Deformation monitoring method and device based on improved ICP point cloud registration
By improving the ICP point cloud registration method and combining it with 3D laser scanning and feature extraction algorithms, efficient and accurate deformation monitoring of GIS equipment has been achieved, solving the problems of high monitoring cost and low efficiency in existing technologies. This method is suitable for low-cost, large-scale deformation monitoring of power grid equipment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-25
- Publication Date
- 2026-03-17
AI Technical Summary
Existing technologies for monitoring deformation of power grid equipment suffer from problems such as high equipment costs, low monitoring efficiency, and inability to monitor for extended periods. In particular, they cannot achieve large-scale, low-cost, and high-precision monitoring of deformation of GIS equipment.
A deformation monitoring method based on improved ICP point cloud registration is adopted. The three-dimensional point cloud data of GIS equipment is acquired by a 3D laser scanner. Combined with voxelization filtering, FPFH feature extraction, RANSAC coarse registration and improved ICP fine registration algorithm, the point cloud data is efficiently and accurately registered to identify the deformed parts of the equipment.
It enables efficient, accurate, and long-term deformation monitoring of GIS equipment, reduces monitoring costs, expands the monitoring range, reduces manual intervention, is suitable for rapid on-site monitoring of power grids, and improves monitoring accuracy and efficiency.
Smart Images
Figure CN121685404A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial equipment monitoring technology, and more specifically, to a method and equipment for monitoring equipment deformation by collecting and processing point cloud data. Background Technology
[0002] In modern power systems, power grid equipment is subject to various factors during long-term operation, including environmental conditions, load variations, and material aging, which can lead to equipment deformation. This deformation not only affects the normal operation of the power grid but can also trigger safety accidents such as short circuits, equipment failures, and even fires, potentially resulting in personal injury and property damage. Currently, the main switching devices used in power grid equipment are GIS (Gas Insulated Switchgear). Therefore, timely and accurate monitoring of the status of GIS equipment in the power grid, especially its deformation, is crucial for ensuring normal production and safe operation. Based on the structural characteristics of GIS devices, the main monitoring methods currently fall into two categories: manual monitoring and ultrasonic monitoring.
[0003] Manual monitoring is the most traditional method for GIS monitoring. It typically relies on periodic inspections and maintenance, with professional technicians using measuring tools to conduct on-site monitoring, particularly on expansion joints of GIS equipment. This method suffers from low measurement efficiency, discontinuous data collection, and significant subjective influence. Furthermore, manual monitoring is tedious and labor-intensive, making it unsuitable for routine monitoring of large-scale electrical equipment. Ultrasonic-based monitoring methods offer the advantage of detecting minute cracks and defects on the equipment's surface area with high accuracy. However, these methods are costly in terms of equipment and technology, and are only suitable for monitoring and analyzing localized deformations. They also cannot monitor large-scale electrical equipment, thus limiting their monitoring scope. Summary of the Invention
[0004] To address the problems of high equipment cost, low monitoring efficiency, and inability to monitor for extended periods by traditional methods, this invention proposes a device deformation monitoring method based on ICP point cloud registration, which can efficiently, accurately, and for extended periods monitor the deformation of electrical GIS equipment.
[0005] The present invention proposes a deformation monitoring method, device and storage medium based on improved ICP point cloud registration, which can at least solve one of the technical problems in the background art.
[0006] To achieve the above objectives, this invention proposes a deformation monitoring method based on improved ICP point cloud registration, comprising: A 3D laser scanner is used to scan the target device to obtain three-dimensional point cloud data of the GIS device surface.
[0007] The acquired raw point cloud data is preprocessed. Voxelization filtering is used to downsample the point cloud data to reduce the data volume and improve the efficiency of subsequent monitoring. Simultaneously, the farthest point sampling method is used to further optimize the point distribution, ensuring that the data better represents the geometric features of the equipment, thereby improving registration accuracy.
[0008] Fast Point Feature Histogram (FPFH) features are extracted from the preprocessed point cloud data, and PCA component analysis is used to filter the features and extract the main features, which can effectively describe the local geometric structure of the point cloud data.
[0009] The RANSAC (Random Sample Consensus) method, combined with FPFH features, is used to perform coarse registration of the two sets of point cloud data. Bidirectional KD-TREE is used to accelerate the search process for corresponding point pairs in the point cloud. This step can roughly align the two sets of point cloud data, laying the foundation for subsequent fine registration.
[0010] Based on the coarse registration results, the improved ICP (Iterative Closest Point) algorithm is used to perform fine registration on the point cloud data to obtain the final transformation matrix, which is used to describe the spatial transformation relationship between the two sets of point clouds.
[0011] After completing the fine registration, the root mean square error (RMSE) is used as an evaluation metric for registration accuracy. RMSE is calculated by measuring the average error between corresponding points in two sets of point cloud data to determine the registration effect. A smaller RMSE value indicates better registration results and higher registration accuracy.
[0012] Based on the final rotation and translation matrices, the source point cloud is transformed to achieve registration between the two sets of point clouds. By comparing the differences between the two sets of point cloud data, the deformed parts of the equipment can be identified. These differences can intuitively reflect the deformation of the equipment at different times, providing important information for equipment maintenance and repair.
[0013] Furthermore, the process of performing voxelization filtering on the acquired raw point cloud data includes: The collected raw point cloud data is spatially divided according to a predetermined voxel size to form a three-dimensional voxel grid. A representative point is selected within each voxel unit; this invention selects the voxel center to represent all points within that voxel. This step helps reduce data volume, improve computational efficiency, and preserve the geometric features of the point cloud. Statistical analysis is performed on the point cloud within each voxel to calculate the number of points in the neighborhood of each point, and a threshold is set to remove isolated points in low-density areas, thereby improving the quality of the point cloud data. Missing points are appropriately filled in using an interpolation algorithm, further enhancing the integrity and reliability of the point cloud data. The voxelized point cloud data is converted into a data format suitable for subsequent ICP registration algorithms, ensuring that the point cloud data can be matched and calculated efficiently and accurately during the registration process.
[0014] Furthermore, the process of performing farthest point sampling on the acquired raw point cloud data includes: randomly selecting a point from the acquired raw point cloud data as an initial sampling point, which will serve as the starting point of the Farthest Point Sampling (FPS) algorithm; calculating the Euclidean distance from all other points to the initial sampling point, and determining the point farthest from the initial point; selecting the point farthest from the initial point as the next sampling point, recording this new sampling point, and updating the minimum distance from all other points to the currently selected sampling point set; repeating the above steps to iteratively select new farthest points until a preset number of sampling points is reached; through the above iterative selection process, ensuring that the sampled points are evenly distributed throughout the entire point cloud dataset, avoiding local over-density or sparseness.
[0015] Furthermore, the process of obtaining the FPFH features of the point cloud includes: S1. Use the k-NN algorithm to determine the set of neighboring points for each point and select an appropriate search radius to ensure that the neighboring points sufficiently represent the local geometric information of the point. S2. For each point, construct a feature vector based on its relationship with neighboring points. The feature vector mainly contains three features. .in, This represents the angle formed by projecting the angle between the reference point's normal vector and the normal vectors of neighboring points onto the normal plane of the reference point's normal vector. This represents the angle between the vector from the reference point to a neighboring point and the normal vector of the reference point. This represents the Euclidean distance from the reference point to neighboring points; S3. Quantize the calculated feature values and construct a feature histogram. Divide each feature value into several intervals, count the number of points in each interval, and generate a simplified feature histogram, which is the SPFH feature; S4. Based on the simplified feature histogram, for each point, its own SPFH feature is weighted and accumulated with the SPFH features of its neighboring points according to a certain weight, so as to comprehensively consider the geometric features of the local area. S5. Normalize the calculated FPFH features to ensure that all components of the feature vector are within the same scale range. This helps improve the robustness of feature matching and avoids errors caused by differences in feature value ranges.
[0016] FPFH features preserve the local geometric information of point cloud data, providing reliable feature support for subsequent point cloud registration and deformation monitoring.
[0017] Furthermore, the principle and basic steps of the RANSAC algorithm include: During point cloud registration, the sparsity and noise in the collected point cloud data can lead to registration failures, affecting the overall accuracy of the registration result. This invention uses the RANSAC algorithm combined with the FPFH method for coarse registration while eliminating erroneous matching point pairs. The basic steps of RANSAC are as follows: S1. For the point cloud registration problem, randomly select a minimum subset (three points) from the input point cloud data and fit a planar model. S2. For other points, calculate the position of each point according to the fitted model, evaluate its fit with the model, and if the distance between a point and the model is less than the preset critical value, mark it as an interior point; otherwise, mark it as an exterior point. Use distance measurement to determine whether the point conforms to the current model, thereby distinguishing between interior and exterior points. S3. Repeat the process of random sample selection and interior / exterior point calculation until the predetermined number of iterations is reached. During each iteration, record the number of interior points in the current model and compare it with the number of interior points in the previous best model. Select the model with the most interior points as the final model. S4. For the final model, based on its determined set of interior points, the singular value decomposition method is used to solve the transformation matrix of the coarse registration point cloud, laying the foundation for subsequent fine registration.
[0018] Optionally, the principle and basic steps of the ICP algorithm include: The basic idea of the ICP (Iterative Closest Point) algorithm is to align two or more sets of point cloud data so that the two point clouds overlap as much as possible. The specific steps of the ICP algorithm are as follows: S1. Select one point from the source point cloud and find the corresponding point in the target point cloud. That is, calculate the Euclidean distance between the two points so that the absolute value of the distance is minimized. S2. Calculate the rigid transformation matrix by matching the point pairs; S3. Apply the calculated transformation matrix to the source point cloud to generate a new source point cloud, and calculate the mean square error between the new source point cloud and the target point cloud. The calculation formula is as follows:
[0019] Where e represents the mean square error, p represents the target point cloud, q represents the source point cloud, R represents the rotation matrix in the rigid transformation matrix, and T represents the translation matrix in the transformation matrix.
[0020] S4. Set a threshold. When the mean squared error is less than the preset threshold, end the current iteration and use the current rotation matrix and translation matrix as the final transformation matrix. Otherwise, return to step S1 until it is less than the preset threshold or the maximum number of iterations is reached.
[0021] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.
[0022] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.
[0023] As can be seen from the above technical solution, the deformation monitoring method and system based on improved ICP point cloud registration of the present invention belongs to the field of industrial equipment deformation monitoring technology, including the following steps: point cloud data acquisition; data preprocessing using voxelization filtering and farthest point sampling methods; extraction of FPFH features of the point cloud, and screening of features using PCA principal component analysis to extract the main features; coarse registration of the point cloud data using the RANSAC method combined with FPFH features, and optimization of the process of finding corresponding points in the algorithm using the bidirectional KD-TREE method; fine registration of the point cloud using the improved ICP algorithm to obtain the transformation matrix; using RMSE as the point cloud registration accuracy index to evaluate the quality of point cloud registration; point cloud registration based on the transformation matrix, and comparing the differences between the two point clouds to identify the deformed parts of the equipment. This invention addresses the problems of cumbersome manual monitoring steps and the inability to conduct long-term monitoring in GIS, using improved ICP point cloud registration technology to achieve deformation monitoring of GIS equipment, solving the problems of low accuracy, cumbersome monitoring steps, and inability to conduct long-term monitoring in traditional GIS equipment monitoring.
[0024] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention addresses the issues of high computational cost and denser point cloud data in the point cloud registration process by using preprocessed points as feature points, thus covering the entire point cloud structure with a smaller number of points, providing a foundation for subsequent point cloud registration.
[0025] This invention addresses the problems of inaccurate registration on complex surfaces and high requirements for initial pose in the ICP point cloud registration method by proposing an improved point cloud registration method.
[0026] This invention addresses the problems of high cost and limited monitoring range of ultrasonic monitoring equipment, as well as the cumbersome procedures of traditional manual deformation monitoring methods. It uses a 3D laser scanner to scan GIS equipment, thereby reducing the cost of monitoring equipment and expanding the monitoring range. It achieves full-process monitoring without manual intervention, which is more conducive to rapid monitoring in the power grid field and meets the practical needs of low-cost, large-scale monitoring of electrical equipment. Attached Figure Description
[0027] Figure 1 This is a flowchart of the present invention; Figure 2 This is the conventional registration result with high overlap rate in an embodiment of the present invention; Figure 3 This is the registration result with low overlap rate in an embodiment of the present invention. Detailed Implementation
[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.
[0029] like Figure 1 As shown in the figure, the deformation monitoring method based on improved ICP point cloud registration described in this embodiment includes the following steps: A 3D laser scanner is used to scan the target device to obtain three-dimensional point cloud data of the GIS device surface.
[0030] The acquired raw point cloud data is preprocessed. Voxelization filtering is used to downsample the point cloud data to reduce the data volume and improve the efficiency of subsequent monitoring. Simultaneously, the farthest point sampling method is used to further optimize the point distribution, ensuring that the data better represents the geometric features of the device, thereby improving registration accuracy. The farthest point is iteratively selected from the point cloud using the following formula:
[0031] in For the selected point set, This is the original point cloud.
[0032] Fast Point Feature Histogram (FPFH) features are extracted from preprocessed point cloud data, and PCA component analysis is used to filter and extract the main features, which can effectively describe the local geometric structure of the point cloud data. To obtain more accurate local geometric features, this invention uses a neighborhood modeling method based on Gaussian distribution to derive the normal vector and curvature. By analyzing the points... neighborhood point set ( Construct a Gaussian distribution model i s (μ i s ,∑ i s In the Gaussian model, the covariance and mean deviation represent the geometric features at the center point, namely the normal vector and curvature:
[0033]
[0034] Where, μ i s Indicates the location of the centroid of the neighborhood, ∑ i s It represents the distribution characteristics of local curved surface morphology.
[0035] The RANSAC (Random Sample Consensus) method, combined with FPFH features, is used to perform coarse registration of the two sets of point cloud data. Bidirectional KD-TREE is used to accelerate the search process for corresponding point pairs in the point cloud. This step can roughly align the two sets of point cloud data, laying the foundation for subsequent fine registration.
[0036] Based on the coarse registration results, the improved ICP (Iterative Closest Point) algorithm is used to perform fine registration on the point cloud data to obtain the final transformation matrix, which is used to describe the spatial transformation relationship between the two sets of point clouds.
[0037] After completing the fine registration, the root mean square error (RMSE) is used as an evaluation metric for registration accuracy. RMSE calculates the average error between corresponding points in two sets of point cloud data to determine the registration effect. A smaller RMSE value indicates better registration and higher accuracy. RMSE quantifies the deviation between predicted and true values to measure the accuracy of the model or algorithm, thus reflecting the overall error magnitude. Its calculation formula is as follows:
[0038] Based on the final rotation and translation matrices, the source point cloud is transformed to achieve registration between the two sets of point clouds. By comparing the differences between the two sets of point cloud data, the deformed parts of the equipment can be identified. These differences can intuitively reflect the deformation of the equipment at different times, providing important information for equipment maintenance and repair.
[0039] Specifically, the process of performing voxelization filtering on the acquired raw point cloud data includes: The collected raw point cloud data is spatially divided according to a predetermined voxel size to form a three-dimensional voxel grid. A representative point is selected within each voxel unit; this invention selects the voxel center to represent all points within that voxel. This step helps reduce data volume, improve computational efficiency, and preserve the geometric features of the point cloud. Statistical analysis is performed on the point cloud within each voxel to calculate the number of points in the neighborhood of each point, and a threshold is set to remove isolated points in low-density areas, thereby improving the quality of the point cloud data. Missing points are appropriately filled in using an interpolation algorithm, further enhancing the integrity and reliability of the point cloud data. The voxelized point cloud data is converted into a data format suitable for subsequent ICP registration algorithms, ensuring that the point cloud data can be matched and calculated efficiently and accurately during the registration process.
[0040] The process of performing farthest point sampling on the acquired raw point cloud data includes: Randomly select a point from the acquired raw point cloud data as the initial sampling point. This initial point will serve as the starting point for the Farthest Point Sampling (FPS) algorithm. Calculate the Euclidean distance from all other points to the initial sampling point to determine the point farthest from the initial point. Select the point farthest from the initial point as the next sampling point, record this new sampling point, and update the minimum distance from all other points to the currently selected sampling point set. Repeat the above steps iteratively to select new farthest points until the preset number of sampling points is reached. Through the above iterative selection process, ensure that the sampled points are evenly distributed throughout the entire point cloud dataset, avoiding local over-density or sparseness.
[0041] The process of obtaining FPFH features from the point cloud includes: S1: The k-NN algorithm is used to determine the set of neighboring points for each point and to select an appropriate search radius to ensure that the neighboring points are sufficiently representative of the local geometric information of the point. S2: For each point, construct a feature vector based on its relationship with neighboring points. The feature vector mainly contains three features. .in, This represents the angle formed by projecting the angle between the reference point's normal vector and the normal vectors of neighboring points onto the normal plane of the reference point's normal vector. This represents the angle between the vector from the reference point to a neighboring point and the normal vector of the reference point. This represents the Euclidean distance from the reference point to neighboring points; S3: Quantize the calculated feature values and construct a feature histogram. Divide each feature value into several intervals, count the number of points in each interval, and generate a simplified feature histogram, which is the SPFH feature; S4: Based on the simplified feature histogram, for each point, its own SPFH feature is weighted and accumulated with the SPFH features of its neighboring points according to a certain weight, so as to comprehensively consider the geometric features of the local area. S5: Normalize the calculated FPFH features to ensure that all components of the feature vector are within the same scale range. This helps improve the robustness of feature matching and avoids errors caused by differences in feature value ranges. The general form of FPFH features is as follows:
[0042] in, This represents the angle between the normal vectors of two points. This represents the angle between the vector connecting the two points and the normal vector of the source point. This represents the angle between the vector connecting the two points and the normal vector of the neighboring points. , , Point The normal vectors in three-dimensional space are respectively located at... , , Normal vector components on the axis.
[0043] FPFH features preserve the local geometric information of point cloud data, providing reliable feature support for subsequent point cloud registration and deformation monitoring.
[0044] The principle and steps of the PCA principal component analysis method include: This invention uses Principal Component Analysis (PCA) to reduce the dimensionality of the extracted Fast Point Feature Histogram (FPFH) feature descriptors. This method effectively removes vectors with insignificant features, retaining only the most critical feature information. The resulting point cloud data points with rich feature information are selected as feature points for initial registration. The specific steps are as follows: S1: Calculate the average value of the FPFH feature descriptors to determine their central location in the feature space.
[0045] S2: Based on the FPFH feature descriptor, solve for its covariance matrix. The covariance matrix reveals the correlation between different feature dimensions, thus assessing the correlation between features. Perform orthogonal decomposition on the covariance matrix:
[0046] in, , , It is ∑ arranged in ascending order i s The eigenvalues, take the minimum eigenvalue. Corresponding feature vector For normal line Curvature value 0 and 1 reflect the degree of curvature of a local surface. The specific formula is as follows:
[0047] S3: Based on the covariance matrix, find the main directions of change by solving for the eigenvalues and corresponding eigenvectors of the covariance matrix. The directions with larger eigenvalues represent the dimensions of the data changes most significantly, i.e., the principal components; while the directions with smaller eigenvalues often contain less information and may be noise or redundant information.
[0048] S4: Construct a transformation matrix based on the eigenvalues and eigenvectors. This transformation matrix projects the original FPFH feature descriptors into a low-dimensional space, resulting in dimensionality-reduced feature descriptors. These dimensionality-reduced descriptors capture the most critical feature information while removing unnecessary dimensions, thus improving data processing efficiency. Points in the low-dimensional space serve as key feature points for coarse registration.
[0049] The principles and basic steps of the RANSAC algorithm include: During point cloud registration, the sparsity and noise in the collected point cloud data can lead to registration failures, affecting the overall accuracy of the registration result. This invention uses the RANSAC algorithm combined with the FPFH method for coarse registration while eliminating erroneous matching point pairs. The basic steps of RANSAC are as follows: S1: For the point cloud registration problem, randomly select a minimum subset (three points) from the input point cloud data and fit a planar model; S2: For other points, calculate the position of each point according to the fitted model, evaluate its fit with the model, and if the distance between a point and the model is less than the preset critical value, mark it as an interior point; otherwise, mark it as an exterior point. Use distance measurement to determine whether the point conforms to the current model, thereby distinguishing between interior and exterior points. S3: Repeat the process of random sample selection and interior / exterior point calculation until the predetermined number of iterations is reached. During each iteration, the number of interior points in the current model is recorded and compared with the number of interior points in the previous best model. The model with the most interior points is selected as the final model. S4: For the final model, based on its determined set of interior points, the singular value decomposition method is used to solve the transformation matrix of the coarse registration point cloud, laying the foundation for subsequent fine registration.
[0050] The principle and steps of the bidirectional KD-TREE include: Traditional Iterative Closest Point (ICP) methods find corresponding points by searching the entire point cloud data one by one, a time-consuming and inefficient process. To address this issue, this invention employs an optimization strategy using a bidirectional KD-Tree. Specifically, S1: Establish KD-Tree for the point cloud to be registered and the target point cloud respectively, and use this structure to quickly find the nearest neighbor of each point in the other point cloud, thereby generating two sets of corresponding point sets.
[0051] S2: Select the intersection of the two point clouds from these two sets of point sets to form the final valid corresponding point set.
[0052] Compared with the conventional one-way KD-Tree method, the two-way KD-Tree method not only significantly improves the accuracy of the registration point set, but also effectively avoids incorrect matching, thus providing a good initial pose.
[0053] The principle and basic steps of the ICP algorithm include: The basic idea of the ICP (Iterative Closest Point) algorithm is to align two or more sets of point cloud data so that the two point clouds overlap as much as possible. The specific steps of the ICP algorithm are as follows: S1: Select one point from the source point cloud, find the corresponding point in the target point cloud, and calculate the Euclidean distance between the two points to minimize the absolute value of the distance. S2: Calculate the rigid transformation matrix by matching point pairs; S3: Apply the calculated transformation matrix to the source point cloud to generate a new source point cloud, and calculate the mean square error between the new source point cloud and the target point cloud. The calculation formula is as follows:
[0054] Where e represents the mean square error, p represents the target point cloud, q represents the source point cloud, R represents the rotation matrix in the rigid transformation matrix, and T represents the translation matrix in the transformation matrix.
[0055] S4: Set a threshold. When the mean squared error is less than the preset threshold, end the current iteration and use the current rotation and translation matrices as the final transformation matrices. Otherwise, return to step S1 until convergence to a preset threshold or the maximum number of iterations is reached.
[0056] The improved ICP point cloud registration algorithm includes: This invention proposes a method for calculating the distance between point pairs. By changing the formula for calculating the distance between point pairs, specifically by considering only the squared Euclidean distance from a point to its corresponding point on the tangent plane, it eliminates erroneous matching points that may be caused by surface irregularities. The specific steps are as follows: S1: For any point p in the point cloud P, calculate the normal vector and tangent plane of that point; S2: For the corresponding point q, construct a perpendicular line in three-dimensional space that is perpendicular to the tangent plane. The intersection of the two lines is point p'. S3: Calculate the Euclidean distance between point q and point p', minimizing this value as the objective function of the algorithm, which can effectively reduce the adverse effects of complex surface features on registration accuracy.
[0057] The following are examples: Data preparation A 3D laser scanner is used to scan the target device to obtain three-dimensional point cloud data of the GIS device surface.
[0058] Data processing First, the measured point cloud data is voxelized and filtered. The raw point cloud data is spatially divided according to a predetermined voxel size to form a three-dimensional voxel grid. Within each voxel cell, a voxel center is selected to represent all points within that voxel. This step helps reduce the amount of data, improve computational efficiency, and preserve the geometric features of the point cloud. Statistical analysis is performed on the point cloud within each voxel to calculate the number of points in the neighborhood of each point. A threshold is set to remove isolated points in low-density areas, thereby improving the quality of the point cloud data. Through interpolation algorithms, missing points are reasonably filled in, further improving the integrity and reliability of the point cloud data. Next, farthest point sampling is performed. A point is selected as the starting point, and the Euclidean distance from all other points to the initial sampling point is calculated to determine the point farthest from the initial point. The point farthest from the initial point is selected as the next sampling point, this new sampling point is recorded, and the minimum distance from all other points to the currently selected sampling point set is updated. The above steps are repeated iteratively to select a new farthest point until the preset number of sampling points is reached. Point cloud registration Coarse registration: First, calculate the FPFH feature descriptors of the two sets of point clouds. Then, use RANSAC combined with the FPFH feature descriptors to perform point-to-point coarse registration of the two sets of point clouds to achieve approximate alignment of the two sets of point clouds, providing a good positional relationship for subsequent fine registration.
[0059] Fine registration: Based on the coarse registration results, the fine registration algorithm is used to further optimize the position of the point cloud, perfectly register the two sets of point clouds, minimize the registration error between the two sets of point clouds, and improve the accuracy of subsequent deformation monitoring.
[0060] Algorithm Evaluation This invention uses root mean square error (RMSE) to evaluate and compare the accuracy of the proposed algorithm. Fifty sets of point cloud data acquired using a 3D scanner were used as experimental data, and the average value of these 50 sets of evaluation metrics was used to assess the algorithm's accuracy.
[0061] Deformation monitoring For each pair of corresponding points, the distance between the points is calculated. If the distance is greater than zero, the point cloud data is considered to have deformed at the corresponding location. The deformation monitoring results are output to generate a detailed deformation report, including the location and degree of deformation, as well as a visualization, which helps engineers quickly understand the health status of the equipment and take appropriate maintenance measures.
[0062] To comprehensively evaluate the performance of the proposed method in handling complex surfaces and the impact of initial pose on the registration effect, this experiment conducted two registration experiments on a self-built GIS device point cloud dataset, one with high overlap rate and the other with low overlap rate. Comparative experiments were conducted using ICP, SAC-IA, Go-ICP, VGICP, and the algorithm proposed in this study to comprehensively verify the performance and robustness of the algorithm.
[0063] (1) Analysis of conventional registration results with high overlap rate The high overlap rate conventional registration experiment selected point cloud data with an overlap rate of 75% for testing. The visualization results are as follows: Figure 2 As shown in the figure, the initial pose and the performance of each algorithm's registration results are illustrated. Figure 2 (a) is the initial pose, where the point cloud P to be registered is red and the ideal point cloud Q is green.
[0064] The visualization results show that while the SAC-IA algorithm achieves registration, significant local misalignment remains. In contrast, the ICP algorithm achieves better registration, but some local areas at the top fail to fully overlap. Go-ICP and VGICP still exhibit minor misalignment in the fine structural regions at the top of the device. The method of this invention, however, achieves high-precision registration in both global and local structural regions, significantly outperforming other algorithms.
[0065] As shown in Table 1 of the experimental data, in high overlap scenarios, the method of this invention outperforms other algorithms in both registration accuracy and efficiency. Specifically, the registration time of this invention is 14.50 s, the RMSE is 9.23 × 10⁻³ mm, and the LCP is 94.78%. Compared with other methods, the accuracy of this invention is significantly improved, with the RMSE being 51.2%, 94.8%, 46.3%, and 48.2% higher than that of ICP, SAC-IA, Go-ICP, and VGICP algorithms, respectively. Simultaneously, the registration time is reduced by 28.4%, 47.17%, 18.68%, and 20.77%, respectively. These data indicate that the method proposed in this invention can significantly improve registration accuracy in high overlap scenarios, while also demonstrating a significant advantage in computational efficiency. Although the ICP algorithm achieves a relatively high RMSE of 1.89 × 10⁻², its LCP is only 70.84%, indicating a significant number of mismatches during the registration process. This is because the ICP algorithm optimizes only based on point-to-point geometric distance, ignoring the structural information of the point cloud, leading to significant local misalignment. While SAC-IA and Go-ICP introduce more complex global and local structural constraints, they still cannot completely eliminate subtle local misalignments in highly overlapping scenarios, especially in the fine structural regions at the top of the device. In contrast, the method of this invention, by fusing multimodal features, including geometric shape, normal vectors, and curvature information, more effectively captures the details and local features of the point cloud. It not only demonstrates superior global registration accuracy but also avoids misalignment when handling complex surfaces and detailed regions, showing a clear advantage, particularly in the registration of fine structures at the top of the device.
[0066] Table 1 Comparison of Registration Accuracy with High Overlap Rate Conventional Registration
[0067] The proposed method outperforms other methods in registration accuracy and computational efficiency in high overlap scenarios, validating its advantages in complex curved surface scenes. This demonstrates that the proposed method can better handle point cloud registration tasks with high overlap and complex structures, exhibiting strong robustness and practicality.
[0068] (2) Analysis of extreme registration results with low overlap rate To verify the robustness of the proposed method in extreme scenarios with low overlap rates, this experiment selected a set of GIS device point cloud data samples with non-overlapping regions for testing and result analysis. The experimental results are as follows: Figure 3 As shown, Figure 3 (a) shows the initial pose, where red represents the point cloud P to be registered and green represents the ideal point cloud Q.
[0069] Experimental results show that the ICP algorithm fails to achieve effective registration due to iterative divergence caused by the lack of initial matching relationships. Figure 3 As shown in (b), although SAC-IA, Go-ICP, and VGICP successfully registered, there are still varying degrees of non-overlapping regions between the point cloud to be registered and the ideal point cloud, such as... Figure 3 As shown in (c)-(e). In contrast, the method of this invention shows a significant advantage in low overlap scenarios, achieving registration results with high global accuracy and high local detail overlap, almost perfectly matching the ideal point cloud, such as... Figure 3 As shown in (f), the effectiveness and superiority of this method in such complex scenarios are fully verified.
[0070] Table 2 Comparison of Extreme Registration Accuracy with Low Overlap Rate
[0071] Note: "_" indicates that the ICP algorithm did not find the optimal solution within 120 seconds.
[0072] Experimental results show that the ICP algorithm failed to achieve effective registration due to iterative divergence caused by the lack of initial matching relationships, as shown in Figure 3(b). Although SAC-IA, Go-ICP, and VGICP successfully registered, there were still varying degrees of non-overlapping regions between the point cloud to be registered and the ideal point cloud, such as... Figure 3 As shown in (c)-(e). In contrast, the method of this invention exhibits significant advantages in low overlap scenarios, achieving registration results with high global accuracy and high local detail overlap, almost perfectly matching the ideal point cloud, such as... Figure 3 As shown in (f), the effectiveness and superiority of this method in such complex scenarios are fully verified.
[0073] Table 2 shows that the method of this invention exhibits significant advantages in the zero-overlap extreme registration experiment. Specifically, the ICP algorithm fails to register due to excessive initial pose differences. The SAC-IA algorithm mitigates the impact of initial pose differences by utilizing point cloud surface features, but the registration accuracy is still insufficient, with significant deviations. In low overlap scenarios, the RMSE of SAC-IA is 1.13 × 10⁻⁶. - The LCP (Local Point Compatibility) is 83.44%, and the registration time is 31.86 seconds, indicating limited ability to handle complex scenes and low accuracy. Go-ICP and VGICP improve point-to-point matching accuracy by introducing a face-to-face distance formula, but this also increases computational complexity, with registration times of 22.42 seconds and 23.13 seconds respectively, higher than ICP and the method of this invention. In contrast, the RMSE of the method of this invention is only 1.84 × 10⁻⁶. - The method achieves a registration accuracy of 2 mm, an LCP of 90.25%, and a registration time of 19.61 seconds, which is lower than other methods. This invention not only achieves the lowest RMSE in registration accuracy but also optimal performance in LCP value, indicating that its registration effect is more precise.
[0074] In summary, our proposed method achieves good registration results in both high and low overlap scenarios. Compared with mainstream methods such as Go-ICP, VGICP, and SAC-IA, our method not only effectively improves registration accuracy and reduces computation time, but also exhibits stronger stability and robustness in complex low overlap environments.
[0075] In another aspect, the present invention also discloses a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method described above.
[0076] In another aspect, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the method described above.
[0077] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the deformation monitoring methods based on improved ICP point cloud registration in the above embodiments.
[0078] It is understood that the systems, devices, and storage media provided in the embodiments of the present invention correspond to the methods provided in the embodiments of the present invention, and the explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.
[0079] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state disk (SSD)).
[0080] It should be noted that in this invention, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0081] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.
[0082] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A deformation monitoring method based on improved ICP point cloud registration, characterized in that, The method comprises the following steps, Scanning the target equipment by using a 3D laser scanner to obtain three-dimensional point cloud data of the surface of the GIS equipment; Pretreating the collected original point cloud data; Extracting FPFH features from the pretreated point cloud data, and screening the features by using PCA component analysis to extract main features; Using the RANSAC method and combining the FPFH features to coarsely register two groups of point cloud data, and using bidirectional KD-TREE to accelerate the search process of corresponding point pairs in the point cloud; On the coarse registration result, an improved ICP algorithm is used to finely register the point cloud data to obtain a final transformation matrix, which is used to describe the spatial transformation relationship between the two groups of point cloud data; After completing the fine registration, the root mean square error is used as an evaluation index of registration accuracy; According to the final rotation matrix and translation matrix, the source point cloud is transformed so that the two groups of point cloud data are registered, and by comparing the difference between the two groups of point cloud data, the deformation part of the equipment can be identified.
2. The deformation monitoring method based on improved ICP point cloud registration according to claim 1, characterized in that: The pretreatment of the collected original point cloud data includes using a voxelization filtering method to downsample the point cloud data to reduce the data volume and improve the efficiency of subsequent monitoring; at the same time, a farthest point sampling method is used to further optimize the distribution of points to ensure that the data can better represent the geometric features of the equipment, thereby improving the registration accuracy.
3. The deformation monitoring method based on improved ICP point cloud registration according to claim 2, characterized in that: Downsampling the point cloud data by using the voxelization filtering method includes dividing the collected original point cloud data into three-dimensional voxel grids according to a predetermined voxel size; Selecting representative points in each voxel unit; statistically analyzing the point cloud in the voxel, calculating the number of points in the neighborhood of each point, and setting a threshold to remove isolated points in low-density areas, thereby improving the quality of the point cloud data; By using an interpolation algorithm, the missing points are reasonably completed to further improve the integrity and reliability of the point cloud data; Converting the point cloud data after voxelization into a data format suitable for subsequent ICP registration algorithm to ensure that the point cloud data can be efficiently and accurately matched and calculated in the registration process.
4. The deformation monitoring method based on improved ICP point cloud registration according to claim 2, characterized in that: Further optimizing the distribution of points by using the farthest point sampling method, specifically including, Randomly selecting a point from the collected original point cloud data as an initial sampling point, which will be the starting point of the farthest point sampling algorithm; Calculating the Euclidean distance of all other points to the initial sampling point to determine the point farthest from the initial point; selecting the point farthest from the initial point as the next sampling point, recording the new sampling point, and updating the minimum distance of all other points to the current selected sampling point set; Repeat the above steps to iteratively select new farthest points until the preset number of sampling points is reached; through the above iterative selection process, it is ensured that the sampled points are uniformly distributed in the entire point cloud data set, avoiding local over-dense or sparse phenomena.
5. The deformation monitoring method based on improved ICP point cloud registration according to claim 4, characterized in that: Extracting FPFH features from the pretreated point cloud data, specifically including, Using a k-NN algorithm to determine the neighbor point set of each point, and selecting a set search radius to ensure that the neighborhood points sufficiently represent the local geometric information of the point; For each point, a feature vector is constructed according to its relationship with the neighborhood points, the feature vector contains three features ; wherein, denotes the angle formed by the projection of the angle between the reference point normal vector and the neighborhood point normal vector onto the normal plane of the reference point normal vector, denotes the angle between the vector from the reference point to the neighborhood point and the reference point normal vector, denotes the Euclidean distance from the reference point to the neighborhood point; The calculated characteristic value is quantized to construct a characteristic histogram; each characteristic value is divided into several intervals, and the number of points in each interval is counted to generate a simplified characteristic histogram, namely the SPFH characteristic; On the basis of the simplified characteristic histogram, for each point, the SPFH characteristics of the point itself and its neighborhood points are weighted and accumulated according to certain weights, so as to comprehensively consider the geometric characteristics of the local region; The calculated FPFH characteristics are normalized to make the components of the feature vector in the same scale range, which is helpful to improve the robustness of feature matching and avoid errors caused by the difference in the range of characteristic values.
6. The deformation monitoring method based on improved ICP point cloud registration according to claim 5, characterized in that: The PCA component analysis method is used to screen and extract main characteristics, specifically including, The average value of the FPFH feature descriptor is calculated to determine the center position of the feature in the feature space; Based on the FPFH feature descriptor, the covariance matrix is solved, and the covariance matrix reveals the correlation between different feature dimensions to evaluate the correlation between the features; On the basis of the covariance matrix, the eigenvalues and eigenvectors of the covariance matrix are solved to find the main variation direction; According to the eigenvalues and eigenvectors, a transformation matrix is constructed; through the transformation matrix, the original FPFH feature descriptor is projected into a low-dimensional space to obtain the dimension-reduced feature descriptor; wherein the points in the low-dimensional space are used as the key feature points of coarse registration.
7. The deformation monitoring method based on improved ICP point cloud registration according to claim 6, characterized in that: The RANSAC method is used in combination with the FPFH characteristics to perform coarse registration on two groups of point cloud data, specifically including, For the point cloud registration problem, a minimum subset is randomly selected from the input point cloud data to fit a plane model; For other points, the position of each point is calculated according to the fitted model to evaluate its fitness to the model; if the distance between a certain point and the model is less than a predetermined threshold, the point is marked as an inlier, otherwise it is marked as an outlier; the distance metric is used to determine whether the point meets the current model, thereby distinguishing between inliers and outliers; The process of randomly selecting samples and solving inliers and outliers is repeated until a predetermined number of iterations is reached; wherein in each iteration process, the number of inliers of the current model is recorded and compared with the number of inliers of the previous best model, and the model with the most inliers is selected as the final model; For the final model, the singular value decomposition method is used to solve the transformation matrix of the coarse registration point cloud, laying a foundation for subsequent fine registration.
8. The deformation monitoring method based on improved ICP point cloud registration according to claim 7, characterized in that: The bidirectional KD-TREE is used to accelerate the search process of corresponding point pairs in the point cloud, specifically including, KD-Trees are respectively established for the point cloud to be registered and the target point cloud, and this structure is used to quickly find the nearest neighbor of each point in the other point cloud, and then two sets of corresponding point sets are generated; The intersection of the two sets of point sets is screened to form the final effective corresponding point set.
9. The deformation monitoring method based on improved ICP point cloud registration according to claim 8, characterized in that: The improved ICP algorithm is used to perform fine registration on the point cloud data, including: For any point p in the point cloud P, the normal vector of the point is calculated and a tangent plane is constructed; For the corresponding point q, a vertical line perpendicular to the tangent plane is constructed in the three-dimensional space, and the intersection point of the two lines is the point p'; The Euclidean distance between the point q and the point p' is calculated, so that the value is minimized as the objective function of the algorithm, which can effectively reduce the adverse effects of complex surface features on the registration accuracy. 10.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-9. The computer program, when executed by the processor, causes the processor to perform the steps of the method of any one of claims 1 to 9.