generate a sparse 3D point cloud of the foundation pit support structure surface for that monitoring period. Specifically, this includes: filtering the feature point matching pair set, removing feature point matching pairs with a reprojection error greater than 1.5 pixels, and retaining valid feature point matching pairs with a reprojection error less than or equal to a threshold; determining the camera shooting ray direction for each image based on the camera pose parameter set, where the shooting ray direction is determined by the camera's spatial position. With feature point pixel coordinates It is determined that the direction vector of the shooting light is obtained by inversely deriving the collinearity equation. The calculation process of the direction vector is as follows: Vector ,in To capture the direction vector of the light, For camera focal length, , , The third row element of the rotation matrix is ​​used to obtain the pixel coordinates of the feature point in the two adjacent images for each valid feature point matching pair. Combined with the camera pose parameters of the two images, two corresponding shooting rays are determined. These two shooting rays are then spatially intersected; the intersection point is the three-dimensional spatial coordinate of the feature point in the engineering coordinate system. The spatial coordinate calculation process is as follows: Let the camera coordinates of the first image be... The direction vector of the shooting light is The camera coordinates of the second image are: The direction vector of the shooting light is Then the three-dimensional spatial coordinates of the feature points Simultaneously satisfy: ; ;in , Given the ray parameters, by solving this system of linear equations, we obtain... , Substituting the values ​​into the equations will give you the three-dimensional spatial coordinates of the feature points. The above spatial intersection calculation is performed on all valid feature point matching pairs to obtain several three-dimensional spatial coordinate points. These coordinate points are filtered to remove pseudo-points that are too far from the preset range of the foundation pit support structure. The preset range is determined according to the actual size of the foundation pit, generally extending 5 meters outward from the edge of the foundation pit. All the filtered three-dimensional spatial coordinate points are organized according to their corresponding spatial positions and combined to form a sparse three-dimensional point cloud of the foundation pit support structure surface during the current monitoring period. The point density of this point cloud is about 10 points / square meter, which accurately marks the key feature positions of the support structure surface, such as the edge line of the retaining wall, the outline of the capping beam, and the fixing points of the anchor cables.

[0019] Based on the sparse 3D point cloud and camera pose parameter set for the monitoring period, the depth map corresponding to each image is calculated pixel by pixel, and multi-view depth map fusion is performed to obtain the dense 3D point cloud for the monitoring period. Specifically, this includes: calling the sparse 3D point cloud and camera pose parameter set for the current monitoring period, performing pixel-by-pixel depth calculation on each image in the multi-view image sequence, and generating a depth map for a single image. The depth calculation adopts a semi-global matching algorithm guided by sparse point cloud. The specific calculation process is as follows: guided by the sparse 3D point cloud, the approximate depth range corresponding to each pixel in the image is determined. The depth range calculation process is as follows: , ,in The minimum depth value, This is the maximum depth value. For the camera's flight altitude, This represents the maximum height of the foundation pit support structure. To determine the minimum height of the foundation pit support structure and ensure the depth calculation is within a reasonable range, the image is divided into blocks, with each image consisting of several 16×16 pixel blocks. A cost calculation is performed on each pixel block, based on the difference in pixel grayscale values. The census transform is used to calculate the grayscale difference between adjacent pixels. The cost function is as follows: ; in, pixels in the image In depth Matching cost at the location; These are the pixel coordinates in the current image where the depth to be calculated. The depth value is the current assumption, and its range is within... and between; This represents the grayscale value at the corresponding 16×16 position within the pixel block in the current image. , It is the pixel index within the pixel block, with a value range of 0-15, covering the entire 16×16 pixel block; The cost function quantifies the degree of matching between the current pixel and its corresponding pixel in a neighboring image at a certain depth, using the grayscale value within the corresponding pixel block in the adjacent image. After cost calculation, since the cost calculation for a single pixel block is easily affected by local noise and texture loss, a dynamic programming algorithm is used to aggregate the cost of each pixel in multiple directions to compensate for insufficient local matching. Eight aggregation directions are set: horizontal, vertical, 45°, 135°, and their opposite directions. After aggregation, the depth value with the lowest aggregation cost is selected as the final depth value of the pixel using the Winner-Takes-All strategy. ; in, For image pixels To further improve the accuracy of the final depth value, it needs to be verified using the camera's flight altitude, focal length, and image point height. The verification formula is as follows: ,in, This represents the depth value corresponding to the pixel. The camera's flight altitude; The focal length of the camera; The depth value is the pixel height in the image. This formula is used for verification, and abnormal pixels with excessively large deviations from the verification value are removed. After calculating the depth maps for all images, due to differences in shooting angles and lighting conditions in multi-view images, there will be deviations between the depth maps generated from different images. Furthermore, there may be a small number of noise points with abnormal depth values ​​in the depth maps. Therefore, multi-view depth maps need to be fused. The Poisson fusion algorithm is used to complete this operation. The specific fusion process is as follows: All depth maps are registered; the depth maps of different images are transformed to the same engineering coordinate system based on the camera pose parameters; and multiple depth values ​​at the same spatial location are weighted and averaged. The weight calculation process is as follows: ,in, For the first The weight of each depth value For the first The variance of each depth value To find the minimum value, the weighted average formula is: ; in, The final depth value after merging from the same spatial location. For the spatial location corresponding to A depth value, For the corresponding weights, a weighted average is used to effectively balance the reliability of different depth values, eliminate bias, and remove noise. The fused depth data needs to be converted into three-dimensional spatial coordinate points to generate a dense three-dimensional point cloud. The conversion process is based on the camera imaging principle, and the conversion formula is as follows: ; ; ; in, These are the three-dimensional spatial coordinates corresponding to the depth value; The camera spatial coordinates at the current image exposure moment; The pixel coordinates of the current pixel in the image; The coordinates of the camera's principal point; The focal length of the camera; This represents the depth value after fusion. All transformed 3D spatial coordinate points are integrated to form a dense 3D point cloud of the foundation pit support structure surface during the current monitoring period.

[0020] For each monitoring period, feature point extraction and matching, camera pose calculation, sparse 3D reconstruction, depth map calculation, and multi-view fusion are performed sequentially to obtain dense 3D point clouds for each monitoring period. Specifically, this includes: sorting each monitoring period during the construction period in ascending order according to the timestamp of image acquisition, with the timestamp accurate to the second; using the dense 3D point cloud generation process of the current processing period as a standard, developing a standardized processing manual; and sequentially processing the multi-view image sequence and airborne positioning and pose determination data corresponding to the next monitoring period. The processing strictly follows the operating procedure of the current period. The multi-view image sequence was preprocessed, and Gaussian filtering was used to eliminate image noise with a kernel size of 3×3 and σ=1.0. Distortion correction was performed in conjunction with camera intrinsic calibration data. Subsequently, the difference-of-gaussian pyramid algorithm was used to extract feature points with 6 pyramid layers and a scaling factor of 0.8. False feature points were removed by using an edge response value threshold of 0.01. The FLANN matching algorithm was used to match corresponding feature points with a matching threshold of 200 and a false match rejection ratio of 0.8, resulting in a set of feature point matching pairs for the monitoring period. The camera pose parameters for the monitoring period were then calculated using a bundle adjustment optimization algorithm. The projection error threshold is 1.5 pixels, the number of iterations is 50, and the convergence condition is that the sum of the reprojection errors of two adjacent iterations is less than 0.01. The camera spatial position and pose parameters of each image are calculated and integrated into a set of camera pose parameters. Based on this set and the set of feature point matching pairs, sparse 3D reconstruction is performed using the triangulation principle. Matching pairs with reprojection errors greater than 1.5 pixels are removed, and false points outside the preset range of the pit are filtered out to generate a sparse 3D point cloud for the monitoring period. Then, based on the sparse point cloud and the camera pose parameters, the SGM algorithm is used to calculate the depth map pixel by pixel. The depth range is determined according to the current... The flight altitude and pit height were recalculated, and multi-view depth map fusion was performed using the Poisson fusion algorithm. Weights were calculated based on the variance of depth values. The fused depth data was then converted into 3D spatial coordinate points, generating a dense 3D point cloud for that monitoring period. After processing each monitoring period, the generated dense 3D point cloud underwent quality verification. Verification indicators included point density, point cloud accuracy, and completeness. Point density needed to be at least 1000 points / square meter; point cloud accuracy needed to be ±1cm for planar accuracy and ±2cm for elevation accuracy; and completeness needed to cover the entire area of ​​the pit support structure without omissions. If the verification failed, the data for that monitoring period was reprocessed, and relevant parameters were adjusted until a qualified dense 3D point cloud was generated. This processing and verification process was repeated until all monitoring periods during the construction period were traversed, ultimately obtaining dense 3D point clouds of the pit support structure surface for all monitoring periods. All dense 3D point clouds for all periods were stored using a unified engineering coordinate system and accuracy standard. The acquisition time, flight parameters, and processing parameters for each monitoring period were also recorded, forming a full-time dense 3D point cloud dataset.

[0021] In this embodiment of the invention, by employing a pre-set circular route and high-overlap image acquisition, Gaussian filtering for noise reduction and distortion correction preprocessing, Gaussian difference pyramid feature point extraction and FLANN matching, bundle adjustment to solve camera pose, triangulation sparse reconstruction, and semi-global matching and Poisson fusion to generate dense point clouds, and by traversing all monitoring periods to generate a full-time dataset, the technical problems of high image noise, low feature point matching accuracy, reliance on absolute control points for camera pose calculation, and insufficient density of sparse point clouds leading to loss of deformation details are overcome. This achieves the technical effects of accurate registration of multi-view images and pose data, high-density and high-precision reconstruction of dense point clouds, and a reusable full-time point cloud coordinate system.

[0022] In a preferred embodiment of the present invention, point clouds from previous and subsequent monitoring periods are extracted from dense 3D point clouds of each monitoring period. Using the coordinate system of the previous point cloud as a reference, a rigid transformation matrix between the previous and subsequent point clouds is calculated. Coordinate mapping is then performed on the subsequent point cloud using the rigid transformation matrix to obtain a registered point cloud pair, including: Based on the dense 3D point clouds of each monitoring period, two adjacent monitoring periods are selected in chronological order of acquisition time. The dense 3D point cloud corresponding to the monitoring period acquired earlier is designated as the previous period point cloud, and the dense 3D point cloud corresponding to the monitoring period acquired later is designated as the next period point cloud, thus obtaining a pair of point clouds for adjacent monitoring periods. Specifically, based on the dense 3D point cloud dataset of the entire time period generated above, the dense 3D point clouds corresponding to all monitoring periods and their respective acquisition timestamps are called. All monitoring periods are sorted in ascending order according to the timestamp of image acquisition. After sorting, two adjacent monitoring periods are selected in sequence to determine and distinguish the previous period point cloud from the next period point cloud: the dense 3D point cloud corresponding to the monitoring period acquired earlier is defined as the previous period point cloud, and the dense 3D point cloud corresponding to the monitoring period acquired later is defined as the next period point cloud, thus forming a pair of point clouds for adjacent monitoring periods. During the selection process, the quality parameters of the two sets of point clouds need to be verified simultaneously: point density ≥ 1000 points / square meter, planar accuracy ±1cm, elevation accuracy ±2cm, and complete coverage of the foundation pit support structure. If the point cloud fails to meet the standards for any monitoring period, the original data for that period, multi-view image sequences, and airborne positioning and attitude data need to be retrieved again. A qualified dense 3D point cloud needs to be regenerated according to the standardized processing procedure. Then, the selection of adjacent point cloud pairs is carried out, and the above selection and verification process is repeated to obtain the point cloud pairs corresponding to all adjacent monitoring periods during the construction period.

[0023] Using the 3D spatial coordinate system of the previous point cloud as a reference coordinate system, the corresponding point pairs with the smallest spatial distance are searched within the overlapping area of ​​the previous and subsequent point clouds to obtain a set of point pairs with the same name. Specifically, after selecting a set of point cloud pairs for adjacent monitoring periods, the reference coordinate system is determined: the 3D spatial coordinate system of the previous point cloud, i.e., the previously uniformly set engineering coordinate system, with X representing east, Y representing north, and Z representing the sky, is used as the reference coordinate system to determine the overlapping area of ​​the previous and subsequent point clouds. The overlapping area is the foundation pit support structure area jointly covered by the two sets of point clouds, determined through spatial range comparison. First, extract the extreme values ​​of the spatial coordinates of the point cloud from the previous period. , , , , , ; , , , , , The coordinate range of the overlapping region is as follows: ; ; ; Only points within the coordinate range of the two point clouds are retained as valid points within the overlapping region. Within the overlapping region, for each valid point in the previous period's point cloud, the valid point in the next period's point cloud with the smallest spatial distance is searched to form a pair of corresponding points. This ultimately forms a set of corresponding point pairs. The spatial distance is calculated using the Euclidean distance formula. The specific calculation process is as follows: For any valid point within the overlapping region of the previous period's point cloud... Its coordinates in the reference coordinate system are ,in, 1a Represents any valid point in the previous point cloud 1. a Traverse all valid points within the overlapping area of ​​the point cloud in the next phase. Its coordinates are ,in, 2b Represents any valid point in the next point cloud 2. b ,calculate With each Euclidean distance between The calculation formula is as follows: ; in, for and Spatial distance between them The correspondence between valid point a in the previous period and valid point b in the next period; , , valid points from the previous point cloud The three-dimensional coordinates , , For the next phase of point cloud valid points The three-dimensional coordinates are traversed, and then the distance is selected. smallest As Corresponding points with the same name form a pair of points with the same name. Set distance threshold Based on the point cloud elevation accuracy set at ±2cm, if the minimum distance If the same point pair is retained, the above operation is repeated to traverse all valid points in the overlapping area of ​​the previous point cloud, and finally a set of same point pairs consisting of multiple valid same point pairs is obtained.

[0024] Based on the spatial coordinate differences of each corresponding point pair in the reference coordinate system, the rotation and translation components that align the later point cloud with the earlier point cloud are calculated to obtain the rigid transformation matrix. Specifically, after obtaining a set of corresponding point pairs that meets the accuracy requirements, the rigid transformation matrix between the earlier and later point clouds can be calculated based on this set. This matrix is ​​the core for achieving accurate alignment between the two point clouds and describes the spatial position change of the later point cloud relative to the earlier point cloud. The expression for the rigid transformation matrix is ​​as follows: ,in It is a 3×3 rotation matrix used to describe the change in rotation angle of the point cloud in the later stage relative to the point cloud in the previous stage. The translation vector is 3×1, used to describe the change in translation distance between the later point cloud and the previous point cloud. 0 represents the zero vector (1×3), and 1 represents a constant term to determine the validity of the matrix operation. The solution process closely relies on the spatial coordinate differences of each pair of corresponding points in the reference coordinate system. The rotation matrix and translation components are obtained through multiple precise calculations. The specific steps are as follows: Calculate the mean coordinates of the set of pairs of corresponding points. Calculating the mean coordinates is fundamental for subsequent decentralized coordinate calculation and covariance matrix construction. Let the set of pairs of corresponding points contain a total of... Group of valid identical point pairs, the first In the group of points with the same name, the points in the previous point cloud are: Its coordinates are The next point cloud point is Its coordinates are ,in Calculate the average coordinates of all points with the same name in the previous point cloud. Calculate the average coordinates of all points with the same name in the point cloud of the next phase. , , , For the first Point cloud points from the previous period in the same group of points The three-dimensional coordinates , , For the first Point cloud points in the next phase of the same group of points The three-dimensional coordinates , , The coordinates of the corresponding points in the previous point cloud are the average values. , , To calculate the mean coordinates of corresponding points in the point cloud for the next period, we need to center the coordinates. Since the mean coordinates can interfere with the calculation of rotation components, after obtaining the mean coordinates, we need to center the coordinates of each pair of corresponding points to eliminate the influence of the mean. For each pair of corresponding points, we calculate the mean coordinates of the previous period's point cloud points. Decentralized coordinates And the next point cloud Decentralized coordinates The calculation formula is: ; ; ; in , , For the first Decentralized coordinates of point cloud points from the previous phase of the group. , , For the first The decentralized coordinates of the point cloud points in the next phase are used to solve for the rotation matrix R. The rotation matrix is ​​a core component of the rigid transformation matrix, and its solution requires constructing a covariance matrix based on the decentralized coordinates obtained in the second step, followed by accurate solution through singular value decomposition. The covariance matrix C is constructed using the following formula: The covariance matrix is ​​analyzed through singular value decomposition. Decompose to obtain ,in and All are 3×3 orthogonal matrices. It is a 3×3 diagonal matrix. Representation matrix Transpose, rotation matrix The solution formula is ,in Representation matrix Transpose of the vector and solve for the translation vector. The translation vector and rotation matrix together form a rigid transformation matrix, the solution of which requires combining the coordinate mean and the obtained rotation matrix, and the translation vector. The calculation formula is: ,in The translation vector is 3×1, and its components are as follows: ; ; ; in , , , , , , , , Rotation matrix The 9 elements, , , They are translation vectors The translation components in the X, Y, and Z directions are integrated to obtain the rigid transformation matrix. After completing the rotation matrix Translation vector After accurate calculation, the two are integrated according to the standard form of a rigid transformation matrix to obtain a complete rigid transformation matrix. The calculated rotation matrix is ​​then used as the basis for this process. Translation vector Substituting into the expression for the rigid transformation matrix, we obtain the complete rigid transformation matrix. ,in It is a 3×3 rotation matrix. The matrix is ​​a 3×1 translation vector, 0 is a 1×3 zero vector in [0,0,0], and 1 is a constant term. This matrix can completely describe the spatial rigid transformation of the point cloud in the later stage relative to the point cloud in the previous stage.

[0025] The rigid transformation matrix is ​​applied point-by-point to the 3D coordinates of each data point in the subsequent point cloud to obtain the transformed point cloud. The previous point cloud and the transformed point cloud are then spatially superimposed to obtain a registered point cloud pair. Specifically, after the rigid transformation matrix is ​​solved, a coordinate mapping operation is performed on the subsequent point cloud to align it with the previous point cloud. The specific operation process is as follows: traverse all data points in the subsequent point cloud, and for each data point... The coordinates are , rigid transformation matrix By applying the transformation to each of its three-dimensional coordinates, the corresponding points in the transformed point cloud are calculated. The coordinates are The formula for calculating coordinate mapping is as follows: ; ; ; in For any data point in the point cloud of the next phase, for Three-dimensional coordinates in the original coordinate system for The corresponding points after rigid transformation for Three-dimensional coordinates in the reference coordinate system Rotation matrix elements, , , Translation vector The components are calculated to map all data points of the subsequent point cloud to the reference coordinate system of the previous point cloud. After the coordinate mapping is completed, the transformed point cloud is obtained, that is, all data points have been transformed to the reference coordinate system. The previous point cloud and the transformed point cloud are then spatially superimposed: keeping the coordinates of the previous point cloud unchanged, the transformed point cloud is placed in the same reference coordinate system as the previous point cloud according to its mapped coordinates. Non-overlapping areas are reasonably distributed according to the actual spatial position. After the superposition is completed, a registered point cloud pair is obtained. In this point cloud pair, the previous point cloud is the reference point cloud, and the transformed point cloud is the point cloud aligned to the reference coordinate system. The spatial positions of the two sets of point clouds correspond one-to-one. The above four steps are repeated to complete the registration operation of all adjacent monitoring period point cloud pairs in turn, and the registered point cloud pairs corresponding to all adjacent monitoring periods are obtained, forming a complete registered point cloud dataset.

[0026] In this embodiment of the invention, by sorting adjacent point clouds according to acquisition time, searching for pairs of corresponding points with the smallest Euclidean distance based on spatial overlapping areas and setting a distance threshold to eliminate mismatches, and generating a rigid transformation matrix by solving the rotation matrix and translation vector using decentralized coordinates and singular value decomposition, the technical problems of inconsistent coordinate systems of point clouds in multiple periods, registration error accumulation caused by dependence on absolute control points, interference of mismatched points with the accuracy of rigid body transformation, and easy failure of reference points in dynamic construction are overcome. The invention achieves the technical effects of sub-centimeter-level high-precision registration of point clouds in adjacent monitoring periods, robust and reliable solution of rigid transformation matrix, and one-to-one correspondence of the spatial position of the registered point cloud for repeated use.

[0027] In a preferred embodiment of the present invention, spatial neighborhood matching is performed on the registered point cloud pairs to obtain a set of neighborhood point pairs; the geometric difference value of each set of neighborhood point pairs is calculated based on the set of neighborhood point pairs to obtain a set of geometric difference values; the corresponding neighborhood point pairs with difference values ​​lower than a preset stability threshold are selected based on the set of geometric difference values ​​to obtain a set of stable point clusters for the support structure, including: Centered on each data point of the previous point cloud in the registered point cloud pair, a spherical search domain with a preset radius is constructed within the subsequent point cloud in the registered point cloud pair. Data points of the subsequent point cloud falling within the spherical search domain are extracted, resulting in neighborhood point pairs corresponding to each center point. Specifically, this involves: based on the registered point cloud pair obtained above, with the previous point cloud as the reference point cloud and the transformed point cloud as the subsequent point cloud aligned to the reference coordinate system, centering on each data point of the previous point cloud in the registered point cloud pair, constructing a spherical search domain with a preset radius within the subsequent point cloud, extracting all subsequent point cloud data points falling within this search domain, and establishing a correspondence with the center point of the previous point cloud to obtain each set of neighborhood point pairs, ultimately forming a set of neighborhood point pairs. The specific operation is as follows: Determine the preset search radius. Based on the point cloud accuracy, planar accuracy ±1cm, elevation accuracy ±2cm, and the surface flatness requirements of the support structure, the following settings are made: Iterate through all data points of the previous point cloud in the registered point cloud pair, and use each data point of the previous point cloud as the center point. Its coordinates in the reference coordinate system are ,in Representing the center point, with the center point Center of the ball, preset radius Construct a spherical search domain with radius , the spatial extent of which satisfies the formula: ,in, Given the 3D coordinates of any data point within the subsequent point cloud, iterate through all data points in the registered point cloud that are aligned with the subsequent point cloud. Determine whether the coordinates of each data point satisfy the aforementioned spherical search domain. If they do, include the subsequent point cloud data point in the neighborhood point set corresponding to the center point, thus obtaining the coordinates relative to the center point. Corresponding neighborhood point pairs ,in To determine the validity of neighboring point set pairs, a neighborhood point set is defined as the set of neighboring point data points that fall within the spherical search domain in the previous period. Minimum number of points threshold If a certain center point The corresponding neighborhood point set Midpoint number less If the number of points is zero, the neighborhood point set is deemed invalid and removed; if the number of points is zero, the neighborhood point set is deemed invalid and removed. If the neighborhood point pair is not found, then retain that neighborhood point pair. Repeat the above operation to traverse all data points of the previous point cloud, and finally obtain a set of neighborhood point pairs consisting of multiple sets of valid neighborhood point pairs.

[0028] Based on the spatial distribution of point cloud data points from the previous and subsequent periods in each pair of neighboring point sets, the angle between the surface normal vectors of the two periods' point clouds within that neighborhood and the deviation of the point-to-surface distance are calculated to obtain the geometric difference value of each pair of neighboring point sets. Specifically, after obtaining the set of neighboring point sets, for each pair in the set, based on the spatial distribution characteristics of the center point of the previous period's point cloud and the neighboring point sets of the subsequent period's point cloud, the angle between the surface normal vectors of the two periods' point clouds within that neighborhood is calculated. Deviation between point and surface Then, the geometric difference values ​​of each set of neighborhood point pairs are obtained by weighted summation. The specific calculation process is as follows: For any set of neighborhood point pairs Select the center point of the previous point cloud. The previous point cloud data points and their five adjacent neighboring points form a neighborhood subset of the previous point cloud. Simultaneously, select the neighborhood point set of the next phase point cloud. All data points in the next phase form a subset of the point cloud neighborhood. To each and Constructing the covariance matrix and The formula for calculating the covariance matrix is ​​as follows: ; in The number of points within the neighborhood subset. For the first in the neighborhood subset The three-dimensional coordinate vector of a point Let the mean vector of coordinates of all points in the neighborhood subset be given by the covariance matrix. and Perform singular value decomposition on each, and obtain , ,in , It is an orthogonal matrix. Given a diagonal matrix, take the diagonal matrix. , The eigenvectors corresponding to the minimum singular values ​​in the middle are used as the neighborhood of the previous point cloud. Surface normal vector and the next phase of point cloud neighborhood Surface normal vector The normal vectors all point outwards from the support structure, based on the surface normal vectors of the previous point cloud neighborhood. and the normal vector of the neighborhood surface of the point cloud in the next phase The angle between the two is calculated using the vector dot product formula. The formula is as follows: ,in The dot product of two normal vectors. , Let the magnitudes of the two normal vectors be the values ​​obtained using the inverse cosine function. Previous point cloud neighborhood Surface normal vector and center point Construct the neighborhood plane of the previous point cloud The equation of the plane is ,Right now ,in , , Normal vector Calculate the neighborhood point set of the next phase point cloud for the components in the X, Y, and Z directions. From each point in the plane The distance to a single point is calculated using the following formula: ; in for The Middle The coordinates of the points For the point to the plane Distance, take all The average value is used as the point-to-surface distance deviation of the neighborhood point set pair. ,Right now ,in for The number of points in the calculation; the geometric difference value. The surface normal vector angle is integrated by weighted summation. Deviation between point and surface Considering that both have the same weight in influencing geometric differences, the weighting coefficient is set to 0.5 for both, and the calculation formula is as follows: ;in, This is the maximum permissible attitude deviation threshold determined based on historical stable region statistics. This is the maximum permissible positional deviation threshold determined based on historical stable region statistical calibration. Specifically, during calibration, multiple sets of point cloud data are collected under the condition of no deformation of the support structure. For the marked stable regions, the local plane angle is calculated according to the same procedure as in this embodiment. Deviation between point and surface The distribution is used, and the upper bound of the 95% confidence interval is taken as the value. and The calibration value, in this embodiment, is... , The values ​​obtained by the above calibration method are typical values, which are then normalized to reduce the geometric difference value. The range of values ​​is The smaller the value, the smaller the geometric difference of the neighborhood point pair, and the more stable the point cloud. Repeat the above calculation process to obtain the geometric difference values ​​of all neighborhood point pairs, forming a geometric difference value set.

[0029] The geometric difference values ​​of each set of neighboring point pairs are compared with a preset stability threshold. Neighboring point pairs with geometric difference values ​​lower than the preset stability threshold are retained to obtain a stable set of neighboring point pairs. Specifically, this includes: after generating the geometric difference value set, the geometric difference values ​​of each set of neighboring point pairs are... Compared with the preset stability threshold One-by-one comparisons are performed to select neighborhood point pairs whose geometric difference value is lower than a preset stability threshold, forming a stable neighborhood point pair set. The specific implementation process is as follows: First, determine the preset stability threshold. To enhance the reproducibility of the threshold, this embodiment adopts a calibration method based on the identification requirements of point cloud noise and support structure deformation: multiple sets of point cloud data are collected under the condition of no deformation of the support structure, the geometric difference distribution of neighborhood point pairs in the no-deformation region is calculated, and the standard deviation of the difference values ​​is statistically obtained. Based on the minimum identifiable deformation requirement for early warning of support structure deformation, the following settings are made: Using three standard deviations as the criterion for distinguishing between stable and potentially deformable regions, this method can adaptively adjust according to the point cloud accuracy and deformation early warning requirements of different engineering scenarios. In typical engineering scenarios, combining the deformation early warning requirements of the support structure, point cloud accuracy, and engineering practice experience, The range of values ​​is ,by This is a commonly used typical value. The selection criteria at this value are moderate, ensuring the reliability of stable point sets while retaining a sufficient number of valid samples. The smaller the value, the stricter the selection criteria, and the higher the reliability of stable point pairs. Iterate through each neighborhood point pair in the neighborhood point pair set and extract its corresponding geometric difference value. ,Will and Comparison: If If the region corresponding to the neighborhood point pair is determined to be without significant deformation, it belongs to a stable neighborhood point pair and is retained; if If a pair of neighboring points is deemed unstable, it is determined that the region corresponding to that pair has potential deformation and is therefore removed. During the selection process, key information for each stable pair of neighboring points is recorded simultaneously, including the coordinates of the center point of the previous point cloud, the number of neighboring points in the next point cloud, and the geometric difference value. After screening, the quality of the stable neighborhood point pair sets is checked to ensure that the number of stable neighborhood point pairs in the set is not less than 60% of the total number of neighborhood point pair sets. If it is lower than this proportion, the geometric difference value calculation process needs to be re-examined or the preset stability threshold needs to be adjusted. After the verification is passed, the final set of stable neighborhood point pairs is obtained.

[0030] Spatial clustering is performed on the previous period point cloud data points contained in each neighborhood point pair in the stable neighborhood point pair set to obtain a stable point cluster set for the support structure. Specifically, this includes: based on the stable neighborhood point pair set, using the previous period point cloud data points contained in each neighborhood point pair in the stable neighborhood point pair set, i.e., the center point of each stable neighborhood point pair... As a clustering object, a spatial density clustering algorithm is used for systematic clustering analysis. The specific implementation process is as follows: Combining the spatial distribution characteristics of the stable points of the support structure, a density clustering algorithm is used as the spatial clustering algorithm. Based on the point cloud density (preset to 50 points / cm²), the support structure dimensions (single support structure length ≥ 1m), and engineering practice experience, two core parameters of the algorithm are scientifically calibrated to ensure that the parameter settings fit the actual engineering scenario. The specific parameter calibration and implementation process is as follows: The two core parameters are the neighborhood radius... , minimum number of points Its calibration is strictly determined by combining the preset point cloud density, support structure dimensions, and point cloud distribution characteristics, as follows: Neighborhood radius Based on the preset point cloud density of 50 points / cm², the average distance between adjacent points in the point cloud is approximately 3cm. Setting the neighborhood radius to 5cm, which is a slight increase compared to the average point spacing, results in a minimum number of points. This parameter setting is based on the minimum number of effective points in the neighborhood point set, combined with the dimensional characteristics of a single-segment support structure with a length ≥1m. The score was increased to 8 points. To further verify the rationality of the parameters, a parameter sensitivity analysis was conducted to compare them. Take 3cm, 5cm, and 7cm. The clustering results at values ​​of 6, 8, and 10 were used to determine the final clustering outcome. , To achieve the optimal parameter combination, after parameter calibration, the clustered objects are preprocessed to extract the center points of all stable neighborhood point pairs. Three-dimensional coordinates form the clustering input dataset ,in To stabilize the total number of neighborhood point pairs and remove center points with abnormal coordinates from the input dataset, such as points whose coordinates exceed the preset range of the support structure, the following determination operation is performed for core points, boundary points, and isolated points: using the current center point... Center of the sphere, neighborhood radius Construct a spherical neighborhood with radius r, the spatial extent of which satisfies the following formula: ; in Let be the 3D coordinates of other center points in the dataset; count the number of other center points contained within this spherical neighborhood, denoted as . ; Determine the point type based on the statistical results: if Then the current center point Marked as a core point, the core point is the fundamental basis for forming a stable point cluster, and all points in its neighborhood are density-reachable points; if Then the current center point Marked as boundary points, boundary points themselves do not possess the density conditions of core points, but are within the neighborhood of core points, and need to be subsequently assigned to the corresponding core point clusters; if Then the current center point Points marked as isolated (not adjacent to any other centroids) are likely to be point cloud noise or locally unstable and are directly removed from subsequent clustering operations. After traversal, the core point set is obtained. Boundary point set The initial determination of point types is completed; density reachability analysis and temporary point cluster generation are performed. Based on the core point set and boundary point set, temporary point clusters are constructed according to the density reachability rules of the spatial density clustering algorithm: from the core point set... Select an unlabeled core point Starting from this core point, recursively traverse its... All core points within the neighborhood are grouped into the same temporary cluster. Simultaneously mark all core points included in the temporary point cluster; traverse the temporary point cluster. All the core points Neighborhood is the set of all boundary points within the neighborhood. The boundary points belong to this temporary point cluster. Repeat the above two operations until the core point set is reached. All core points are marked and included in the corresponding temporary point clusters, resulting in multiple temporary point clusters. Each temporary point cluster consists of a core point and its associated boundary points, initially reflecting the spatial distribution of the stable area of ​​the support structure. After obtaining multiple temporary point clusters through clustering, each temporary point cluster needs to undergo quality screening to remove invalid point clusters and optimize valid point clusters. The specific operations are as follows: Point count screening: Set a minimum point count threshold for each stable point cluster. Iterate through each temporary point cluster one by one, and count the total number of center points within the cluster. If the number of points in a certain temporary point cluster is... If the number of points is low, the cluster is considered a candidate stable cluster and is retained; otherwise, the cluster is considered stable. If the clusters are not found to be stable, they are considered invalid and are removed. Spatial continuity screening: The spatial dispersion of each candidate stable cluster is calculated using the following formula: ; in, The number of points within the candidate point cluster. For the first in the cluster The coordinates of the center point The mean coordinates of all center points within the candidate point cluster; set the dispersion threshold. ,like If the candidate point cluster has good spatial continuity, it is determined to be retained; otherwise... If the candidate point cluster is spatially discrete, it is determined to be discarded. Point cluster parameter calculation: For the effective stable point clusters after screening, two core parameters are calculated: one is the coordinates of the center point of the point cluster, that is, the average coordinates of all center points in the cluster. First, locate the center of the stable region; second, the point density. Point density Must meet After all effective stable point clusters have been screened and optimized, they are integrated to form a set of stable point clusters for the support structure. ,in To effectively stabilize the number of point clusters.

[0031] In this embodiment of the invention, a spherical search domain is constructed with a preset radius centered on a reference point cloud data point to extract the next-phase neighborhood point set. The deviation between the angle between the surface normal vectors and the distance between the points and the surface is calculated. Finally, a stable point cluster set is generated through spatial density clustering and triple verification of point count, spatial continuity, and point density. This overcomes the technical problems of local pseudo-differential noise caused by temporary facilities and surcharges under dynamic construction interference and the inability of a single geometric index to fully characterize stability. It achieves the technical effects of automatic identification of stable regions, effective suppression of pseudo-differential noise, and accurate and reliable spatial location of stable point clusters in the support structure.

[0032] In a preferred embodiment of the present invention, based on the spatial position of each point cluster in the set of stable point clusters of the support structure on the surface of the foundation pit support structure, the distance change of the registered point cloud pair along the surface normal direction at that position is calculated to obtain a three-dimensional deformation vector field, including: The three-dimensional coordinates of the center of each point cluster are extracted from the set of stable point clusters of the support structure to obtain the set of spatial locations of stable point clusters. Based on each spatial coordinate point in the set of spatial locations of stable point clusters, neighborhood point clusters centered on each coordinate point are extracted from the previous point cloud of the registered point cloud pair to obtain the local neighborhood point set corresponding to each spatial coordinate point. Specifically, this includes: the set of stable point clusters of the support structure obtained through clustering. ;in, , To determine the total number of stable point clusters, traverse each stable point cluster one by one. The average coordinates calculated during the preprocessing stage of this point cluster are called. ;in, Representing the The average coordinates of each stable point cluster are used as the center three-dimensional coordinates of the corresponding point cluster. The spatial location set of the stable point clusters is obtained by integrating the center three-dimensional coordinates of all stable point clusters. , ;in, For the first The center coordinates of each stable point cluster are recorded synchronously, along with auxiliary information such as the cluster number and number of points corresponding to each center coordinate; this is for the spatial location set of stable point clusters. Each spatial coordinate point Only in the previous point cloud of the registered point cloud pair, using the previous point cloud as the reference point cloud, can the selected neighboring points reflect the original surface morphology of the support structure. A spherical search domain is constructed around the center. Based on the point cloud density (preset to 50 points / cm²), stable point cluster size, and local surface fitting requirements, the search radius is set. The spatial extent of the spherical search domain satisfies the following formula: ; in, Given the 3D coordinates of any data point within the previous point cloud; iterate through all data points in the previous point cloud, checking whether the coordinates of each data point satisfy the above spherical search domain. If they do, include the data point in the corresponding domain. The local neighborhood point set is obtained to get the first The local neighborhood point set corresponding to each spatial coordinate point Set a minimum number of points threshold According to the requirements of the tangent plane fitting algorithm, too few points will lead to insufficient fitting accuracy. corresponding Midpoint number less If the search radius is adjusted to 5 cm and the truncation operation is repeated, and the condition is still not met, the stable point cluster is removed, finally obtaining the local neighborhood point set. ;in, This represents the number of effective spatial coordinate points, which correspond one-to-one with the spatial location of the stable point cluster.

[0033] By fitting a local tangent plane to the local neighborhood point set corresponding to each spatial coordinate point and extracting the unit normal vector of the tangent plane, the surface normal direction corresponding to each spatial location is obtained. Specifically, this includes: retrieving the local neighborhood point set. Each local neighborhood point set in ,use The criteria are used for noise reduction preprocessing, among which... To calculate the standard deviation, the specific steps are as follows: Calculate the mean of the three-dimensional coordinates of all valid data points in the point set. Simultaneously calculate the standard deviation of each coordinate axis direction, based on The criteria define the range for outlier detection, eliminating discrete outliers whose coordinate components exceed any range. The detection range is as follows: ; ; ; in, Standard deviation , , The mean of the three-dimensional coordinates of the valid data points is given after preprocessing. The remaining The number of valid data points is determined sequentially. The coordinates of the data points are labeled as follows: ;in The least squares method is used to fit the local tangent plane. Let the equation of the fitted tangent plane be... ,in , , , Let be the coefficients of the equation for the tangent plane, where Corresponding to the Add constraints to a local neighborhood set of points. Simultaneously, a least squares objective function is constructed. By finding the objective function respectively , , , Taking the partial derivatives and setting them to zero yields a system of linear equations. Solving this system yields the four coefficients of the plane equation, completing an accurate fit to the local tangent plane. After fitting, the normal vector of the tangent plane is directly defined based on the coefficients of the plane equation. Normalize the normal vector to obtain the unit normal vector. To achieve unified calibration of the normal vector direction, the outer direction vector of the foundation pit is determined. This vector can be determined in two ways: 1) Based on the global coordinate system of the foundation pit design drawings, the outer direction vector is directly set according to the design orientation of the support structure; 2) Through principal component analysis of the point cloud, the overall point cloud data of the support structure is decomposed into principal components, and the direction of the principal component with the largest variance in the point cloud distribution is taken as the outer direction vector. In this embodiment, for ease of calculation, the outer direction vector is preset to... This vector parameter can be adjusted according to the actual orientation of the foundation pit. During calibration, it is calculated... The dot product of the normal vector with the preset outer direction vector is used. If the dot product is negative, the direction of the normal vector is reversed, ensuring that all unit normal vectors point outward from the support structure. The unit normal vectors corresponding to all local neighborhood points are then summed to obtain the set of surface normal directions. This set is related to the set of spatial locations of stable point clusters. Local neighborhood point set Strict one-to-one correspondence.

[0034] Starting with the data point at the stable point cluster location of the previous point cloud after registration, a projection ray is constructed along the surface normal direction corresponding to that location. The corresponding projection point for the next period is obtained by finding the intersection of this projection ray with the next point cloud after registration. Specifically, this involves: starting with the data point at the stable point cluster location of the previous point cloud after registration, a projection ray is constructed along the surface normal direction corresponding to that location. The corresponding projection point for the next period is obtained by finding the intersection of this projection ray with the next point cloud after registration. Since the previous point cloud has already been registered and aligned with the next point cloud, the coordinates of the previous point cloud corresponding to the center of the stable point cluster are selected. ,coordinate Points that completely overlap are used as the starting point for projection. This starting point can precisely correspond to the same spatial location of the support structure, and synchronously match the unit normal vector corresponding to that spatial location. It is decomposed into components in the X, Y, and Z directions, and labeled as follows: , , Based on the projection origin and normal components, the parametric equations of the projection ray are constructed, and the equations are in the following form: ; in, The parameter is the distance between the rays. Considering the typical deformation magnitude of the foundation pit support structure, which generally does not exceed 5cm, the parameter is... The range of values ​​is strictly set as follows: Iterate through all valid data points in the next phase point cloud after registration, and calculate the value of each individual data point in the next phase point. Its coordinates are The spatial perpendicular distance to the projected ray is calculated using the vector cross product formula, specifically expressed as follows: ,in, Let the vectors from the next data point to the projection starting point be vectors. The magnitude of the cross product is the perpendicular distance between the two points. The vector is a unit normal with a magnitude of 1, so no additional magnitude normalization is needed. To select the point cloud points that precisely correspond to the projected ray in the next phase, a distance threshold is set. This threshold, combined with the point cloud acquisition accuracy of ±1cm, can effectively eliminate irrelevant points that are too far away, retaining only those with vertical distance. The points are selected as candidate projection points. A differentiated classification process is performed based on the number of candidate projection points to determine their uniqueness and accuracy: if a projection ray corresponds to only 1 candidate projection point, then that candidate point is the projection point for the corresponding spatial location in the next period. The three-dimensional coordinates are recorded directly; if the number of candidate projection points is greater than 1, the parameters corresponding to each candidate point are solved by inverse equation of projection ray parametric equation. The calculation formula is: ,in, Select the coordinates of the candidate points. The candidate point with the smallest positive value is selected as the projection point. If the number of candidate projection points is 0, the projection of that spatial location is deemed unsuccessful and the corresponding point is directly discarded. , and All qualified projection points for the next period, along with relevant data, will be collected and grouped together to form a projection point set. ,in The set represents the number of effective projection points, and it corresponds one-to-one with the spatial location of the effective stable point cluster.

[0035] The directed distance between the starting point and the corresponding projection point in the next phase along the surface normal direction is calculated to obtain the distance change at each spatial location. The three-dimensional coordinates of each spatial location in the set of stable point clusters are correlated with the corresponding distance change to obtain a three-dimensional deformation vector field. Specifically, this includes retrieving the projection starting points of each matched set. With the next period of projection points Construct the spatial offset vector between two points ; in, For the next period of projection points The three-dimensional coordinates As the starting point of the projection The three-dimensional coordinates of the vector are given, with its three components corresponding to the offsets in the X, Y, and Z directions, respectively. Since the deformation of the support structure mainly occurs along its surface normal direction, the offset vector is calculated in relation to the corresponding unit normal vector. The dot product yields the directed distance along the normal direction. This directed distance is the change in distance at the corresponding spatial location, and the specific calculation formula is as follows: ; when When, it means that the point cloud of the later period shifts outward along the normal direction relative to the point cloud of the previous period, that is, the corresponding position of the support structure deforms outward; when When, it means that the point cloud of the later period is offset inward along the normal direction relative to the point cloud of the previous period, that is, the corresponding position of the support structure is deformed inward; when At this time, it means there is no offset between the two point clouds, that is, there is no significant deformation of the support structure at this location. The distance changes of all effective spatial locations are summarized to obtain the set of distance changes. This set and the set of projection points One-to-one correspondence, stabilizing the spatial location of the point cluster. 3D coordinates The corresponding change in distance and unit normal vector It includes three directional components, which are correlated one-to-one to construct a standardized three-dimensional deformation vector. This vector contains three core pieces of information: spatial location, deformation magnitude, and deformation direction. By integrating all effective three-dimensional deformation vectors, a complete three-dimensional deformation vector field of the foundation pit support structure is obtained. Set strict data validation rules: calculate all valid data. The absolute value of the mean, i.e. If the mean value is greater than 2 cm, combined with the verification threshold set by the allowable deformation of the support structure design, the vector field is judged to be abnormal. It may be caused by incorrect projection point selection or distance calculation error. The entire process of obtaining projection points and calculating distance needs to be reviewed again to find out the cause of the abnormality and correct it. If the mean value is ≤ 2 cm, the vector field verification is judged to be qualified.

[0036] In this embodiment of the invention, the spatial position is located by using the center coordinates of a stable point cluster, and the surface normal direction is obtained by fitting a tangent plane to the local neighborhood point set in the previous point cloud. The projection ray is constructed along the normal direction starting from the previous point and intersects with the next point cloud to obtain the corresponding projection point. Then, the directed distance along the normal direction is calculated as the distance change and associated with the spatial coordinates to generate a three-dimensional deformation vector field. This overcomes the technical problems of deformation calculation distortion caused by the accumulation of traditional global registration errors, difficulty in reflecting the true normal deformation of the structure by relying solely on absolute coordinate displacement, and inability to distinguish the inward and outward deformation directions. It achieves the technical effects of accurate spatial position of deformation vector field, reliable normal deformation quantification, distinguishable inward and outward deformation directions, and data verification to ensure the consistency of vector field.

[0037] In a preferred embodiment of the present invention, the distance change at each spatial location in the three-dimensional deformation vector field is used as the deformation displacement value at that spatial location, and compared with a preset warning threshold to generate a deformation warning level for that spatial location, thereby completing the deformation warning during the construction period, including: The three-dimensional coordinates and corresponding distance changes of each spatial location are extracted from the three-dimensional deformation vector field to obtain a set of deformation data points. The distance change of each spatial location in the set of deformation data points is assigned as the deformation displacement value of that spatial location, resulting in a set of deformation data points with displacement labels. Specifically, this includes: traversing the three-dimensional deformation vector field globally. Each three-dimensional deformation vector in Extract the core information from the vector separately to stabilize the three-dimensional coordinates of the point cluster's spatial location. and the corresponding distance change Remove normal direction components , , Redundant parameters are used to combine the extracted core information into a single deformation data point. All deformation data points are aggregated and a basic set of deformation data points is constructed. Due to the change in distance The positive and negative signs are only used to distinguish the direction of deformation, while the core of deformation early warning is to determine whether the magnitude of deformation exceeds the safe range. Therefore, deformation displacement values ​​are set for each spatial location. That is, the absolute value of the distance change is taken as the basis for determining the actual deformation amplitude. This displacement value is non-negative and only represents the actual deformation magnitude of the support structure surface. The deformation data points are then set together. Each data point Internal Replace with the corresponding deformation displacement value Generate new data units with uniform displacement markers. To further ensure data validity, abnormal data points caused by point cloud noise and projection errors are removed, and an abnormal deformation screening threshold is set. This threshold is set in conjunction with the maximum allowable deformation of the support structure. If the deformation displacement value of a certain data point... If the data point is an outlier, it will be removed; if If a data point is found to be valid, it is retained. All valid deformation data points with displacement markers are then integrated to form a set of deformation data points with displacement markers. ,in, To ensure an effective number of data points with displacement markers, each data point in this set contains defined spatial coordinates and a corresponding deformation displacement value.

[0038] The deformation displacement value at each spatial location in the set of deformation data points with displacement markers is compared with preset first-level and second-level warning thresholds. Spatial locations with deformation displacement values ​​below the first-level warning threshold are extracted and marked as safe; spatial locations with deformation displacement values ​​between the first-level and second-level warning thresholds are extracted and marked as attention-level; spatial locations with deformation displacement values ​​above the second-level warning threshold are extracted and marked as warning level. Specifically, based on the relevant specifications of the "Technical Specification for Foundation Pit Support", on-site construction safety management standards, and point cloud acquisition accuracy of ±1cm, first-level, second-level warning thresholds and abnormal critical thresholds are set in layers. All thresholds are determined using a reproducible calibration method. The specific calibration process is as follows: Multiple sets of point cloud data are collected under the condition of no deformation of the foundation pit support structure. Through the same neighborhood truncation, plane fitting, and deformation calculation process, the distribution of deformation displacement values ​​in the no-deformation area is obtained, and the standard deviation of this distribution is calculated. Based on the minimum identifiable deformation of the support structure and the engineering safety redundancy requirements, a first-level early warning threshold is set. Level II warning threshold Abnormal threshold In typical engineering scenarios, based on the above calibration method, the first-level early warning threshold is... The range of values ​​is Level II warning threshold The range of values ​​is Abnormal threshold The range of values ​​is Among them , , These are commonly used typical values. While ensuring early warning sensitivity, these values ​​can be adaptively adjusted in practical applications based on engineering parameters such as the excavation depth, soil and rock properties, and support structure material. The set of deformation data points with displacement markers is iterated sequentially. Each data point in Extract its corresponding deformation displacement value Each data point is compared with a preset threshold, and a unique warning level is assigned to each data point based on the comparison results. The specific judgment rules are as follows: When When the structural deformation in the monitored area is determined to be within the safe range allowed by the specifications, the overall support structure is stable, and there are no safety hazards, it is uniformly marked as a safety level and labeled as SAFE. Subsequent routine daily monitoring will be maintained once a day, and no additional control measures are required. At that time, it was determined that the area showed slight to moderate deformation. The overall support structure was under control, but there was a risk of continued deformation development. Close monitoring of the trend was necessary, and the area was uniformly marked as ATTENTION. The monitoring frequency was subsequently increased to once every 4 hours, and the deformation development pattern was analyzed regularly. When the deformation in the area is deemed excessive, posing a significant construction safety hazard, and failure to address it promptly could lead to further deformation, it is uniformly designated as a warning level and marked as WARNING. An immediate on-site inspection is organized, and control measures such as localized reinforcement and additional support are implemented. Heavy-load construction operations in the surrounding area are restricted, and monitoring frequency is increased to once every hour. When the area is determined to have severe deformation, indicating a potential risk of instability and collapse of the support structure, it is uniformly marked as a severe warning level, designated CRITICAL. Emergency response procedures are immediately initiated, all surrounding construction activities are halted, personnel and equipment are evacuated to a safe area, and emergency reinforcement and support measures are implemented. After the warning level is marked, each data point is... Add warning level labels and update accordingly. ,in, To correspond to the warning level, the specific value of deformation displacement, threshold comparison difference, judgment time, and operator information for each data point are recorded synchronously.

[0039] The three-dimensional coordinates of each spatial location are associated with and stored with the corresponding warning level labels to complete the deformation warning during the construction period. Specifically, this includes: structuring the deformation monitoring data points with warning level labels; forming a standardized monitoring ledger according to the correspondence between three-dimensional coordinates, deformation displacement values, warning levels, judgment criteria, and judgment time; archiving the data in the engineering monitoring database; establishing a multi-dimensional retrieval index based on warning level, spatial location, and time; automatically generating a phased construction period deformation warning report after data archiving; pushing the report to relevant management terminals and displaying the warning information in real time at the construction site; conducting a second review of the associated stored data and warning information; and completing the deformation warning process for the support structure during the construction period after confirming that there are no errors. The monitoring data is updated regularly and the deformation trend is continuously analyzed. The monitoring frequency and response measures are dynamically adjusted according to changes in the warning level until the support structure reaches a stable state, thus completing the deformation warning for the construction period.

[0040] In this embodiment of the invention, by extracting the spatial coordinates and distance changes from the three-dimensional deformation vector field and taking the absolute value as the deformation displacement value, the technical problems of the inability of a single threshold to distinguish the degree of deformation, the coarse division of warning levels, the lack of spatial location-based hierarchical response, and the lack of data archiving are overcome. The invention achieves the technical effects of accurate quantification of deformation displacement value, automatic determination of four-level warning level, spatial location of high-risk areas, structured archiving of monitoring data, and dynamic tracking.

[0041] like Figure 2 As shown, embodiments of the present invention also provide a construction engineering deformation early warning system based on UAV oblique photography, comprising: The acquisition module is used to acquire multi-view image sequences and airborne positioning and attitude data of the foundation pit support structure collected by UAVs equipped with oblique photography cameras along preset routes during multiple monitoring periods in the construction period; and to reconstruct three-dimensional point clouds from the multi-view image sequences and airborne positioning and attitude data of the same monitoring period to obtain dense three-dimensional point clouds of the foundation pit support structure surface in each monitoring period. The processing module is used to extract the point cloud of the previous period and the point cloud of the next period from the dense 3D point cloud of each monitoring period. Using the coordinate system of the previous period point cloud as the reference, it calculates the rigid transformation matrix between the previous period point cloud and the next period point cloud. The rigid transformation matrix is ​​used to perform coordinate mapping on the next period point cloud to obtain the registered point cloud pair. The matching module is used to perform spatial neighborhood matching on the registered point cloud pairs to obtain a set of neighborhood point pairs; calculate the geometric difference value of each set of neighborhood point pairs based on the set of neighborhood point pairs to obtain a set of geometric difference values; filter the corresponding neighborhood point pairs whose difference values ​​are lower than the preset stability threshold based on the set of geometric difference values ​​to obtain a set of stable point clusters of the support structure. The calculation module is used to calculate the distance change of the registered point cloud pair along the surface normal direction of the surface at the location of each point cluster in the set of stable point clusters of the support structure, based on the spatial position of each point cluster on the surface of the foundation pit support structure, and to obtain the three-dimensional deformation vector field. The evaluation module is used to take the distance change of each spatial location in the three-dimensional deformation vector field as the deformation displacement value of that spatial location, compare it with the preset warning threshold, generate the deformation warning level of that spatial location, and complete the deformation warning during the construction period.

[0042] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0043] Embodiments of the present invention also provide a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when executed by the processor, performs the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0044] Embodiments of the present invention also provide a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the method described above. All implementations in the above method embodiments are applicable to this embodiment and can achieve the same technical effects.

[0045] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for early warning of deformation during the construction period of building engineering based on UAV oblique photography, characterized in that, The method includes: Acquire multi-view image sequences and airborne positioning and attitude data of the foundation pit support structure collected by UAVs equipped with oblique photography cameras along a preset flight path during multiple monitoring periods of the construction period; reconstruct three-dimensional point clouds from the multi-view image sequences and airborne positioning and attitude data of the same monitoring period to obtain dense three-dimensional point clouds of the foundation pit support structure surface in each monitoring period. Based on the dense 3D point clouds of each monitoring period, extract the point clouds of the previous period and the point clouds of the next period of adjacent monitoring periods. Using the coordinate system of the previous period point cloud as the reference, calculate the rigid transformation matrix between the previous period point cloud and the next period point cloud. Perform coordinate mapping on the next period point cloud through the rigid transformation matrix to obtain the registered point cloud pair. Spatial neighborhood matching is performed on the registered point cloud pairs to obtain a set of neighborhood point pairs; the geometric difference value of each set of neighborhood point pairs is calculated based on the set of neighborhood point pairs to obtain a set of geometric difference values; based on the set of geometric difference values, the corresponding neighborhood point pairs with difference values ​​lower than the preset stability threshold are selected to obtain a set of stable point clusters of the support structure. Based on the spatial position of each point cluster in the set of stable point clusters of the support structure on the surface of the foundation pit support structure, the distance change of the registered point cloud pair along the surface normal direction at that position is calculated to obtain the three-dimensional deformation vector field. The change in distance at each spatial location in the three-dimensional deformation vector field is taken as the deformation displacement value at that spatial location, and compared with a preset warning threshold to generate the deformation warning level at that spatial location, thus completing the deformation warning during the construction period.

2. The method for early warning of deformation during construction of building projects based on UAV oblique photography according to claim 1, characterized in that, Acquire multi-view image sequences and airborne positioning and attitude data of the foundation pit support structure collected by a drone equipped with an oblique photography camera along a preset flight path during multiple monitoring periods in the construction period; reconstruct 3D point clouds from the multi-view image sequences and airborne positioning and attitude data of the same monitoring period to obtain dense 3D point clouds of the foundation pit support structure surface for each monitoring period, including: Acquire multi-view image sequences of the foundation pit support structure and corresponding airborne positioning and attitude data collected by a drone equipped with an oblique photography camera along a preset route during multiple monitoring periods in the construction period. Select a multi-view image sequence and airborne positioning and attitude determination data corresponding to a monitoring period, extract feature points from the multi-view images within the monitoring period, and perform matching of the same-name feature points between adjacent images to obtain a set of feature point matching pairs for the monitoring period. Based on the feature point matching set and airborne positioning and attitude data for the monitoring period, the camera spatial position and attitude at each image exposure time are calculated to obtain the set of camera pose parameters for the monitoring period. Based on the camera pose parameter set and feature point matching during the monitoring period, sparse three-dimensional reconstruction is performed on the set to generate a sparse three-dimensional point cloud on the surface of the foundation pit support structure during the monitoring period. Based on the sparse 3D point cloud and camera pose parameter set of the monitoring period, the depth map corresponding to each image is calculated pixel by pixel and multi-view depth map fusion is performed to obtain the dense 3D point cloud of the monitoring period. For each monitoring period, feature point extraction and matching, camera pose calculation, sparse 3D reconstruction, depth map calculation and multi-view fusion operations are performed sequentially to obtain dense 3D point clouds for each monitoring period.

3. The method for early warning of deformation during construction of building projects based on UAV oblique photography according to claim 2, characterized in that, Based on the dense 3D point clouds of each monitoring period, the point clouds of the previous and subsequent periods of adjacent monitoring periods are extracted. Using the coordinate system of the previous point cloud as a reference, the rigid transformation matrix between the previous and subsequent point clouds is calculated. Coordinate mapping is then performed on the subsequent point cloud using the rigid transformation matrix to obtain registered point cloud pairs, including: Based on the dense 3D point clouds of each monitoring period, two adjacent monitoring periods are selected in the order of collection time. The dense 3D point cloud corresponding to the monitoring period with the earlier collection time is taken as the previous period point cloud, and the dense 3D point cloud corresponding to the monitoring period with the later collection time is taken as the next period point cloud, thus obtaining the point cloud pair of adjacent monitoring periods. Using the three-dimensional spatial coordinate system of the previous point cloud as the reference coordinate system, search for the corresponding point pairs with the smallest spatial distance in the overlapping area of ​​the previous and subsequent point clouds to obtain a set of point pairs with the same name. Based on the spatial coordinate difference of each pair of points in the reference coordinate system, the rotation and translation components that align the point cloud of the later period with the point cloud of the previous period are calculated to obtain the rigid transformation matrix. The rigid transformation matrix is ​​applied point by point to the three-dimensional coordinates of each data point in the next point cloud to obtain the transformed point cloud; the previous point cloud and the transformed point cloud are spatially superimposed to obtain the registered point cloud pair.

4. The method for early warning of deformation during construction of building projects based on UAV oblique photography according to claim 3, characterized in that, Spatial neighborhood matching is performed on the registered point cloud pairs to obtain a set of neighborhood point pairs. The geometric difference value of each neighborhood point pair is calculated based on this set to obtain a set of geometric difference values. Based on this set, neighborhood point pairs with difference values ​​lower than a preset stability threshold are selected to obtain a set of stable point clusters for the support structure, including: Using each data point of the previous point cloud in the registered point cloud pair as the center, a spherical search domain is constructed within the next point cloud in the registered point cloud pair with a preset radius. Data points of the next point cloud that fall within the spherical search domain are extracted to obtain the neighborhood point set pairs corresponding to each center point. Based on the spatial distribution of point cloud data points from the previous period and the next period in each set of neighboring point sets, the angle between the surface normal vectors of the two periods of point clouds in the neighborhood and the deviation of the point-to-surface distance are calculated to obtain the geometric difference value of each set of neighboring point sets. The geometric difference value of each set of neighborhood point pairs is compared with a preset stability threshold. Neighbor point pairs with geometric difference values ​​lower than the preset stability threshold are retained to obtain a set of stable neighborhood point pairs. Spatial clustering of the previous point cloud data points contained in each neighborhood point pair in the stable neighborhood point pair set is performed to obtain the stable point cluster set of the support structure.

5. The method for early warning of deformation during construction of building projects based on UAV oblique photography according to claim 4, characterized in that, Based on the spatial position of each point cluster in the stable point cluster set of the support structure on the surface of the foundation pit support structure, the distance change of the registered point cloud pair along the surface normal direction at that position is calculated to obtain the three-dimensional deformation vector field, including: The three-dimensional coordinates of the center of each point cluster are extracted from the set of stable point clusters of the support structure to obtain the set of spatial locations of stable point clusters. Based on each spatial coordinate point in the set of spatial locations of stable point clusters, the neighborhood point cluster centered on the coordinate point is extracted from the previous point cloud of the registered point cloud pair to obtain the local neighborhood point set corresponding to each spatial coordinate point. The local tangent plane is fitted based on the local neighborhood point set corresponding to each spatial coordinate point, and the unit normal vector of the tangent plane is extracted to obtain the surface normal direction corresponding to each spatial position. Starting from the data point at the stable point cluster spatial position of the previous point cloud after registration, a projection ray is constructed along the surface normal direction corresponding to that position. The corresponding projection point of the next period is obtained through the intersection point of the projection ray and the next point cloud after registration. The directed distance between the starting point and the corresponding projection point in the next period along the surface normal direction is calculated to obtain the distance change of each spatial position; the three-dimensional coordinates of each spatial position in the set of spatial positions of the stable point cluster are correlated with the corresponding distance change to obtain the three-dimensional deformation vector field.

6. The method for early warning of deformation during construction of building projects based on UAV oblique photography according to claim 5, characterized in that, The change in distance at each spatial location in the three-dimensional deformation vector field is used as the deformation displacement value at that location. This value is then compared with a preset warning threshold to generate a deformation warning level for that location, thus completing the deformation warning during the construction period. This includes: The three-dimensional coordinates and corresponding distance changes of each spatial location are extracted from the three-dimensional deformation vector field to obtain a set of deformation data points; the distance change of each spatial location in the set of deformation data points is assigned as the deformation displacement value of that spatial location to obtain a set of deformation data points with displacement labels. The deformation displacement value of each spatial location in the set of deformation data points with displacement markers is compared with the preset first-level warning threshold and second-level warning threshold. Spatial locations with deformation displacement values ​​lower than the first-level warning threshold are extracted and marked as safe; spatial locations with deformation displacement values ​​between the first-level warning threshold and the second-level warning threshold are extracted and marked as attention level; spatial locations with deformation displacement values ​​higher than the second-level warning threshold are extracted and marked as warning level. The three-dimensional coordinates of each spatial location are associated with and stored with the corresponding warning level markers to complete the deformation warning during the construction period.

7. A construction deformation early warning system based on UAV oblique photography, wherein the system implements the method as described in any one of claims 1 to 6, characterized in that, include: The acquisition module is used to acquire multi-view image sequences of the foundation pit support structure and airborne positioning and attitude data collected by UAVs equipped with oblique photography cameras along preset routes during multiple monitoring periods in the construction period. Three-dimensional point cloud reconstruction was performed on the multi-view image sequence and airborne positioning and attitude data of the same monitoring period to obtain dense three-dimensional point clouds on the surface of the foundation pit support structure in each monitoring period. The processing module is used to extract the point cloud of the previous period and the point cloud of the next period from the dense 3D point cloud of each monitoring period. Using the coordinate system of the previous period point cloud as the reference, it calculates the rigid transformation matrix between the previous period point cloud and the next period point cloud. The rigid transformation matrix is ​​used to perform coordinate mapping on the next period point cloud to obtain the registered point cloud pair. The matching module is used to perform spatial neighborhood matching on the registered point cloud pairs to obtain a set of neighborhood point pairs; calculate the geometric difference value of each set of neighborhood point pairs based on the set of neighborhood point pairs to obtain a set of geometric difference values; and filter the corresponding neighborhood point pairs whose difference values ​​are lower than the preset stability threshold based on the set of geometric difference values ​​to obtain a set of stable point clusters of the support structure. The calculation module is used to calculate the distance change of the registered point cloud pair along the surface normal direction of the surface at the location of each point cluster in the set of stable point clusters of the support structure, based on the spatial position of each point cluster on the surface of the foundation pit support structure, and to obtain the three-dimensional deformation vector field. The evaluation module is used to take the distance change of each spatial location in the three-dimensional deformation vector field as the deformation displacement value of that spatial location, compare it with the preset warning threshold, generate the deformation warning level of that spatial location, and complete the deformation warning during the construction period.

8. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images