Pole tower deformation feature extraction method and device combining image and point cloud, and medium

Abstract

This invention relates to a method, device, and medium for extracting tower deformation features by combining images and point clouds, and pertains to the field of power tower deformation detection technology. The method includes acquiring multiple periods of original images and laser point clouds of transmission towers to obtain dense point clouds of the images and laser point clouds of the towers; registering and fusing the data from the same period to obtain a fused point cloud of the towers; cross-period registration of the fused point clouds of the towers from multiple periods to unify the coordinate space; extracting stable feature points from the original images from multiple periods; transforming the stable feature points from one period to the space of the fused point cloud of another period to obtain the three-dimensional position of the stable feature points in the other period; determining a search area based on the three-dimensional position; searching for and matching multiple candidate feature points within the search area to obtain a candidate matching set containing one-to-many candidate matches; and then refining the set by local geometric consistency constraints to obtain dense deformation feature point pairs. This invention solves the problems of data sparsity and lack of model validation in tower deformation detection.

Description

Technical Field

[0001] This invention relates to the field of power pole deformation detection technology, and in particular to a method, device and storage medium for extracting pole deformation features by combining images and point clouds. Background Technology

[0002] As a critical infrastructure for power transmission, the structural health of transmission towers directly affects the operational safety of the power grid. Under harsh environments such as ice loads and wind loads, towers are prone to non-uniform deformations such as bending and torsion. Therefore, accurately obtaining three-dimensional deformation data of towers is of great significance for structural safety assessment and prevention of tower collapse accidents.

[0003] However, existing technologies face two major technical bottlenecks in tower deformation detection:

[0004] (1) Lack of model verification: Numerical simulation based on finite element method (such as ANSYS) is the main analysis method, but it relies on idealized loads and boundary conditions and lacks high-density, full-coverage measured data for effective verification and calibration.

[0005] (2) Data sparsity: Whether it is sensor placement or laser point cloud, the density of three-dimensional measurement points provided is relatively limited, making it difficult to capture the continuous deformation of rods such as small bending and torsion.

[0006] The two major bottlenecks mentioned above manifest themselves in the specific implementation of existing technologies as follows:

[0007] Firstly, numerical simulation methods based on finite element simulation are difficult to verify due to the lack of high-density measured data.

[0008] Currently, researchers typically use finite element analysis software (such as ANSYS) to perform mechanical modeling and simulation analysis of tower structures. For example, the existing scheme 1 (Liu Wenfeng, Yuan Xiang, He Baocheng, Liu Jun. "Analysis of the influence of uneven icing on the stress of 110kV towers" [J]. Guangdong Electric Power, 2019, 32(11): 136-143) simulated the nodal displacement under icing conditions using ANSYS software (mentioning that the nodal displacement can reach 100+ mm). However, this type of method is essentially based on theoretical deduction of idealized loads and boundary conditions. As stated in the existing scheme 1, its research relies on idealized models and does not provide quantitative real measured data for calibration. Due to the lack of high-density, full-coverage real three-dimensional deformation data as a verification basis, the simulation results often deviate from the actual situation and are difficult to truly reflect the subtle damage or cumulative deformation of the tower structure.

[0009] Secondly, sensor-based physical monitoring methods, while providing measured data, have sparse data points and are costly.

[0010] To obtain accurate data, existing technologies propose deploying sensor clusters at key parts of power poles. As shown in patent documents CN111895965A and CN109373890A, calculating power pole deformation using IoT technology can theoretically achieve millimeter-level accuracy. However, in practical engineering applications, this method faces numerous challenges:

[0011] First, the number of sensor deployment points is limited, typically covering only a few key sections or nodes. For example, the existing scheme 2 (Wu Jun, Zheng Jiaxuan, Gan Yan, et al. "Simulation method of transmission line tower strain under thermo-coupling action" [J]. Journal of Harbin University of Science and Technology, 1-8) introduces measured strain data to correct the model, but its sensor deployment density is only 16 points, resulting in highly sparse monitoring data and difficulty in capturing the continuous deformation field of the entire pole. Second, the sensors, acquisition equipment, and their long-term installation and maintenance costs in the field are high, which is not conducive to large-scale promotion and application.

[0012] In summary, in existing technologies, data sparsity prevents physical monitoring from providing full-field deformation information, and the lack of model verification prevents numerical simulation from reflecting real working conditions; a significant gap exists between the two. The lack of a low-cost technique for acquiring high-density, high-precision three-dimensional deformation data has long been a problem of difficult finite element model calibration and low accuracy in structural safety assessment. Summary of the Invention

[0013] Therefore, it is necessary to address the issues of data sparsity and lack of model validation in existing tower deformation detection technologies by providing a method, device, and medium that combines images and point clouds for tower deformation feature extraction. This method solves the problem of insufficient density from a single data source by fusing high-density features from images with high-precision geometric information from point clouds, achieving low-cost, high-density, and high-precision extraction of tower deformation, and providing full-field measured data for finite element model validation.

[0014] Firstly, a method for extracting tower deformation features by combining images and point clouds includes:

[0015] The original image data and laser point cloud data of the transmission tower were acquired in multiple periods, and the data of each period were preprocessed to obtain the dense point cloud of the image and the laser point cloud of the tower for each period.

[0016] The dense point cloud of the image and the laser point cloud of the tower are registered and fused to obtain the tower fused point cloud of the corresponding period.

[0017] Cross-period registration is performed on the point cloud of multi-phase tower fusion to unify the point cloud of multi-phase tower fusion into the same coordinate space;

[0018] Stable feature points are extracted from multiple periods of original image data; the stable feature points have two-dimensional image coordinates in the original image and corresponding three-dimensional spatial coordinates in the dense point cloud of the image.

[0019] For any two images, based on the registration relationship between the dense point cloud of the image and the laser point cloud of the tower during the same period, the stable feature points of one image are transformed to the coordinate space of the fused point cloud of the tower in the other image, so as to obtain the three-dimensional position of the stable feature points in the fused point cloud of the tower in the other image.

[0020] Based on the three-dimensional position, the search area of ​​the stable feature point in another image is determined. Multiple candidate feature points corresponding to the stable feature point are searched within the search area, and a candidate matching set is constructed based on the stable feature point and the multiple candidate feature points.

[0021] The candidate matching set is purified by local geometric consistency constraints to obtain dense pairs of deformable feature points.

[0022] This invention acquires multiple periods of original images and laser point clouds, processes them separately to obtain dense point clouds (high density) from the images and laser point clouds (high precision) from the towers, and then registers and fuses the two to generate a fused point cloud for each period of the towers. This fused point cloud simultaneously possesses the high density characteristics of the dense point cloud from the images and the high-precision geometric information of the laser point cloud, providing a high-quality data foundation for subsequent deformation analysis and initially alleviating the contradiction between density and precision from a single data source.

[0023] By performing cross-period registration on the point clouds of multi-period towers, all point clouds from different periods are unified into the same coordinate space, eliminating the overall rigid displacement introduced by different acquisition times. This allows data from different periods to be compared under the same benchmark, providing the necessary conditions for accurately separating the tower's own deformation from its overall displacement.

[0024] The process of extracting, transforming, and determining the search region of stable feature points establishes a precise correspondence between the same feature point and point clouds in different time periods, providing accurate geometric constraints for subsequent matching.

[0025] By narrowing the search range for feature points from the global image to a search region determined by 3D location, the computational cost of matching is significantly reduced. Searching for multiple candidate feature points within this region preserves various matching possibilities. Furthermore, the candidate matching set is refined through local geometric consistency constraints, filtering out correct matches from multiple candidate feature points and eliminating false matches, ultimately obtaining dense and precise deformation feature point pairs. These feature point pairs cover the tower surface, capturing continuous deformations such as minute bends and torsions of the members, directly solving the technical bottleneck of "data sparsity."

[0026] In one embodiment, the cross-period registration of the multi-period tower fusion point cloud is achieved by using one period as a baseline and performing two-period registrations with the other periods separately. Each two-period registration uses a bottom-first adaptive registration method, specifically including:

[0027] The point cloud of the two towers to be registered is divided into multiple segments at predetermined height intervals along the height direction;

[0028] Starting from the bottom segment, proceed upwards segment by segment for evaluation: Calculate the first inter-segment deviation of the current segment. If the first inter-segment deviation is less than a preset deviation threshold, the current segment is determined to be undeformed and added to the registration reference set. Then, determine whether the registration reference set contains several consecutive segments. If not, continue processing the next segment upwards; if it does, stop evaluation. If the first inter-segment deviation of the current segment is not less than the preset deviation threshold, stop evaluation. This completes the initial construction of the registration reference set.

[0029] Based on the initially constructed registration reference set, the current transformation matrix is ​​calculated using a point cloud registration algorithm;

[0030] The current transformation matrix is ​​used to transform the point cloud of the two phases of tower fusion to the same coordinate space, and the second inter-segment deviation is calculated for the next segment above the registration reference set. It is then determined whether the second inter-segment deviation of the current segment falls within the normal error range.

[0031] If so, the current segment is added to the registration reference set, the transformation matrix is ​​updated, and the point cloud of the fusion of the two towers is transformed to the same coordinate space again, and the next segment is processed upwards; otherwise, it is determined that the current segment and the segments above it have deformed, the cross-period registration is terminated, and the point cloud of the fusion of the two towers is transformed to the same coordinate space according to the transformation matrix calculated last time.

[0032] Among them, "several" refers to at least three; the normal error range is dynamically constructed based on the inter-segment deviation of each segment in the registration reference set, and the inter-segment deviation includes at least the first inter-segment deviation.

[0033] This invention employs a bottom-up, segment-by-segment approach, slicing along the height direction and evaluating segments sequentially starting from the bottom. Segments with inter-segment deviations less than a preset threshold are included in a registration reference set. After calculating the transformation matrix based on this reference set, the upper segments are then inspected segment by segment. A normal error range is dynamically constructed using the statistical characteristics of the already registered segments to determine whether the current segment has undergone deformation. This avoids the problem of upper-level deformation contaminating the registration results in traditional overall registration. Through bottom-up, segment-by-segment inspection and dynamic error range construction, adaptive identification of rigid references is achieved, ensuring the accuracy and robustness of cross-period registration and providing a precise coordinate reference for subsequent deformation analysis.

[0034] In one embodiment, the point cloud of the two towers to be registered is divided into multiple segments along the height direction at predetermined height intervals, including:

[0035] For each phase of the tower fusion point cloud, it is divided along the height direction according to a predetermined height interval to obtain multiple segments; for the tower fusion point cloud in each segment, the pole structure axis or center line is fitted to obtain the axis line segment, and the coordinates of the two ends of the axis line segment in the height direction are recorded.

[0036] The segmentation method described above preserves the geometric features of the members while significantly reducing the data dimensionality, providing a regularized comparison object for calculating inter-segment deviations. Based on the coordinate records of the bottom and top endpoints of the segmented axis segments, segment-by-segment analysis in the height direction is realized, enabling the registration reference set to be constructed from the bottom up, starting from the undeformed bottom segment, avoiding contamination of the global registration by upper deformation. The first inter-segment deviation (horizontal distance from the top after bottom alignment) and the second inter-segment deviation (vector difference L2 norm) can be quantitatively calculated through the axis segments, providing clear numerical indicators for deformation determination. At the same time, the axis fitting is insensitive to local point cloud missingness, improving the robustness of the method in complex field environments.

[0037] In one embodiment, the first inter-segment deviation is calculated as follows: for the same segment, extract the point cloud subset belonging to the segment from the corresponding two-phase tower fusion point cloud; fit the axis line segments to the two point cloud subsets respectively to obtain two axis line segments; align the bottom endpoints of these two axis line segments, calculate the horizontal distance between the top endpoints of the two axis line segments at this time, and use the horizontal distance as the first inter-segment deviation of the segment.

[0038] The second segment deviation is calculated as follows: For the same segment, extract the point cloud subset belonging to the segment from the corresponding two-phase tower fusion point cloud; fit the axis line segments to the two point cloud subsets respectively to obtain two axis line segments; regard these two axis line segments as the first vector and the second vector respectively;

[0039] Calculate the vector difference between the first vector and the second vector, and obtain the L2 norm of the vector difference. Use the L2 norm as the second inter-segment deviation of the segment.

[0040] The calculation methods for the first and second inter-segment deviations described above serve different stages of cross-period registration: the first inter-segment deviation, calculated by the horizontal distance between the top and bottom after bottom alignment, eliminates overall displacement interference and quickly filters undeformed segments to construct a registration reference set; the second inter-segment deviation, calculated by the L2 norm of the vector difference, quantifies the three-dimensional comprehensive offset for accurate deformation determination. Both deviation calculations share the same set of segment and axis fitting results, balancing computational efficiency with the comprehensiveness of deformation detection.

[0041] In one embodiment, the normal error range is constructed using a statistical anomaly detection method, specifically including:

[0042] Based on the inter-segment deviations of each segment in the registration reference set, calculate its statistical central value and dispersion.

[0043] When the number of segments in the registration reference set is greater than the first number threshold and less than the second number threshold, the normal error range is constructed with the median as the center and the multiple of the absolute median difference as the dispersion.

[0044] When the number of segments in the registration reference set is greater than or equal to the second quantity threshold, the normal error range is constructed with the mean as the center and the multiple of the standard deviation as the dispersion.

[0045] This invention achieves dynamic adaptive construction of the normal error range: initially constructed based on the first inter-segment deviation, and continuously supplemented by the second inter-segment deviation during subsequent updates, thus avoiding the problem that a fixed threshold cannot adapt to changes in data distribution; at the same time, different statistical methods are used according to the sample size (absolute median difference method is suitable for small samples and has strong robustness; 3σ principle is suitable for large samples and has high efficiency), ensuring the scientific nature and accuracy of the error range construction and improving the reliability of deformation detection.

[0046] In one embodiment, stable feature point extraction is performed on multiple periods of original image data, specifically including:

[0047] Based on the established feature point matching relationship between images and the correspondence between feature points and 3D points in dense point clouds of images during the 3D reconstruction process of multi-view images, feature points that have the same name in at least O original images from different viewpoints are selected as stable feature points.

[0048] The stable feature point screening process described above in this invention ensures the repeatability and reliability of the extracted feature points and eliminates interference from temporary or non-rigid objects such as vegetation and noise. Stable feature points simultaneously possess the dual attributes of two-dimensional image coordinates and three-dimensional spatial coordinates in dense point clouds of images, providing a complete information foundation for subsequent inter-period conversion and matching.

[0049] In one embodiment, searching for multiple candidate feature points corresponding to the stable feature point within the search area, and constructing a candidate matching set based on the stable feature point and the corresponding multiple candidate feature points, specifically includes:

[0050] Based on the three-dimensional position of the stable feature point in the fused point cloud of another tower phase and the camera parameters of the other phase image, calculate the projection point of the stable feature point in the other phase image.

[0051] The search area is determined within a buffer of a predetermined pixel width, centered on the projection point.

[0052] Search for multiple candidate feature points in another image within the search area, and calculate the similarity between the stable feature point and each candidate feature point; the candidate feature points in the other image are stable feature points extracted from the other image.

[0053] A preset number of candidate feature points are selected from high to low similarity, or multiple candidate feature points with similarity exceeding a preset threshold are selected. The selected candidate feature points are paired with the corresponding stable feature points to generate one-to-many candidate matching pairs, thereby constructing a candidate matching set.

[0054] This invention narrows the matching search range of feature points from the global image to a search area defined near the projection point, significantly reducing the amount of matching computation and improving matching efficiency. At the same time, the introduction of projection point constraints ensures the geometric correctness of the matching. By searching for multiple candidate feature points within the search area and matching stable feature points with multiple candidate feature points, a high-quality candidate matching set is provided for subsequent purification through local geometric consistency constraints.

[0055] In one embodiment, the candidate matching set is purified by local geometric consistency constraints, specifically including:

[0056] For a candidate matching pair (p1, p2) in the candidate matching set, p1 is the 3D position of a stable feature point of one phase image after transformation in the fused point cloud of another phase, and p2 is the 3D position of a candidate feature point of another phase image in the fused point cloud of its corresponding phase.

[0057] Centered on p1, select S neighboring points within a predetermined radius to form its local neighborhood. For any one of these neighboring points q1, if q1 is a stable feature point in the candidate matching set, then obtain the corresponding candidate feature point in the candidate matching set, denoted as q2; otherwise, skip the neighboring point q1.

[0058] Verify whether the distance consistency condition is satisfied: distance(p1,q1) ≈ distance(p2,q2), where distance() represents the Euclidean distance between two points in space;

[0059] The RANSAC algorithm is used to iteratively filter candidate matching pairs in the candidate matching set based on the distance consistency condition, eliminating mismatches that do not meet the distance consistency condition, and obtaining precisely matched feature point pairs.

[0060] This invention refines the candidate matching set by utilizing the similarity of the local neighborhood geometric structure, effectively eliminating mismatched points that cannot be distinguished by similarity alone; the distance consistency condition reflects the physical characteristic that the relative positions within the local neighborhood remain unchanged after deformation, ensuring the geometric correctness of the fine matching, and finally obtaining high-density, high-precision deformation feature point pairs, directly solving the technical bottleneck of data sparsity.

[0061] In one embodiment, the method further includes the step of generating a three-dimensional deformation field of the tower based on the deformation feature point pair, specifically including:

[0062] Each pair of deformation feature points is treated as a deformation vector to obtain a discrete deformation vector field;

[0063] The discrete deformation vector field is interpolated by radial basis function interpolation or Gaussian process regression to generate a continuous and smooth three-dimensional deformation field.

[0064] This invention transforms discrete pairs of deformation feature points into a continuous, smooth three-dimensional deformation field, filling the information gaps between discrete points and providing full-field deformation data covering the entire surface of the tower. The continuous deformation field can be directly input into finite element analysis software as displacement boundary conditions to verify and calibrate numerical simulation results, directly solving the technical bottleneck of missing model verification.

[0065] Secondly, the present invention also provides an electronic device, including a memory, a processor, and a computer program / instructions stored in the memory, wherein the processor executes the computer program / instructions to implement the tower deformation feature extraction method as described above.

[0066] Thirdly, the present invention also provides a computer-readable storage medium having a computer program / instruction stored thereon, which, when executed by a processor, implements the tower deformation feature extraction method as described above. Attached Figure Description

[0067] Figure 1 This is a flowchart of the tower deformation feature extraction method in an embodiment of the present invention;

[0068] Figure 2 This is a flowchart of the bottom-first adaptive registration method in an embodiment of the present invention;

[0069] Figure 3 This is a schematic diagram of the tower structure in an embodiment of the present invention;

[0070] Figure 4 This is a schematic diagram of the bottom-priority adaptive registration method in an embodiment of the present invention;

[0071] Figure 5 This is a block diagram of the electronic device structure in an embodiment of the present invention. Detailed Implementation

[0072] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0073] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0074] This invention provides a method for extracting tower deformation features by combining images and point clouds, such as... Figure 1 As shown, it includes the following steps:

[0075] Step 1: Acquire raw image data and laser point cloud data of the transmission tower in multiple periods, and preprocess the data of each period to obtain the dense point cloud of the image and the laser point cloud of the tower for each period.

[0076] In this embodiment, the acquisition of raw image data and laser point cloud data of the transmission tower over multiple periods is specifically implemented as follows: A drone equipped with a camera and lidar is used to collect data from the target tower along a planned flight path over multiple periods. Each period of data collection includes simultaneously acquired raw images and laser point clouds. Preprocessing is performed on the data from each period:

[0077] For the original image data: dense point clouds are extracted using aerial triangulation. Specifically, high-density 3D point clouds are reconstructed from multi-view images using SFM (Structure of Motion) and MVS (Multi-view Stereo) methods, denoted as image dense point clouds.

[0078] For laser point cloud data: outlier removal, cloth simulation filtering, and deep learning methods are used to extract the laser point cloud of the tower after vegetation removal.

[0079] Step 2: Register and fuse the dense point cloud of the image from the same period with the laser point cloud of the tower to obtain the fused point cloud of the tower for the corresponding period.

[0080] In this embodiment, the dense point cloud of the image and the laser point cloud of the tower are registered and fused, specifically including:

[0081] First, by combining the IMU (Inertial Measurement Unit) and GNSS (Global Navigation Satellite System) data synchronously recorded during UAV acquisition, initial position and attitude information is provided for dense point clouds in images and laser point clouds on towers, serving as a priori reference for registration.

[0082] Then, a point cloud registration algorithm (such as the Iterative Closest Point (ICP) algorithm) is used to register the dense point cloud of the image with the laser point cloud of the tower. During the registration process, anchor points are flexibly selected according to the data acquisition scenario.

[0083] If the laser point cloud of the tower covers the ground area, the ground points are extracted first as global anchor points. The consistency of the ground points in different data sources is used to improve the robustness and accuracy of registration.

[0084] If dense vegetation renders ground points unusable, then only the exposed tower body or tower base point cloud is used as the registration reference, combined with the geometric information provided by the dense point cloud in the image for auxiliary registration.

[0085] The above registration transforms the dense point cloud of the image and the laser point cloud of the tower to a unified coordinate system. After registration, the two point clouds are fused to generate the fused point cloud of the tower for this phase. This fused point cloud of the tower possesses both the high-density characteristics of the dense point cloud of the image and the high-precision geometric information of the laser point cloud of the tower.

[0086] Step 3: Perform cross-period registration on the point cloud of the multi-period tower fusion to unify the point cloud of the multi-period tower fusion into the same coordinate space.

[0087] In this embodiment, the cross-period registration of multi-period tower fusion point clouds is achieved by using one period as a reference and performing two-period registration with each of the other periods separately. Specifically, using the first-period tower fusion point cloud as a reference, the second, third, and subsequent periods' tower fusion point clouds are sequentially registered with the first period, so that the tower fusion point clouds of all periods are unified to the same coordinate space.

[0088] like Figure 3 As shown, a typical tower structure includes tower legs, tower body, and tower head. Under ice loads, wind loads, etc., the tower head and tower body are prone to non-uniform deformation, while the tower legs, being fixed to the ground, experience minimal deformation. Therefore, this invention, based on the assumption that the bottom is not easily deformed, employs an adaptive registration method that prioritizes the bottom in each two-stage registration design. This method utilizes the characteristic of the tower bottom being less prone to deformation, such as the fixed ground around the tower legs, to avoid interference from upper deformation during overall registration. Figure 2 As shown, this method proceeds in stages from bottom to top, specifically including an initialization stage and a formal processing stage.

[0089] Step 3.1: Slicing.

[0090] Let the point clouds of the two towers to be registered be x. i and x j Where i ≠ j. With x i For example, to form a point cloud x i The M points are denoted as Each point It has three coordinate values ​​in a spatial rectangular coordinate system, with z representing its height. According to a predetermined height interval Δh (e.g., 5cm), it will... The points in the array are divided according to the height z of each point. Let d be the nth segment that satisfies (n-1)·Δh≤z≤n·Δh. in For x j Using the same processing method, segment d is obtained. jn .

[0091] For each segment d in For points within the segment, fit the structural axis (for slender members) or centerline (for square towers) to obtain an axial line segment representing the geometric center of that segment. The endpoint with the smaller z-value on this axial line segment is denoted as s. in The endpoint with the larger z value is denoted as e. in For x j Segmentation d jn Perform the same fitting process to obtain the axis line segment (s) jn ,e jn ).

[0092] Step 3.2: Initialization phase.

[0093] Step 3.2.1: Calculate the first segment deviation:

[0094] For one of the point cloud phases x i Segmentation d in The fitted (s) in , e in ) and another point cloud x j Segmentation d jn The fitted (s) jn , e jn ), through horizontal displacement, make s in With s jn Overlap, at this time e in With e jn The horizontal distance between them is the first inter-segment deviation of this segment, denoted as c. n .

[0095] Step 3.2.2: Initially construct the registration reference set.

[0096] Starting from the bottom segment, continuously evaluate each segment upwards: calculate the first inter-segment deviation c of the current segment. n If the deviation between the first segment is c n If the deviation is less than the preset deviation threshold, the current segment is considered to have not undergone deformation, and the first segment deviation c is calculated. n The two segments d used in With d jnAdd the data to the registration reference set and determine if the registration reference set contains at least 5 consecutive segments. If it does not contain at least 5 consecutive segments, continue processing the next segment; if it contains at least 5 consecutive segments, stop the process; if the deviation between the first segments is c... n If the deviation is not less than the preset deviation threshold, the judgment will stop.

[0097] The preset deviation threshold is set based on one or more of the following factors: material elastic modulus, the range of elastic deformation under the estimated maximum wind load or ice load, and the noise level of point cloud measurement. For example, in a simulation, when the tower undergoes irreversible deformation, the horizontal deviation of the central axis at the bottom of the irreversibly deformed part is 5 centimeters with a height difference of 1 meter. Therefore, in this method, for a Δh of 1 meter, the preset deviation threshold is set to 5 centimeters.

[0098] Since the preset deviation threshold is a relatively large value, it is mainly used at the beginning when there is no data accumulation to identify data problems. Based on the simulation and experimental assumptions that the bottom is not easily deformed, and assuming that the fitting does not fail, the inter-segment deviation c1 of the bottom segment can be considered very small.

[0099] The registration reference set includes point cloud x i Multiple consecutive segments, including point cloud x j A series of consecutive segments.

[0100] Step 3.3: Formal processing stage.

[0101] Step 3.3.1: Based on the registration reference set, calculate the current transformation matrix T using the Iterative Closest Point (ICP) algorithm, and then use the current transformation matrix T to transform the point cloud x i With point cloud x j Transform to the same coordinate space.

[0102] Step 3.3.2: Determine the initial processing segment.

[0103] Suppose that the registration reference set already contains consecutive segments d1, d2, ..., d2 starting from the bottom. k (including d) ik With d jk Then, the first segment to be processed in the formal processing stage is the next segment above the registration reference set, i.e., d. k+1 .

[0104] Step 3.3.3: Segment-by-segment detection and judgment.

[0105] For the current segment d to be processed k+1 Perform the following operations:

[0106] Calculate the deviation between the second segment: [Calculate the point cloud x] i Segmentation di,k+1 The fitted (s) i,k+1 ,e i,k+1 Treating x as vector 1, the point cloud x j Segmentation d j,k+1 The fitted (s) j,k+1 ,e j,k+1 Treating the first two vectors as vector 2, calculate the L2 norm of the difference between them, which is taken as the second inter-segment deviation of the segment, denoted as c. k+1 In another implementation, the first inter-segment deviation and the second inter-segment deviation are calculated in the same way, such as both using the endpoint alignment method of the first inter-segment deviation, or both using the L2 norm method of the second inter-segment deviation.

[0107] Determine if deformation has occurred: The current registration reference set contains segments whose first inter-segment deviations are {c1,c2,c3,…,c…}. k}. Determine c k+1 Is it within the normal error range? If c k+1 Within the normal error range, then the segments d will be divided. k+1 Add the registration reference set, then return to step 3.3.1, that is, recalculate the transformation matrix T based on the updated registration reference set, and use this transformation matrix to transform the point cloud x i With point cloud x j Transform to the same coordinate space, increment k by 1, and continue processing the next segment d to be processed. k+2 .

[0108] If c k+1 If it is outside the normal error range, then segment d is determined. k+1 If deformation occurs in segments 1 and 2, inter-period registration is terminated.

[0109] The normal error range is constructed based on the inter-segment deviations of each segment in the registration reference set, using a statistical anomaly detection method. c is initially determined during the formal processing stage. k+1 When determining whether the error is within the normal error range, this range is initially constructed based on the first inter-segment deviation of each segment in the registration reference set. As the formal processing stage progresses, the second inter-segment deviation is continuously added to the registration reference set, and the normal error range is dynamically constructed based on the latest registration reference set. The specific construction process is as follows:

[0110] When the number of segments in the registration reference set is greater than the first threshold (e.g., 5) and less than the second threshold (e.g., 20), anomaly detection is performed using the median absolute difference (MAD). The normal error range is constructed with the median as the center and the multiple of the absolute median absolute difference as the dispersion.

[0111] When the number of segments in the registration reference set is greater than or equal to the second quantity threshold, the 3σ principle is used for anomaly detection, with the mean as the center and multiples of the standard deviation as the dispersion, to construct the normal error range.

[0112] like Figure 4 As shown, the bottom-first adaptive registration method includes an initialization phase and a formal processing phase. The initialization phase starts from the bottom segment and continuously filters upwards, including segments with an inter-segment deviation less than a preset deviation threshold into the registration reference set, until at least five consecutive segments are obtained or a segment that does not meet the condition is encountered. The formal processing phase calculates the transformation matrix based on the registration reference set, transforming the two point clouds to the same coordinate space. Then, it calculates the second inter-segment deviation for the next segment above the registration reference set and determines whether it falls within the normal error range. If it does, the segment is added to the registration reference set, the transformation matrix is ​​updated, and processing continues upwards; otherwise, registration is terminated, and the coordinate space is unified using the last calculated transformation matrix.

[0113] Step 3.3.4: Coordinate unification

[0114] After termination, the transformation matrix calculated last before termination is used to unify the point clouds x from the two periods. i With x j The coordinate space is then used. For multiple periods of data, the above process is repeated, using the first period's fused tower point cloud as the reference, and registering each of the other periods with the first period in turn, so that the fused tower point clouds of all periods are ultimately unified into the same coordinate space.

[0115] Step 4: Extract stable feature points from the original image data from multiple periods.

[0116] The extraction of stable feature points aims to obtain repeatable feature points on tower structures or fixed textures, providing reliable input for subsequent inter-period matching. This step does not rely on the point cloud after inter-period registration, but directly filters based on the original image data and the correspondences established during multi-view image 3D reconstruction. The specific extraction process includes:

[0117] Step 4.1: 3D reconstruction of multi-view images.

[0118] For each data set, the original image data was processed using multi-view image 3D reconstruction technology during the data preprocessing stage. Specifically, Structure of Motion (SFM) and Multi-View Stereo Vision (MVS) methods were employed to reconstruct the 3D structure of the scene from multiple original images from different perspectives, generating a dense point cloud. During this reconstruction process, the following correspondences were simultaneously established:

[0119] Image feature point matching relationship: the matching relationship of the same feature points between original images from different viewpoints;

[0120] Two-dimensional to three-dimensional correspondence: The correspondence between the two-dimensional image coordinates of each feature point in the original image and its three-dimensional spatial coordinates in the dense point cloud of the image.

[0121] Step 4.2: Screening of stable feature points.

[0122] Based on the established correspondence, feature points that meet the stable observation conditions are selected from the original image data as stable feature points. The stable observation conditions are: the feature point has corresponding points in at least O original images from different viewpoints, and the corresponding points correspond to the same three-dimensional point in the dense point cloud of the image, where O is a preset quantity threshold (e.g., O≥3).

[0123] The specific screening process is as follows:

[0124] For each 3D point in the dense point cloud of the image, based on the established 2D-3D correspondence, the projection point (i.e., image point) of the 3D point in the original image from each viewpoint is obtained; the number of images containing the projection of the 3D point is counted; if this number is greater than or equal to a preset threshold O, the image point corresponding to the 3D point is taken as a stable feature point; the above process is repeated to traverse all 3D points in the dense point cloud of the image to obtain all stable feature points. This selection condition ensures that the extracted stable feature points are located on tower structures or fixed textures, rather than on temporary or non-rigid objects such as vegetation or noise, because these unstable objects are difficult to stably reproduce under multiple different viewpoints.

[0125] Step 4.3: Dual attributes of stable feature points.

[0126] The stable feature points obtained through the above screening process possess both of the following attributes:

[0127] Two-dimensional image coordinates: Stable feature points have corresponding two-dimensional image coordinates in the original image, meaning that the feature point can be located in the original image from multiple different viewpoints.

[0128] Three-dimensional spatial coordinates: Stable feature points have corresponding three-dimensional spatial coordinates in the dense point cloud of the image, that is, the precise three-dimensional position obtained by SFM / MVS reconstruction.

[0129] Step 4.4: Repeated extraction of data from multiple periods.

[0130] Repeat the above process for each period of data (period 1, period 2, ...) to extract the corresponding stable feature point set for each period. Since the coordinate space of the data from multiple periods has been unified through step 3, the stable feature points extracted from each period are all located in the same coordinate system, providing a direct correspondence basis for subsequent cross-period matching.

[0131] Step 5: For any two images, based on the registration relationship between the dense point cloud of the image and the laser point cloud of the tower in the same period, the stable feature points of one image are transformed to the coordinate space of the fused point cloud of the tower in the other image, so as to obtain the three-dimensional position of the stable feature points in the fused point cloud of the tower in the other image.

[0132] This step aims to utilize the registration relationships from the same period and the coordinate space unified by cross-period registration to map stable feature points from one period's image to the coordinate space of the fused point cloud of another period's towers, thereby obtaining their 3D positions within the fused point cloud of the other period's towers. Specifically, it includes the following process:

[0133] Let the two images to be matched be the first image and the second image, where the first image corresponds to the first phase of the tower fusion point cloud, and the second image corresponds to the second phase of the tower fusion point cloud. Using the stable feature points of the first image as a reference, find their corresponding points in the second image.

[0134] For a stable feature point in the first phase image, since the feature point has been confirmed to have corresponding three-dimensional spatial coordinates in the dense point cloud of the image in step 4, the three-dimensional position of the feature point in the dense point cloud of the first phase image can be directly obtained, denoted as b1.

[0135] Since step 2 has already registered and fused the dense point cloud of the image in the first phase with the laser point cloud of the tower, a registration relationship has been established between the dense point cloud of the image and the fused point cloud of the tower in the first phase. Based on this registration relationship, the three-dimensional coordinates b1 of the stable feature point in the dense point cloud of the image are transformed to the fused point cloud of the tower in the first phase, so as to obtain its three-dimensional coordinates b1′ in the fused point cloud of the tower in the first phase.

[0136] Step 3 has unified the point clouds of the first and second phase tower fusion into the same coordinate space. Therefore, b1′ and the point cloud of the second phase tower fusion are located in the same coordinate system. Based on this unified coordinate space, b1′ is taken as the corresponding three-dimensional position of the stable feature point in the point cloud of the second phase tower fusion, denoted as p1. This p1 is the three-dimensional position of the stable feature point in the point cloud of the second phase tower fusion.

[0137] For each stable feature point in the first phase image, repeat the above steps to obtain the 3D position of each stable feature point in the second phase tower fusion point cloud. For the third phase and subsequent phase images, use the same method to transform the stable feature points of the first phase image to the coordinate space of the corresponding phase tower fusion point cloud.

[0138] Step 6: Determine the search region of the stable feature point in another image based on the 3D position, search for multiple candidate feature points corresponding to the stable feature point within the search region, and construct a candidate matching set based on the stable feature point and the corresponding multiple candidate feature points.

[0139] Based on the 3D position obtained in step 5, this step determines the search region in another image and searches for multiple candidate feature points within that search region. Through feature descriptor similarity matching, a candidate matching set containing one-to-many candidate matches is generated.

[0140] For a stable feature point in the first phase image, its 3D position in the fused point cloud of the second phase towers is denoted as p1. Based on p1 and the camera parameters (including intrinsic and extrinsic parameters) of the second phase image, the projection point of this stable feature point in the second phase image is calculated. Using the calculated projection point as the center, a search region is determined within a buffer of a predetermined pixel width (e.g., 2-5 pixels). This search region is the search range for subsequent matching; it is much smaller than the entire image, thus significantly improving matching efficiency.

[0141] Within the defined search area, all candidate feature points in the second phase image are searched. Candidate feature points refer to stable feature points extracted in step 4 in the second phase image, i.e., points that have two-dimensional image coordinates and corresponding three-dimensional coordinates in the dense point cloud of the second phase image.

[0142] Extract feature descriptors from stable feature points and candidate feature points within the search region, and calculate the similarity between them. Feature descriptors are used to quantify the texture information of the local neighborhood of feature points. Deep learning feature descriptors such as SuperPoint and LoFTR, or traditional feature descriptors such as SIFT and ORB, can be used. Similarity measures can be Euclidean distance (smaller values ​​indicate greater similarity) or cosine similarity (larger values ​​indicate greater similarity).

[0143] Based on the similarity calculation results, a preset number of candidate feature points (e.g., the first 3) are selected from high to low, or candidate feature points with similarity exceeding a preset threshold are selected. The selected candidate feature points are paired with the corresponding stable feature points. Each stable feature point corresponds to multiple candidate matching points, forming a one-to-many candidate matching pair.

[0144] A candidate matching set is generated by summing up all one-to-many candidate matching pairs of stable feature points. This candidate matching set preserves multiple matching possibilities, providing sufficient candidates for subsequent fine matching.

[0145] Repeat the above steps for the first and second periods, the first and third periods, and so on, to obtain the candidate matching set for all periods relative to the first period.

[0146] Step 7: Purify the candidate matching set by applying local geometric consistency constraints to obtain dense pairs of deformable feature points.

[0147] This step utilizes the similarity of local neighborhood geometric structures to refine the candidate matching set generated in step 6, eliminating mismatches and obtaining accurate deformation feature point pairs. Let the candidate matching pairs in the candidate matching set be (p1, p2), where: p1 is the 3D position of a stable feature point from one phase image after transformation in the fused point cloud of another phase; p2 is the 3D position of a candidate feature point from another phase image in its corresponding phase fused point cloud of the tower.

[0148] Centered on p1, select S neighboring points within a predetermined radius r (e.g., 1 meter) to form a local neighborhood point set of p1. This neighborhood point set reflects the local geometric structure around p1.

[0149] For any neighboring point q1 in the local neighborhood point set, if it is a stable feature point in the candidate matching set generated in step 6, then obtain its corresponding candidate feature point in the candidate matching set, denoted as q2. q2 is the candidate matching point obtained by matching q1 with the feature descriptor in step 6.

[0150] Based on the principle of local invariance, the relative positional relationships within the local neighborhood point set should remain essentially unchanged after deformation occurs. Therefore, for the correct matching pair (p1, p2) and (q1, q2), the distance consistency condition should be satisfied:

[0151] distance(p1,q1) ≈ distance(p2,q2);

[0152] Here, distance() represents the Euclidean distance between two points in space. That is, the distance from p1 to q1 should be approximately equal to the distance from p2 to q2.

[0153] The RANSAC (Random Sample Consensus) algorithm is used to iteratively filter the candidate matching set based on the distance consistency condition. The specific process is as follows:

[0154] Randomly select a set of matching pairs from the candidate matching set as the inlier hypothesis; calculate the allowable error threshold based on the hypothesis; traverse all candidate matching pairs to verify whether they meet the distance consistency condition (i.e., within the error threshold range); count the number of inliers that meet the condition; repeat the above steps multiple times and select the set with the most inliers as the final model; remove matching pairs that do not meet the model (i.e., do not meet the distance consistency condition).

[0155] By using RANSAC iterative filtering to eliminate mismatches, a unique correct matching point is retained for each stable feature point, resulting in precisely matched deformation feature point pairs.

[0156] After the above purification, each precisely matched deformation feature point pair (p1, p2) represents the correspondence of the same physical point in the two phases of data. Since the coordinate space of the fused point cloud of the two phases of towers has been unified through step 3, the coordinate difference between p1 and p2 reflects the deformation vector that occurred at that point. Therefore, these precisely matched deformation feature point pairs are the dense deformation feature point pairs output in this step.

[0157] Repeat steps 5 to 7 above for phase 1 and phase 2, phase 1 and phase 3, ... to obtain all dense deformation feature point pairs relative to phase 1 for all phases.

[0158] Step 8: Generate the three-dimensional deformation field of the tower based on the deformation feature point pairs.

[0159] This step aims to transform the high-density deformation feature point pairs obtained in step 7 into a continuous three-dimensional deformation field covering the entire surface of the tower, providing input data for structural safety assessment and finite element verification. Specifically, it includes the following processes:

[0160] Each precisely matched deformation feature point pair (p1, p2) obtained in step 7 is treated as a deformation vector, where p1 belongs to the first-phase tower fusion point cloud and p2 belongs to the second-phase tower fusion point cloud. Since the coordinate space of the two-phase tower fusion point clouds has been unified through step 3, the coordinate difference between p1 and p2 is the deformation vector of that point, calculated as: v = p2 - p1, where v represents the three-dimensional deformation vector of that point, containing components in three directions (Δx, Δy, Δz). This operation is performed on all deformation feature point pairs to obtain a discrete deformation vector field covering the tower surface. This discrete field consists of a series of deformation vectors with three-dimensional spatial coordinates, reflecting the magnitude and direction of the deformation occurring at the corresponding positions of the tower.

[0161] Discrete deformation vector fields only contain deformation information at the locations of feature points, while other areas on the tower surface lack data. To obtain continuous deformation information covering the entire tower surface, interpolation processing of the discrete vector field is required.

[0162] In this embodiment, radial basis function (RBF) interpolation or Gaussian process regression (GPR) method is used to interpolate the discrete deformation vector field to generate a continuous and smooth three-dimensional deformation field.

[0163] Radial basis function (RBF) interpolation is a distance-based interpolation method that predicts deformation at arbitrary spatial locations by representing the deformation values ​​at discrete points as a linear combination of radial basis functions. Commonly used radial basis functions include Gaussian kernels, multiple quadratic functions, and inverse multiple quadratic functions. RBF interpolation ensures that the generated deformation field passes precisely through known points and transitions smoothly at unknown points, making it suitable for situations with complex deformation field distributions.

[0164] Gaussian process regression (GPR) is a probability-based nonparametric regression method that treats the deformation field as a Gaussian process prior, predicting the deformation value and its uncertainty at unknown points using observed values ​​at known points. GPR can simultaneously output the expectation and variance of the deformation value, making it suitable for quantifying the uncertainty of the deformation field. This method is particularly suitable for situations with few discrete deformation points or high data noise.

[0165] The three components (Δx, Δy, Δz) of the discrete deformation vector field are interpolated using the above method to obtain a continuous three-dimensional deformation field covering the entire surface of the tower, denoted as D(x, y, z), which represents the three-dimensional deformation vector at any position (x, y, z) on the tower.

[0166] The generated three-dimensional deformation field can be directly used as a displacement boundary condition and input into finite element analysis software (such as ANSYS) to inversely calculate the distribution of external loads causing the deformation. Specifically: the three-dimensional deformation field is applied as a nodal displacement constraint to the finite element model of the tower; the equivalent loads acting on the tower (such as the distribution and magnitude of ice loads and wind loads) are inversely calculated through finite element analysis; the inverted loads are compared with the design loads to verify the accuracy of the numerical simulation and to calibrate the simulation model with measured data.

[0167] Through the above process, the dense three-dimensional deformation field generated by the method of the present invention makes up for the lack of experimental verification in traditional numerical simulation, solves the technical bottleneck of missing model verification, and provides reliable data support for the safety assessment of tower structures.

[0168] This invention also provides an electronic device, such as... Figure 5 As shown, the electronic device includes: a memory, a processor, and a computer program / instructions stored in the memory, wherein the processor executes the computer program / instructions to implement the tower deformation feature extraction method mentioned above.

[0169] Electronic devices include a processor that performs various appropriate operations and processes based on programs and / or data stored in read-only memory (ROM) or loaded from storage into random access memory (RAM). The processor can be a multi-core processor or may contain multiple processors. In some embodiments, the processor may include a general-purpose main processor and one or more specialized coprocessors, such as a central processing unit, graphics processing unit (GPU), neural network processor (NPU), digital signal processor (DSP), etc. Various programs and data required for device operation are also stored in RAM. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0170] The processor and memory described above are used together to execute programs / instructions stored in the memory. When the program / instructions are executed by the computer, they can implement the methods, steps, or functions described in the above embodiments.

[0171] This invention also provides a computer-readable storage medium storing a computer program / instructions thereon, which, when executed by a processor, implements the tower deformation feature extraction method mentioned above.

[0172] Readable storage media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0173] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0174] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for extracting tower deformation features by combining images and point clouds, characterized in that, include: The original image data and laser point cloud data of the transmission tower were acquired in multiple periods, and the data of each period were preprocessed to obtain the dense point cloud of the image and the laser point cloud of the tower for each period. The dense point cloud of the image and the laser point cloud of the tower are registered and fused to obtain the tower fused point cloud of the corresponding period. Cross-period registration is performed on the point cloud of multi-phase tower fusion to unify the point cloud of multi-phase tower fusion into the same coordinate space; Stable feature point extraction from multiple periods of raw image data; The stable feature points have two-dimensional image coordinates in the original image and corresponding three-dimensional spatial coordinates in the dense point cloud of the image; For any two images, based on the registration relationship between the dense point cloud of the image and the laser point cloud of the tower during the same period, the stable feature points of one image are transformed to the coordinate space of the fused point cloud of the tower in the other image, so as to obtain the three-dimensional position of the stable feature points in the fused point cloud of the tower in the other image. Based on the three-dimensional position, the search area of ​​the stable feature point in another image is determined. Multiple candidate feature points corresponding to the stable feature point are searched within the search area, and a candidate matching set is constructed based on the stable feature point and the multiple candidate feature points. The candidate matching set is purified by local geometric consistency constraints to obtain dense pairs of deformable feature points.

2. The method of claim 1, wherein the method of extracting tower shape deformation features by combining images and point clouds is characterized by, The cross-period registration of the multi-period tower fusion point cloud is achieved by using one period as a baseline and performing two-period registrations with the other periods separately. Each two-period registration uses a bottom-first adaptive registration method, specifically including: The point cloud of the two towers to be registered is divided into multiple segments at predetermined height intervals along the height direction; Starting from the bottom segment, proceed upwards segment by segment for evaluation: Calculate the first inter-segment deviation of the current segment. If the first inter-segment deviation is less than a preset deviation threshold, the current segment is determined to be undeformed and added to the registration reference set. Then, determine whether the registration reference set contains several consecutive segments. If not, continue processing the next segment upwards; if it does, stop evaluation. If the first inter-segment deviation of the current segment is not less than the preset deviation threshold, stop evaluation. This completes the initial construction of the registration reference set. Based on the initially constructed registration reference set, the current transformation matrix is ​​calculated using a point cloud registration algorithm; The current transformation matrix is ​​used to transform the point cloud of the two phases of tower fusion to the same coordinate space, and the second inter-segment deviation is calculated for the next segment above the registration reference set. It is then determined whether the second inter-segment deviation of the current segment falls within the normal error range. If so, the current segment is added to the registration reference set, the transformation matrix is ​​updated, and the point cloud of the fusion of the two towers is transformed to the same coordinate space again, and the next segment is processed upwards; otherwise, it is determined that the current segment and the segments above it have deformed, the cross-period registration is terminated, and the point cloud of the fusion of the two towers is transformed to the same coordinate space according to the transformation matrix calculated last time. Among them, "several" refers to at least three; the normal error range is dynamically constructed based on the inter-segment deviation of each segment in the registration reference set, and the inter-segment deviation includes at least the first inter-segment deviation.

3. The method for extracting tower deformation features by combining images and point clouds according to claim 2, characterized in that, The point cloud of the two towers to be registered is divided into multiple segments at predetermined height intervals along the height direction, including: For each phase of the tower fusion point cloud, it is divided along the height direction according to a predetermined height interval to obtain multiple segments; for the tower fusion point cloud in each segment, the pole structure axis or center line is fitted to obtain the axis line segment, and the coordinates of the two ends of the axis line segment in the height direction are recorded.

4. The method for extracting tower deformation features by combining images and point clouds according to claim 2, characterized in that, The first segment deviation is calculated as follows: for the same segment, extract the point cloud subset belonging to the segment from the corresponding two-phase tower fusion point cloud; fit the axis line segments to the two point cloud subsets respectively to obtain two axis line segments; Align the bottom endpoints of these two axis segments, calculate the horizontal distance between the top endpoints of the two axis segments, and use this horizontal distance as the first inter-segment deviation of the segment. The second segment deviation is calculated as follows: for the same segment, extract the point cloud subset belonging to the segment from the corresponding two-phase tower fusion point cloud; fit the axis line segments to the two point cloud subsets respectively to obtain two axis line segments; Consider these two line segments as the first vector and the second vector, respectively; Calculate the vector difference between the first vector and the second vector, and obtain the L2 norm of the vector difference. Use the L2 norm as the second inter-segment deviation of the segment.

5. The method for extracting tower deformation features by combining images and point clouds according to claim 2, characterized in that, The normal error range is constructed using a statistical anomaly detection method, specifically including: Based on the inter-segment deviations of each segment in the registration reference set, calculate its statistical central value and dispersion. When the number of segments in the registration reference set is greater than the first number threshold and less than the second number threshold, the normal error range is constructed with the median as the center and the multiple of the absolute median difference as the dispersion. When the number of segments in the registration reference set is greater than or equal to the second quantity threshold, the normal error range is constructed with the mean as the center and the multiple of the standard deviation as the dispersion.

6. The method for extracting tower deformation features by combining images and point clouds according to claim 1, characterized in that, Searching for multiple candidate feature points corresponding to the stable feature point within the search area, and constructing a candidate matching set based on the stable feature point and the corresponding multiple candidate feature points, specifically including: Based on the three-dimensional position of the stable feature point in the fused point cloud of another tower phase and the camera parameters of the other phase image, calculate the projection point of the stable feature point in the other phase image. The search area is determined within a buffer of a predetermined pixel width, centered on the projection point. Search for multiple candidate feature points in another image within the search area, and calculate the similarity between the stable feature point and each candidate feature point; the candidate feature points in the other image are stable feature points extracted from the other image. A preset number of candidate feature points are selected from high to low similarity, or multiple candidate feature points with similarity exceeding a preset threshold are selected. The selected candidate feature points are paired with the corresponding stable feature points to generate one-to-many candidate matching pairs, thereby constructing a candidate matching set.

7. The method for extracting tower deformation features by combining images and point clouds according to claim 1, characterized in that, The candidate matching set is purified by local geometric consistency constraints, specifically including: For a candidate matching pair (p1, p2) in the candidate matching set, p1 is the 3D position of a stable feature point of one phase image after transformation in the fused point cloud of another phase, and p2 is the 3D position of a candidate feature point of another phase image in the fused point cloud of its corresponding phase. Centered on p1, select S neighboring points within a predetermined radius to form its local neighborhood. For any one of these neighboring points q1, if q1 is a stable feature point in the candidate matching set, then obtain the corresponding candidate feature point in the candidate matching set, denoted as q2; otherwise, skip the neighboring point q1. Verify whether the distance consistency condition is satisfied: distance(p1,q1) ≈ distance(p2,q2), where distance() represents the Euclidean distance between two points in space; The RANSAC algorithm is used to iteratively filter candidate matching pairs in the candidate matching set based on the distance consistency condition, eliminating mismatches that do not meet the distance consistency condition, and obtaining precisely matched feature point pairs.

8. The method for extracting tower deformation features by combining images and point clouds according to any one of claims 1 to 7, characterized in that, It also includes the step of generating a three-dimensional deformation field of the tower based on the deformation feature point pairs, specifically including: Each pair of deformation feature points is treated as a deformation vector to obtain a discrete deformation vector field; The discrete deformation vector field is interpolated by radial basis function interpolation or Gaussian process regression to generate a continuous and smooth three-dimensional deformation field.

9. An electronic device comprising a memory, a processor, and a computer program / instructions stored in the memory, characterized in that, The processor executes the computer program / instructions to implement the tower deformation feature extraction method as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the tower deformation feature extraction method as described in any one of claims 1 to 8.

Images