Transmission line point cloud fitting method

Abstract

The power transmission line point cloud fitting method breaks through the information limitation of single point cloud feature by fusing point cloud geometric features and image semantic features, provides comprehensive feature support for inlier region mining, accurately locates the high-confidence inlier region by a composite swarm intelligence optimization algorithm, generates a sampling weight, guides the RANSAC algorithm to perform weighted non-uniform sampling, replaces the traditional pure random sampling mode, reduces the probability of outliers and invalid points participating in model construction from the root, and then reduces the number of invalid iterations, improves the fitting efficiency, guarantees the accuracy of the fitted model, and meets the dual requirements of fitting accuracy and processing efficiency in the power transmission line inspection scene.

Description

Technical Field

[0001] This application relates to the field of power system equipment condition monitoring, specifically to a method for fitting point clouds of transmission lines. Background Technology

[0002] With the deepening of smart grid construction, UAV lidar inspection technology has been applied on a large scale in the operation and maintenance of power transmission lines. Accurate and efficient fitting of the three-dimensional point cloud of power transmission lines is the premise and foundation for carrying out core tasks such as line morphology reconstruction, intelligent defect identification, and electrical safety distance verification. The industry has put forward stringent engineering requirements for the accuracy, processing efficiency and anti-interference ability of the power transmission line point cloud fitting algorithm in complex environments.

[0003] In related technologies, the Random Sample Consensus (RANSAC) algorithm is usually used as the core fitting algorithm, supplemented by preprocessing methods such as point cloud filtering and noise reduction, and multimodal data feature screening, to achieve point cloud model fitting of key components such as transmission line conductors, towers, and insulators.

[0004] However, the RANSAC algorithm used in existing solutions relies on pure random sampling to select fitting primitives. Under the condition that transmission line point clouds generally have high noise and a high proportion of outliers, it is easy to select outliers or invalid points to participate in model construction. This not only leads to insufficient accuracy of the fitting model, but also requires a large number of invalid iterations to obtain an effective model. The fitting efficiency is low and cannot meet the dual core requirements of high accuracy and high efficiency for fitting transmission line point clouds in smart grid inspection scenarios. Summary of the Invention

[0005] This application provides a point cloud fitting method for transmission lines, which can solve the technical problems of point cloud fitting algorithms for transmission lines in related technologies, such as strong sampling blindness, insufficient fitting accuracy, low iteration efficiency, and inability to adapt to transmission line inspection scenarios with high noise and high proportion of outliers.

[0006] In a first aspect, embodiments of this application provide a method for fitting point clouds of transmission lines, the method comprising: Geometric features corresponding to the point cloud data of the transmission line and semantic features corresponding to the image data of the transmission line are extracted. Feature fusion is performed on the geometric features and semantic features to obtain the multimodal fusion features corresponding to each point cloud. Based on the multimodal fusion features, a composite swarm intelligent optimization algorithm is used to mine high-confidence in-point regions in the point cloud of transmission lines, and to generate sampling weights corresponding to each high-confidence in-point region. Guided by the aforementioned sampling weights, weighted non-uniform sampling is performed on the sampling process of the random sampling consensus algorithm to obtain the target sampling point set; Based on the target sampling point set, the geometric model corresponding to the transmission line component is fitted to complete the point cloud fitting of the transmission line.

[0007] The beneficial effects of the technical solutions provided in this application include: By fusing point cloud geometric features with image semantic features, the limitations of single point cloud features are overcome, providing comprehensive feature support for interior point region mining. A composite swarm intelligence optimization algorithm is used to accurately locate high-confidence interior point regions and generate sampling weights, thereby guiding the RANSAC algorithm to perform weighted non-uniform sampling, replacing the traditional purely random sampling method. This fundamentally reduces the probability of outliers and invalid points participating in model construction, thereby reducing the number of invalid iterations, improving fitting efficiency, and ensuring the accuracy of the fitted model. This meets the dual requirements of fitting accuracy and processing efficiency in power transmission line inspection scenarios. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the first embodiment of the point cloud fitting method for transmission lines in this application; Figure 2 This is a flowchart illustrating another embodiment of the point cloud fitting method for transmission lines in this application. Detailed Implementation

[0009] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0010] First, some of the technical terms used in this application will be explained to help those skilled in the art understand this application.

[0011] RANSAC Algorithm: RANSAC is a classic iterative robust estimation algorithm. Its core objective is to accurately estimate mathematical model parameters from a dataset containing a large number of noise and interference data points. Through the iterative logic of "random sampling, model fitting, interior point determination, and optimal model selection", it effectively shields out-point interference and has extremely strong noise resistance. It is a fundamental core algorithm in the field of point cloud fitting and model estimation.

[0012] Each time, the algorithm randomly selects the smallest subset from the dataset to fit the initial model and counts the number of inliers that meet the model error threshold. After repeating the iteration a preset number of times, the model with the most inliers is selected as the optimal model. Finally, the model parameters are re-optimized based on all inliers to achieve robust fitting.

[0013] The OOA algorithm, or Octopus Optimization Algorithm, is a metaheuristic algorithm inspired by the intelligent behavior of octopuses. These creatures, often called "ocean strategists," possess not only three hearts and eight tentacles, but also adaptive intelligence rarely seen in nature. They can dynamically change their skin color and texture for camouflage and flexibly use their tentacles to explore their environment and manipulate objects. Researchers have leveraged this unique movement and perception mechanism to propose the OOA algorithm, simulating three key behavioral strategies of octopuses in the search space: a tentacles-based exploration mechanism based on random scaling, herd-oriented movement towards the optimal position, and a multi-tentacle collaborative problem-solving approach. These strategies are transformed into efficient mathematical models that demonstrate excellent global exploration capabilities and convergence stability when dealing with complex optimization problems. GWO Algorithm: The Grey Wolf Optimizer (GWO) is a metaheuristic optimization algorithm inspired by the pack hunting behavior of grey wolves. It simulates the social hierarchy and hunting strategies of grey wolves in nature and is one of the classic intelligent algorithms for solving complex optimization problems.

[0014] With its hierarchical guided convergence mechanism, this algorithm combines excellent local mining accuracy with iterative convergence speed. It does not rely on a large number of parameter adjustments and is widely used in fields such as function optimization, feature selection, parameter optimization, and engineering optimization. It is especially suitable for complex optimization scenarios that require precise location of the optimal solution and high convergence efficiency.

[0015] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0016] In a first aspect, embodiments of this application provide a method for fitting point clouds of transmission lines.

[0017] In one embodiment, reference is made to Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the point cloud fitting method for transmission lines in this application. Figure 1 As shown, the point cloud fitting methods for transmission lines include: S100: Extract the geometric features corresponding to the point cloud data of the transmission line and the semantic features corresponding to the image data of the transmission line, and perform feature fusion on the geometric features and semantic features to obtain the multimodal fusion features corresponding to each point cloud. S200: Based on the multimodal fusion features, a composite swarm intelligent optimization algorithm is used to mine high-confidence in-point regions in the transmission line point cloud, and sampling weights corresponding to each high-confidence in-point region are generated. S300: Guided by the sampling weights, perform weighted non-uniform sampling on the sampling process of the random sampling consensus algorithm to obtain the target sampling point set; S400: Based on the target sampling point set, fit the geometric model corresponding to the transmission line component to complete the point cloud fitting of the transmission line.

[0018] In this embodiment, by fusing point cloud geometric features and image semantic features, the information limitations of single point cloud features are overcome, providing comprehensive feature support for interior point region mining. The composite swarm intelligent optimization algorithm accurately locates high-confidence interior point regions and generates sampling weights, thereby guiding the RANSAC algorithm to perform weighted non-uniform sampling, replacing the traditional pure random sampling method. This fundamentally reduces the probability of outliers and invalid points participating in model construction, thereby reducing the number of invalid iterations, improving fitting efficiency, and ensuring the accuracy of the fitted model, meeting the dual requirements of fitting accuracy and processing efficiency in power transmission line inspection scenarios.

[0019] In one implementation, step S300 includes the following steps: S301: Allocate the number of samplings for each region according to the sampling weight of each high-confidence in-point region; S302: Within each high-confidence inner point region, perform sampling within the region based on spatial grid division to obtain a preliminary set of sampling points; S303: Perform semantic and geometric feature verification on the preliminary sample point set, eliminate potential outliers, and obtain the final target sample point set.

[0020] In this embodiment, the sampling frequency of each region is allocated based on the sampling weight, so as to tilt the sampling resources towards the high confidence inlier region and avoid the ineffective waste of sampling resources; the sampling within the region is performed by dividing the space grid to prevent the sampling points from being over-concentrated and to ensure the rationality of the spatial distribution of the sampling points; and potential outliers are eliminated a second time through semantic verification and geometric feature verification, which further improves the purity of the inlier of the target sampling point set and provides a high-quality sampling foundation for subsequent model fitting.

[0021] In one implementation, step S400 includes the following steps: S401: Based on the multimodal features of the high-confidence in-point region, determine the corresponding transmission line component type and match the geometric model that is compatible with the component's structural characteristics; S402: Fit the initial parameters of the corresponding geometric model based on the target sampling point set, and calculate the spatial distance from each point in the point cloud to the initial geometric model; S403: Calculate the dynamic inlier determination threshold based on the point cloud distance distribution characteristics, and perform inlier filtering by combining semantic verification and geometric feature verification to obtain a pure inlier set; S404: Optimizes the parameters of the geometric model based on the pure interior point set. After model validation, it outputs the optimal geometric model and fitting results.

[0022] In this embodiment, the type of transmission line component is determined based on multimodal features and a suitable geometric model is matched to avoid fitting deviations caused by a single geometric model, thereby improving the adaptability of the model to the actual component structure of the transmission line. By dynamically calculating the interior point determination threshold, it can adaptively adapt to point cloud data with different noise levels. Combined with multi-layer interior point verification of semantic and geometric features, the purity of the interior point set is ensured. Based on the pure interior point set, the model parameters are optimized to further improve the accuracy of the fitting model and output the optimal fitting result that adapts to the characteristics of the transmission line component.

[0023] In one implementation, in S404, the parameters of the geometry model optimized based on the pure interior point set include the following steps: S404-1: When the proportion of interior points and the fitting error of the model validation do not meet the preset requirements, start the iterative optimization of the geometric model parameters; S404-2: If the requirements are still not met after iterative optimization, the sampling strategy is dynamically adjusted, the sampling point set is optimized, and the model fitting and parameter optimization are re-executed.

[0024] In this embodiment, for cases where model validation fails to meet the standards, the model parameters are first optimized through iterative optimization of the interior point set to improve the model fitting accuracy. If the parameter iteration still fails to meet the requirements, the sampling strategy is adjusted from the sampling source to optimize the quality of the sampling point set, forming a two-layer optimization mechanism of model parameters and sampling strategy to ensure that the final fitting result can meet the preset accuracy requirements.

[0025] In one implementation, in S401, the geometric model that matches and adapts to the structural characteristics of the component includes: S401-1: When the component type is determined to be a conductor, match the catenary geometry model; S401-2: When the component type is determined to be a tower, match the geometric model of the combination of a cylindrical element and multiple straight lines. S401-3: When the component type is determined to be an insulator, match the geometric model of multiple cylinders connected in series.

[0026] In this embodiment, a dedicated geometric model is matched to the structural characteristics of different core components of the transmission line, which accurately adapts to the catenary sag characteristics of the conductor, the combined structure characteristics of the column and crossarm of the tower, and the multi-porcelain disc series structure characteristics of the insulator. This avoids fitting deviations at the model level and significantly improves the fitting accuracy of various transmission line components.

[0027] In one embodiment, S405 further includes the following steps in S400: S405-1: When the proportion of interior points and the fitting error of the model validation do not meet the preset requirements, start the iterative optimization of the geometric model parameters; S405-2: If the requirements are still not met after iterative optimization, the sampling strategy is dynamically adjusted, the sampling point set is optimized, and the model fitting and parameter optimization are re-executed.

[0028] In this embodiment, an iterative optimization and dynamic adjustment mechanism is set up for the model fitting process. For scenarios where the fitting accuracy is not up to standard, the model parameters are first corrected through iterative optimization. If the parameter optimization still fails to meet the standard, the sampling strategy is adjusted in reverse to optimize the input data quality from the sampling source, forming a closed-loop optimization logic to ensure that the final fitting result can meet the accuracy requirements.

[0029] In one implementation, step S200 includes the following steps: S201: Initialize the population parameters and iteration parameters of the composite swarm intelligent optimization algorithm. The composite swarm intelligent optimization algorithm integrates the global exploration capability of the octopus optimization algorithm and the hierarchical guided convergence capability of the gray wolf optimization algorithm to complete the generation of the initial population and the group division of hunters and scouts. S202: Construct a fitness function for point mining within the adaptive point cloud, calculate the fitness value of each initial individual in the population based on the multimodal fusion features, and determine the initial hierarchical guiding individuals; S203: Enter the iteration process. For each hunter, calculate the relative distance between the tentacles and compare it with a preset proximity threshold. Based on the comparison results, dynamically switch between the global exploration stage and the local utilization stage. Use the corresponding optimization algorithm to update the tentacles' positions and simultaneously update the hunter's head position. S204: Perform global exploration update for the scout group. Based on the fitness comparison results between the scouts and the corresponding reference hunters, complete the population conversion of scouts to hunters and the initialization of new hunter individuals. S205: After each iteration, update the global hierarchical guiding individual and iteration parameters, check the iteration termination condition, and if the termination condition is not met, proceed to the next iteration; if the termination condition is met, stop the iteration. S206: After the iteration terminates, a set of high-confidence in-point regions is obtained by filtering based on the fitness values ​​of individuals in the population. Based on the fitness values ​​corresponding to each high-confidence in-point region, the sampling weights corresponding to each region are calculated and generated.

[0030] In this embodiment, a composite swarm intelligence optimization algorithm combining the Octopus Optimization Algorithm and the Gray Wolf Optimization Algorithm is adopted. By combining multimodal fusion features, the algorithm achieves accurate mining of high-confidence in-point regions. After initializing the basic configuration of the population and parameters, the fitness function adapted to the mining of in-point clouds is used as the optimization criterion. Through the dual-group collaborative iteration of hunters and scouts, combined with the dynamic switching mechanism of global exploration and local utilization, the algorithm balances the global search capability and local convergence accuracy. Finally, high-confidence in-point regions are accurately selected and corresponding sampling weights are generated, providing accurate guidance for the subsequent sampling optimization of the RANSAC algorithm.

[0031] In one implementation, step S202 includes the following step: constructing the fitness function for adapting point cloud point mining includes: S202-1: Construct a single-objective fitness function that integrates spatial density, geometric consistency, and historical reward dimensions, set differentiated weighting coefficients for each dimension, and add semantic mask hard constraint rules.

[0032] In this embodiment, the constructed fitness function comprehensively evaluates the probability of a candidate region being an interior point region from three dimensions: spatial density, geometric consistency, and historical reward. It highlights the core judgment role of geometric consistency and directly eliminates background interference regions through semantic mask hard constraints, ensuring the adaptability of fitness evaluation results to the interior point characteristics of transmission line point clouds. This provides an accurate and reliable evaluation basis for the iterative optimization of composite swarm intelligent optimization algorithms.

[0033] In one implementation, step S203 includes the following steps: S203-1: When the relative distance between the tentacles is less than the proximity threshold, it is determined that the local utilization stage has been entered, and the position update of the tentacles is dominated by the three-level hierarchical guidance mechanism of the Grey Wolf Optimization Algorithm. S203-2: When the relative distance between the tentacles is greater than or equal to the proximity threshold, the global exploration phase is determined, and the position update of the tentacles is dominated by the octopus optimization algorithm for tentacles and the Levy flight mechanism. S203-3: After completing the tentacle position update, the head position of the corresponding hunter individual is updated simultaneously.

[0034] In this embodiment, the dynamic switching between the global exploration phase and the local utilization phase is achieved by comparing the relative distance of the tentacles with the proximity threshold. When approaching a high-potential area, the hierarchical guidance mechanism of the gray wolf optimization algorithm enables precise local mining, while when far from a high-potential area, the octopus optimization algorithm enables large-scale global exploration. This effectively balances the algorithm's global exploration capability and local utilization capability, avoiding the algorithm from getting stuck in local optima and improving the algorithm's convergence speed and optimization accuracy.

[0035] In one implementation, step S204 includes the following steps: S204-1: Select a reference position from the head position of the hunter group, and perform the position update of the scout based on the global exploration logic of the octopus optimization algorithm; S204-2: Calculate the fitness value of the updated scout and compare it with the fitness value of the referenced hunter head position; S204-3: When a scout's fitness value is better than the head position fitness value of its reference hunter, convert the scout into a new hunter individual and complete the head position update and tentacle regeneration of the new hunter. S204-4: When a scout's fitness value is not better than the fitness value of the hunter's head position it references, the scout's global exploration attribute is retained, and no population conversion operation is performed.

[0036] In this embodiment, global exploration is performed by a group of scouts to maintain population diversity and prevent the algorithm from missing potential high-confidence in-point areas. By comparing the fitness of scouts with that of reference hunters, the results of high-quality exploration are solidified and the population is transformed. The better areas discovered by scouts are quickly incorporated into the fine mining range of the hunter group, which improves the mining efficiency of high-quality areas while ensuring the coverage of global exploration.

[0037] In one implementation, step S100 includes the following steps: S101: The geometric features corresponding to the point cloud data of the transmission line and the semantic features corresponding to the image data of the transmission line are extracted. Feature fusion is performed on the geometric features and semantic features to obtain the multimodal fusion features corresponding to each point cloud, including: S102: Extract geometric features from the preprocessed point cloud data point by point, and perform normalization processing on the geometric features; S103: Perform semantic segmentation on the preprocessed image data to generate a corresponding semantic mask, extract semantic features, perform normalization processing on the semantic features, and map the semantic features to the corresponding point cloud based on the registration relationship; S104: Simultaneously extract the visual features of the image data, perform dimensionality reduction and normalization processing on the visual features, and then map them to the corresponding point cloud; S105: Set differentiated fusion weights based on the contribution of various features to the fitting, and perform weighted splicing fusion on the normalized geometric features, semantic features and visual features to generate multimodal fusion features corresponding to each point cloud.

[0038] In this embodiment, geometric features of point clouds, semantic features of images, and visual features are extracted simultaneously to achieve complementarity of multi-dimensional features. Normalization is used to eliminate the dimensional differences between different features, and then weighted splicing and fusion are performed based on the feature contribution to generate multimodal fusion features with geometric, semantic, and visual attributes. This comprehensively characterizes the properties of point clouds and provides rich and reliable feature support for the accurate mining of high-confidence in-point regions.

[0039] In one implementation, after S103, the following steps are included: S103-1: The neighborhood majority voting method is used to perform spatial filtering and smoothing on the semantic labels mapped to the point cloud to correct the semantic label noise caused by the registration error.

[0040] In this embodiment, the semantic labels of the mapped point cloud are spatially filtered and smoothed by the neighborhood majority voting method, which can effectively suppress semantic label noise caused by registration error, correct the biased semantic labels, improve the accuracy of point cloud semantic labels, and ensure the reliability of subsequent feature fusion and interior point region mining.

[0041] In one implementation, prior to S100, the following steps are included: S000: Acquire the 3D point cloud data and 2D image data corresponding to the transmission line inspection, complete the spatial registration and preprocessing of the two types of data, and obtain the registered and aligned point cloud data and image data.

[0042] In this embodiment, point cloud and image data of power transmission line inspection are acquired synchronously, and spatial coordinates of the two types of data are aligned through spatial registration. Then, data noise is removed and core feature information is extracted through preprocessing, providing spatially aligned and reliable basic data for subsequent feature fusion and algorithm iteration.

[0043] In one embodiment, after S400, S500 is further included, which includes the following steps: S501: Feed the fitted set of interior points, the proportion of interior points, and the fitting error results back to the composite swarm intelligent optimization algorithm. S502: Based on the feedback results, classify and label the high-confidence in-point regions and adjust the sampling weights of each region accordingly; S503: Based on the adjusted sampling weights, perform a new round of sampling and fitting optimization to complete the closed-loop feedback iteration.

[0044] In this embodiment, the front-end fitting results are fed back to the composite swarm intelligent optimization algorithm. Based on the fitting effect, the high-confidence in-point regions are classified, labeled, and their weights are adjusted to achieve iterative optimization of the sampling weights, forming a closed-loop mechanism of "sampling-fitting-feedback-optimization". Without increasing the computational power consumption, the accuracy of the sampling strategy is continuously improved, and the subsequent fitting accuracy and efficiency are further optimized.

[0045] Secondly, embodiments of this application provide a method for fitting point clouds of transmission lines based on multimodal fusion and the OOA-GWO-RANSAC algorithm.

[0046] In the current field of point cloud fitting for transmission lines, the existing technology that is closest to the technical logic of this solution is the point cloud fitting scheme of depth map-point cloud multimodal fusion. The overall process covers four key links: data acquisition and registration, multimodal feature preprocessing, core calculation of point cloud fitting, and result verification.

[0047] The first step is data acquisition and registration. Existing technologies typically use drones equipped with LiDAR and depth cameras to simultaneously acquire 3D point cloud data of power transmission lines and depth images of the corresponding scene. In the data preprocessing stage, a one-to-one correspondence between depth image pixels and 3D coordinates of point clouds is established using camera intrinsic and extrinsic parameter matrices and feature matching algorithms, completing the spatial registration of multimodal data and laying the foundation for subsequent feature fusion.

[0048] Secondly, there is multimodal feature preprocessing. For depth images, existing solutions extract semantic features of transmission line-related areas using semantic segmentation algorithms and mark non-target areas to assist in point cloud screening. For point cloud data, statistical filtering and voxel grid downsampling are used to remove noise points and redundant data. Some solutions further calculate local geometric features of the point cloud, such as normal vectors and curvature, to initially screen out point sets that may belong to the target.

[0049] Next comes the core step of point cloud fitting. Some techniques employ the RANSAC algorithm or its improved versions to fit transmission lines based on preprocessed data. Taking conductor fitting as an example, candidate regions for conductors are identified using semantic labels from depth images. Then, the RANSAC algorithm is used to randomly sample point clouds within these candidate regions, fitting straight or curved models. Inner points are selected to complete component morphology reconstruction. Some solutions introduce simple optimization strategies to adjust the RANSAC sampling range.

[0050] Finally, the results are verified. Existing technologies evaluate fitting performance by calculating metrics such as the distance error between the fitted model and the original point cloud, the proportion of inliers, and processing time. At the same time, they are compared with traditional single RANSAC algorithms and pure point cloud geometric fitting methods to verify the effect of multimodal fusion on improving fitting accuracy and anti-interference ability.

[0051] However, the existing technology has the following three technical problems: 1. The sampling strategy is highly blind and has low fitting efficiency. The point clouds of transmission lines generally have high noise and a high proportion of outliers. The existing RANSAC algorithm uses a purely random sampling method to select fitting primitives, which easily selects outliers or invalid points to participate in model construction, resulting in a high model fitting error rate. Multiple iterations are required to obtain an effective model, which significantly increases the fitting time. At the same time, its sampling logic does not target the local features of the transmission line to select the point set, which further exacerbates the inefficiency of sampling.

[0052] 2. Lack of model adaptation and dynamic adjustment capabilities Existing RANSAC and its improved versions mostly use single geometric models such as straight lines and simple curves to fit transmission line components. However, transmission line conductors actually have a catenary shape, and single models have poor adaptability and are prone to fitting bias. Furthermore, its sampling range, fitting threshold, and other parameters are statically set and cannot be dynamically adjusted according to the distribution of interior / exterior points during the fitting process, making it difficult to adapt to the dynamic characteristics of point cloud data in complex scenarios.

[0053] 3. Inherent technical problems of OOA algorithms The OOA algorithm has significant inherent flaws. Its core exploration mechanism relies excessively on Levy flight for global search, lacking precise local optimization guidance logic. Individual updates are based solely on the "optimal individual (prey)," leading to ambiguous convergence directions towards the optimal region in later iterations. This results in "oscillations around the optimal region," causing slow convergence and significant fluctuations in the final solution's accuracy. Furthermore, this single fitness guide source is weak against interference; if the optimal individual is affected by noise and deviates, the entire population's update direction will deviate from the optimal solution, resulting in insufficient stability and robustness. In addition, OOA lacks a mechanism for dynamically adjusting "global exploration" and "local utilization." Its excessive reliance on Levy flight for global search in the early stages wastes computational resources in ineffective regions, while the lack of targeted local utilization strategies in later stages makes it prone to prematurely falling into local optima. It fails to achieve an efficient balance between "search range" and "solution accuracy," limiting overall optimization performance.

[0054] This application addresses the shortcomings of existing algorithms for power transmission line point cloud fitting by proposing a method based on multimodal fusion and the OOA-GWO-RANSAC algorithm. First, a drone synchronously collects power transmission line point cloud data and corresponding image data. After multimodal data registration, semantic features from the images and geometric features from the point cloud are extracted and fused into a unified feature vector. Then, the global exploration capability of OOA is deeply integrated with the α / β / δ hierarchical guidance and dynamic convergence adjustment mechanism of GWO to construct the OOA-GWO composite algorithm. Based on the multimodal fusion features, high-potential inlier regions in the point cloud are mined and sampling weights are output. Subsequently, these weights guide RANSAC to perform weighted non-uniform intelligent sampling, replacing traditional pure random sampling, and the target model is fitted by combining the structural characteristics of power transmission line components. Simultaneously, a closed-loop mechanism of "sampling, fitting, and inlier feedback" is introduced to continuously optimize the fitness function and sampling weights of OOA-GWO, ultimately achieving high-precision, high-efficiency, and robust fitting of power transmission line point clouds, meeting the core data processing requirements of smart grid inspection.

[0055] Specifically, refer to Figure 2 , Figure 2 This is a flowchart illustrating the second embodiment of the point cloud fitting method for transmission lines in this application. Figure 2 As shown, the point cloud fitting methods for transmission lines include: Step 1: Multimodal Data Acquisition and Preprocessing 1. Synchronous acquisition of multimodal data A drone inspection platform equipped with LiDAR and a high-definition RGB camera is used to conduct multimodal data synchronous acquisition across the entire transmission line area, balancing data accuracy and engineering feasibility. The LiDAR equipment is used to collect 3D point cloud data of the transmission line. The core acquisition targets include key components such as conductors, towers, and insulators, as well as surrounding environmental information such as trees and terrain. The point cloud density for key components such as insulators and hardware is no less than 50 points / cm², ensuring complete representation of fine component details. The overall point cloud density of the transmission line corridor is no less than 20 points / cm², adapting to the actual needs of large-scale inspection scenarios. The "high-density acquisition of key components and adaptive acquisition of the entire corridor" setting in this solution ensures subsequent fitting accuracy while avoiding excessive costs and data redundancy issues associated with high-density acquisition across the entire area. It is also noted that the above density represents the ideal condition for optimal fitting results; in practice, point cloud data of different accuracies can be adapted based on equipment performance and engineering requirements.

[0056] A high-definition RGB camera synchronously acquires two-dimensional image data of the corresponding scene with a resolution of no less than 4K (3840×2160). To solve the synchronization problem caused by the difference in sampling mechanisms between the LiDAR and the camera, hardware trigger signals are used to achieve hard synchronization between the two, ensuring that the acquisition timestamps of the LiDAR point cloud frame and the image frame are accurately aligned with a synchronization error of less than 10 milliseconds, thus avoiding subsequent registration deviations caused by time asynchrony.

[0057] During data acquisition, real-time spatiotemporal coordinates were recorded using the drone's built-in GPS / IMU module, adding a unified timestamp to the point cloud and image data. Extrinsic parameter calibration of the LiDAR and camera was completed in advance, using a checkerboard calibration method to obtain rotation matrices and translation vectors, ensuring spatial alignment of their acquisition ranges. The acquisition route was planned according to the power transmission line's direction, employing a "round-trip coverage" mode with an overlap rate of no less than 15% between adjacent acquisition areas to avoid data loss. Simultaneously, acquisition operating information was recorded to ensure the dataset covers different operating scenarios, providing data support for subsequent anti-interference fitting. After acquisition, point cloud data was stored in .pcd format, and image data in .png format, synchronously saving spatiotemporal metadata such as coordinates, timestamps, and device parameters for easy subsequent processing and retrieval.

[0058] 2. Data Registration Data registration is based on spatiotemporal metadata such as timestamps and GPS / IMU coordinates collected by UAVs. It achieves high-precision spatial registration between images and multimodal point cloud data through a two-step method of "coarse registration + fine registration". At the same time, a robust optimization strategy is added to solve the problem of insufficient feature point matching in complex scenes, ensuring a one-to-one correspondence between 2D image pixels and 3D point cloud coordinates. (1) Coarse registration: The extrinsic parameters of the LiDAR-camera system, which were previously calibrated, are called, including the rotation matrix R and the translation vector T. Based on the pinhole camera model, the pixel coordinate system of the RGB image is transformed to the three-dimensional world coordinate system of the LiDAR. The formula is as follows:

[0059] in, For the three-dimensional coordinates of the point cloud, For the camera intrinsic parameter matrix, Image pixel coordinates, This is the depth value measured by the LiDAR. At this point, initial spatial alignment is complete, with the coarse registration error controlled within 5cm. For scenes with sparse features such as sky background and repetitive line structures, a ray-based coarse registration is simultaneously adopted as a backup solution. Candidate matching points are selected by the intersection of image pixel rays and the spatial range of the point cloud, avoiding the failure of a single coarse registration scheme.

[0060] (2) Fine registration: To address the residual error of coarse registration, a fusion algorithm of "feature point matching + geometric constraints + ICP (iterative nearest point)" is used for optimization. First, ORB feature points are extracted from the RGB image, and matching point pairs are initially screened through epipolar constraints to eliminate obvious mismatches. Then, known geometric dimensions such as towers and insulators are introduced as additional constraints to perform a second screening of matching point pairs to improve matching accuracy. Finally, using the coarse registration result as the initial value, the Euclidean distance error of the matching point pairs is iteratively minimized to update the rotation matrix. With translation vector The iteration continues until convergence, i.e., the residual is less than the set threshold or the set maximum number of iterations is reached, further reducing the registration error.

[0061] (3) Registration accuracy verification: Randomly select 100 sets of feature point pairs to verify the registration effect. The pixel reprojection error is required to be ≤1 pixel and the 3D point cloud registration error is required to be ≤2cm. If the standard is not met, the fine registration process is re-executed to ensure the consistency of multimodal data space and lay the foundation for subsequent feature fusion.

[0062] 3. Point cloud preprocessing To address the issues of outlier noise, isolated points, data redundancy, and missing features in the original transmission line point clouds, a three-step point cloud preprocessing process is employed: noise removal, voxel downsampling, and geometric feature extraction. This process removes invalid data and reduces the data volume while preserving the geometric structural features of core components such as conductors, towers, and insulators. This lays the data foundation for subsequent multimodal feature fusion and intelligent fitting. The specific process is as follows: (1) Noise Removal: Combining noise types such as equipment acquisition errors of transmission line point cloud, isolated points of environmental interference, and outliers such as trees / birds, a combination of statistical filtering and radius filtering is used for noise removal. First, the mean and standard deviation of the Euclidean distance between each point and its 20 neighboring points are calculated by statistical filtering, and outliers with a distance mean exceeding 3 times the standard deviation are removed to eliminate noise caused by random equipment errors; then, a radius filter is set with a search radius of 5cm to remove isolated points with fewer than 5 points in the neighborhood, and remove environmental interference points such as tree branches and leaves and birds, to ensure that the remaining point cloud is a valid set of points related to the transmission line.

[0063] (2) Voxel downsampling: To solve the problem of large amount of original point cloud data and high computing power consumption, voxel grid downsampling is used to simplify the denoised point cloud. The voxel grid size is set to 1cm×1cm×1cm to adapt to the structural features of fine components such as transmission line conductors and tower hardware. All point clouds in each voxel are replaced with the three-dimensional coordinates of the voxel center, with a compression ratio of about 70%. While significantly reducing the amount of point cloud data, the overall geometric shape and key structural details of the transmission line components are completely preserved, balancing data processing efficiency and feature preservation.

[0064] (3) Geometric feature extraction: Based on the K-nearest neighbor algorithm, three core geometric features are calculated point by point in the downsampled point cloud: normal vector, local curvature, and neighborhood spatial density. The normal vector reflects the orientation of the local surface of the point cloud, the curvature represents the degree of bending of the local surface, and the neighborhood spatial density is the number of points in a unit volume. At the same time, the three features are subjected to Min-Max normalization to map the feature values ​​to the interval [0, 1] to eliminate the difference in dimensions.

[0065] (4) Data regularization: The preprocessed point cloud 3D coordinates are associated with the normalized geometric features and stored in a standardized .pcd format. Feature attributes such as normal vector, curvature, and neighborhood density are added to each point cloud to facilitate direct call for subsequent multimodal feature fusion and fitness function calculation of the OOA-GWO composite algorithm.

[0066] 4. Image preprocessing To address the issues of uneven lighting, background noise interference, and low contrast between the target and background in high-definition RGB images acquired by drones outdoors, this study focuses on extracting semantic features of power transmission line targets, preserving key visual features, and achieving dimensional unification with point cloud features. Image preprocessing is completed through a process of basic denoising enhancement, feature point extraction, semantic segmentation, and feature quantization normalization. Furthermore, a fault-tolerant mechanism is added to address the issue of "semantic label mapping depending on registration accuracy." The extracted semantic and visual features are then fused with the geometric features of the point cloud to support subsequent high-potential interior point region mining. The specific process is as follows: (1) Image basic denoising and enhancement: First, 5×5 Gaussian filtering is used to remove random Gaussian noise in the image, and then bilateral filtering is used for secondary denoising. While eliminating noise, the edge features of target components such as conductors, towers, and insulators are preserved to avoid feature blurring caused by filtering. For problems such as uneven lighting, backlighting, and shadows in outdoor acquisition, adaptive histogram equalization (CLAHE) is used to improve the overall contrast of the image. Combined with brightness contrast adaptive adjustment, the distinction between the transmission line target and the background such as trees, sky, and terrain is strengthened, laying the foundation for subsequent semantic segmentation and feature extraction.

[0067] (2) Extraction of key visual feature points: The ORB feature extraction algorithm is used to extract key feature points and corresponding feature descriptors in the image. On the one hand, the accuracy of the previous image-point cloud data registration can be verified by feature point matching, and the registration error can be made up. On the other hand, the extracted feature descriptors, as image visual features, will participate in multimodal feature fusion together with semantic features to improve the richness of features.

[0068] (3) Semantic segmentation and mask generation of transmission lines: Based on the lightweight semantic segmentation network (UNet / SegFormer), the model was fine-tuned and trained using a labeled dataset of transmission line inspection scenes. Pixel-level semantic segmentation was performed on the denoised and enhanced image to accurately segment the target areas of the transmission lines, such as conductors, towers, and insulators, as well as non-target areas such as trees, sky, terrain, and buildings. After segmentation, a multi-class semantic mask with the same size as the original image was generated. Each pixel was assigned a semantic label. The pixel values ​​of different target areas were assigned corresponding values ​​of 1, 2, 3, etc. according to their categories, while the pixel values ​​of non-target areas were 0. At the same time, morphological closing operations were performed on the segmentation results to eliminate small holes in the segmented areas and optimize the integrity of the semantic mask.

[0069] (4) Image feature quantization and normalization: The semantic mask is converted into numerical semantic features. According to the registration relationship between the spatial coordinates of the image pixels and the three-dimensional coordinates of the point cloud, the corresponding point cloud position of each pixel is assigned a matching semantic feature value. At the same time, the extracted ORB feature descriptors are subjected to L2 normalization. All image feature values ​​are mapped to the [0, 1] interval through Min-Max normalization to achieve dimensional unification with the geometric features of point cloud normal vector, curvature, and spatial density, and to avoid fusion deviation caused by differences in feature numerical magnitude.

[0070] (5) Semantic Label Fault Tolerance Optimization: The mapping effect of semantic segmentation results is highly dependent on the registration accuracy between the image and the point cloud. Registration is to align data from different sources (such as the image and point cloud in this scheme) into the same coordinate system to ensure the correct spatial correspondence. Registration errors may cause positional deviations in semantic information, thus affecting the fusion of subsequent features. To improve the robustness of mapping and reduce noise introduced by registration errors, spatial filtering is performed on the semantic labels of the mapped point cloud. This scheme uses a majority voting method to analyze and smooth the semantic labels of each point in its 20 neighborhoods. If more than half of the points in the neighborhood have the target label, the target label of the point is retained; if the target label is less than half, the label of the point is corrected to a non-target label. This method effectively suppresses semantic noise caused by registration errors and enhances the accuracy of point cloud semantic labels, thus laying a good foundation for subsequent analysis and application.

[0071] (6) Data regularization and association: The processed high-definition image, binary semantic mask, normalized visual features and semantic features are associated with the preprocessed point cloud data one by one according to the timestamp, spatial coordinates and stored in a standardized format to ensure that each point cloud three-dimensional coordinate can be matched with the corresponding image pixel semantic features and visual features when multimodal feature fusion is performed in the future, so as to achieve accurate feature correspondence between point cloud and image.

[0072] Step 2: Multimodal Feature Fusion This step, based on the preprocessed data from step one, performs unified normalization verification and feature-level deep fusion on point cloud geometric features, image semantic features, and image visual features. It constructs a unified feature vector of point cloud and image that combines spatial geometric attributes, semantic label attributes, and visual feature attributes. This provides complete and standardized feature input for the subsequent OOA-GWO composite algorithm to mine high-potential inlier regions. The fusion process strictly follows the principles of "uniform dimension, consistent units, and appropriate weights" to meet the actual needs of point cloud fitting for transmission lines.

[0073] 1. Feature normalization (1) Feature Type and Dimension Analysis: Extract three types of feature sets corresponding to the three-dimensional coordinates of each point cloud, and clarify the feature dimensions and value ranges. The point cloud geometric features are 3-dimensional, namely the normal vector, local curvature, and neighborhood spatial density; the image semantic features are 1-dimensional, which is the normalized value of the binary semantic label; the image visual features are 1-dimensional. N The dimension is the core dimension of the ORB feature descriptor after L2 normalization. This scheme reduces it to 10 dimensions using the PCA method, balancing feature richness and computational efficiency.

[0074] (2) Unified normalization processing: Min-Max normalization is used for secondary verification of all features to strictly map all feature values ​​to the interval [0, 1]. The calculation formula is as follows:

[0075] in F norm These are the normalized eigenvalues. F These are the original eigenvalues. F min , F max These are the minimum and maximum values ​​of the feature in the corresponding feature set of the global point cloud, respectively.

[0076] (3) Outlier truncation and correction: Outlier feature values ​​that exceed the [0, 1] interval due to registration error and filtering deviation are truncated. Values ​​greater than 1 are taken as 1, and values ​​less than 0 are taken as 0, to avoid outlier values ​​from polluting subsequent fusion results; (4) Feature dimension regularization: The dimensions of the ORB visual feature descriptors corresponding to all point clouds are unified, and invalid features with inconsistent dimensions are eliminated to ensure that the visual feature dimensions of the same batch of data are completely identical, thus laying the dimensional foundation for feature fusion.

[0077] 2. Feature fusion A weighted splicing feature-level fusion strategy is adopted, which sets feature weights based on the core requirement of power transmission line point cloud fitting. The geometric features of the point cloud, the semantic features of the image, and the visual features of the image are spliced ​​into a unified feature vector to achieve deep fusion of multimodal features. The specific process is as follows: (1) Design and assignment of fusion weights: Based on the contribution of various features to the fitting of transmission line point clouds, differentiated weighting coefficients are set to satisfy the weight sum of 1, which highlights the leading role of core features and takes into account the supplementary role of auxiliary features. The specific weight allocation is as follows: point cloud geometric feature weight ω1=0.5, core feature, directly represents the geometric structure of transmission line components and is the basis for fitting; image semantic feature weight ω2=0.3, key filtering feature, quickly distinguishes target components from background interference areas; image visual feature weight ω3=0.2, auxiliary positioning feature, enhances the spatial visual recognition of target components; the weight coefficients can be finely adjusted according to the actual acquisition scene. When the occlusion is severe, ω2 can be increased, and when the point cloud geometric features are blurred, ω3 can be increased.

[0078] (2) Feature weighting enhancement: The standardized values ​​of each feature are multiplied by their corresponding weights to complete the feature weighting enhancement. The calculation formula is as follows:

[0079] in The weighted feature values ​​are used to achieve numerical enhancement of core features and reasonable representation of auxiliary features; (3) Construction of unified feature vector: Following the order of point cloud geometric features, image semantic features, and image visual features, all weighted feature values ​​are concatenated in one dimension to construct a unique multimodal unified feature vector for each point cloud. The vector dimension is 3+1+10=14 dimensions, and the vector expression is:

[0080] in, For point cloud geometric features, Image semantic features, Image visual features; (4) Feature vector regularization and associated storage: The constructed 14-dimensional multimodal feature vector is uniquely associated with the corresponding point cloud 3D coordinates and stored in a standardized matrix format. The rows are point cloud numbers and the columns are feature dimensions. Metadata such as point cloud 3D coordinates and spatiotemporal stamps are added to ensure that the subsequent OOA-GWO composite algorithm can directly call the corresponding feature vector according to the point cloud coordinates, while supporting fast retrieval and batch processing of feature vectors.

[0081] (5) Secondary verification of fusion results: Perform overall normalization verification on all constructed unified feature vectors to ensure that all feature values ​​in the vector are still in the range of [0, 1]. If there is a slight offset caused by weighted calculation, correct it again by Min-Max normalization to completely ensure the standardization of feature vectors.

[0082] After this fusion step, each transmission line point cloud is endowed with multimodal attributes that combine geometric structure, semantic labels, and visual features. This breaks through the information limitations of single point cloud geometric features and can provide comprehensive, rich, and standardized feature support for the subsequent OOA-GWO composite algorithm to accurately mine high-potential in-point regions. At the same time, the fused feature vector format is unified and can be directly adapted to the iterative calculation requirements of swarm intelligence algorithms without additional feature conversion processing.

[0083] Step 3: OOA-GWO Composite Optimization Algorithm This step uses the multimodal fusion feature vector output from step two as input to construct an OOA-GWO composite intelligent optimization algorithm, which deeply integrates the Octopus Optimization Algorithm (OOA) and the Grey Wolf Optimization Algorithm (GWO). This algorithm is used to accurately mine high-potential inlier regions from transmission line point clouds and output optimal sampling weights. It addresses the inherent shortcomings of traditional OOA, such as strong global exploration but weak local convergence, a single guiding source, and insufficient exploration-utilization balance. Simultaneously, it leverages the hierarchical guidance of GWO to improve convergence accuracy and algorithm stability. The algorithm optimizes by maximizing fitness in a single objective, with each individual corresponding to a candidate sampling region / sampling strategy in the point cloud.

[0084] 1. Algorithm parameter initialization (1) Definition of population and individual Population size: Set the population size N=50. This size can achieve a balance between the completeness of the point cloud parameter space search and the computational efficiency, avoiding both too few individuals leading to local optima and too many individuals increasing the iteration time.

[0085] Individual coding: Each individual X i ( i =1, 2,…, N The core descriptive parameters corresponding to a set of candidate sampling regions in the point cloud are encoded as follows: X i =[ x i1 , x i2 , x i3 , x i4 ]in x i1 , x i2 ,x i3 The coordinates of the three-dimensional spatial center of the candidate region are: x i4 The area is defined as the spatial search radius, which together uniquely determines a point cloud sampling sub-region.

[0086] Individual spatial boundaries: Based on the spatial bounding box of the point cloud of the entire transmission line, upper and lower limits are set for each dimension. x min , x max ]、[ y min , y max ]、[ z min , z max and radius range r min , r max All individual initializations and iterative updates are constrained within this legal range.

[0087] Initial population generation: Within the aforementioned spatial boundary, all N initial individuals are generated using a uniform random strategy and divided into hunters according to the following formula. N h With reconnaissance soldiers N s :

[0088] Each hunter consists of one head (core position) and eight tentacles (exploration branches), ensuring that the initial population is evenly distributed throughout the region and avoiding premature convergence caused by initial aggregation.

[0089] (2) Initialization of OOA core parameters OOA simulates the biological behavior of an octopus with multiple tentacles exploring, preying, moving, energy decaying, and experiencing random perturbations. It is the global exploration entity of the composite algorithm, and its native parameters are set according to the point cloud search characteristics: Number of tentacles M =8, corresponding to the octopus's original 8-tentacle structure. Each tentacle corresponds to an independent exploration branch, which is used to search different point cloud sub-regions in parallel, improving multi-directional exploration capabilities. Field of vision vr =3, used to constrain the tentacle's exploration radius. The formula is:

[0090] Approaching the threshold ll =0.8, used to determine the distance relationship between the tentacle and the optimal region; Levi's flight parameters b =1.5, controls the intensity of random jumps during global exploration, the formula is:

[0091]

[0092] in u , v ~ N (0, 1), that is, it follows a standard normal distribution, and Γ(·) is the gamma function; Exploration Factors α =0.3, which controls the intensity of an individual's tendency to move toward areas with high fitness, balancing following high-quality areas with autonomous random exploration; Exploration Factors λ =0.3, which controls the intensity of an individual's tendency to move toward areas with high fitness, balancing following high-quality areas with autonomous random exploration; random disturbance coefficient β =0.1, introduce a small random perturbation in the wrist update to prevent the algorithm from getting trapped in local optima.

[0093] (3) Initialization of GWO kernel parameters GWO Simulates Gray Wolf Population α , β , δ The behavior of the three-level leader individuals in guiding the hunting is used to compensate for the shortcomings of weak local convergence and single source of guidance in OOA, and to undertake the functions of local refinement and stable convergence.

[0094] Convergence factor a The initial value is 2, which decreases linearly to 0 with the number of iterations, dynamically controlling the switching between global exploration and local utilization. The formula is:

[0095] in, This represents the current iteration number. This represents the maximum number of iterations. Random coefficient vectors A and C: These are randomly generated in each iteration to simulate the random wandering and prey-encircling behavior of a wolf pack, enhancing the robustness of local search. The formula is as follows:

[0096]

[0097]

[0098]

[0099]

[0100]

[0101] in rand 1. rand 2. rand 3∈[0, 1], A Control the exploration area, | A |>1 strengthens exploration, | A When |<1, enhanced utilization is achieved, and C simulates the random distribution characteristics of high-potential in-point regions; Third-level guide individuals: In each iteration, the top three individuals with the highest fitness are selected and marked accordingly. α (Optimal high-potential area) β (Second-best high-potential area) δ (Third high-potential area) replaces the single best individual guidance of OOA, improving the stability of the update direction.

[0102] ④ Common Iteration Control Parameters Maximum number of iterations T max =100; Phase switching determination logic: relative distance of the tentacles based on OOA trans With threshold ll The relationship dynamically triggers the switching between the "local utilization / global exploration" phases; Fitness convergence threshold ε =10 5 If the change in optimal fitness is less than this value after 15 consecutive iterations, the iteration can be terminated early. Stable counter stable count =0 is used to count the number of iterations when the optimal solution is not improved.

[0103] 2. Fitness Function Construction (1) Fitness function design The design of the fitness function takes into account space density. D ( X i Geometric consistency G ( X i Historical Awards H ( X i The three core dimensions are integrated into a single-objective maximization function. The higher the value, the greater the probability that the corresponding area is a high-potential inlier region. This fully adapts to the single-objective optimization characteristics of OOA. The formula is as follows:

[0104] Among them, the spatial density weighting coefficient ω1=0.25, the geometric consistency weighting coefficient ω2=0.5, and the historical reward weighting coefficient ω3=0.25, satisfying ω1+ω2+ω3=1. ω2 is the largest, highlighting the core judgment role of geometric consistency. All dimensions are normalized and mapped to [0, 1] to ensure no interference from quantitative differences.

[0105] (2) Calculation logic for each dimension spatial density D ( X i ): Measures the density of point clouds within the candidate region. The point cloud density of target components of transmission lines is significantly higher than that of background noise. The formula is:

[0106] in, The point cloud density within the region is calculated by dividing the number of points by the region's volume. , The minimum and maximum densities of the global point cloud are given.

[0107] Geometric consistency G ( X i ): Core dimension, measuring the uniformity of point cloud normals and curvature within a region, adapting to regular geometric structures, the formula is:

[0108] in, The number of point clouds in the region. Let be the angle between the normal vector of point k and the average normal vector of the region. Let k be the local curvature. The mean curvature of the region. θmax is the global preset threshold.

[0109] Historical Awards Iterative feedback is introduced to assign rewards to regions with a historically high percentage of internal points, using the following formula:

[0110] in, This represents the number of times the region has been identified as a valid interior point in its historical iterations. This represents the maximum historical hit count across all regions.

[0111] (3) Post-processing constraints Semantic mask hard constraint: If the proportion of non-target point clouds in the region is ≥50%, directly mask the target point cloud. Set to 0 to force the exclusion of background interference; Outlier truncation: Calculation results exceeding [0, 1] are truncated, taking 0 for values ​​less than 0 and 1 for values ​​greater than 1, to ensure the validity of the function.

[0112] 3. OOA-GWO Iterative Optimization (1) Initial preparation Calculate the fitness value of each individual in the initial population and select... α , β , δ Three-level guiding entity, initializing the number of iterations. t =1. Stable counter stable count =0.

[0113] (2) Hunter's Wrist Dual Scene Update For each hunter, the relative distance between the tentacles is first calculated using a formula. trans And then according to trans and ll Different relationships trigger different update logic:

[0114] Scenario 1: When trans < ll When this occurs, it indicates that the tentacle has approached a high-potential intima region, similar to an octopus's tentacles sensing prey and entering a localized exploitation phase. Once this phase is determined, the GWO algorithm completely dominates the tentacle position update, because GWO's... α , β , δ Hierarchical guidance mechanisms are inherently suited to the need for "refined mining around known high-quality regions." However, target components such as conductors and towers in transmission line point clouds often exhibit continuous clusters of high-potential internal points. The algorithm needs to focus on these regions to reduce randomness and improve convergence accuracy. The specific update formula is as follows:

[0115] in, α , β , δ These point cloud regions, which correspond to the top three in global fitness, are the core regions most likely to be points within the target component in the transmission line. As hierarchical guidance sources for GWO, they can provide a clear convergence direction for the tentacle. Hunters i .Tgroup j This indicates the position of the j-th tentacle of the i-th octopus; C 1. C 2. C 3 represents the random coefficients regenerated in each iteration, used to simulate minute positional fluctuations in high-potential in-point regions, preventing the tentacles from getting stuck in absolutely fixed search paths; while A1. A 2. A 3 is a convergence factor that decreases linearly from 2 to 0 with each iteration. Its function is to dynamically shrink the search range in the early stages of iteration. A Small-scale exploration is allowed when the value is large, and later... A When the value approaches 0, the focus is on the core region. Additionally, in the formula, |·| represents taking the absolute value independently of each dimension of the vector, indicating that the tentacle is calculated separately along each axis in three-dimensional space. α , β , δ The distance ensures the comprehensiveness of the encirclement search.

[0116] After the tentacles are updated, the coordinates corresponding to the average fitness values ​​of all eight new tentacle positions of the hunter are taken and directly updated to the hunter's head position. Hunters i .head The reason for this design is that all the tentacles are in this stage. α , β , δ Guided by the algorithm, a refined search is conducted. While the position of a single tentacle may fluctuate locally, the mean position more stably reflects the hunter's core location in a high-potential area. This avoids head position oscillations caused by random deviations of a single tentacle, and also aligns with the core idea of ​​GWO population collaboration towards the elite center, allowing the hunter group to quickly converge towards the optimal inner point region. It should be noted that if the diversity advantage of OOA is to be preserved, the position with the highest fitness among the eight tentacles can be selected to update the head. This method retains the randomness of the local search and is suitable for scenarios where the features of target components in the transmission line point cloud exhibit subtle differences, but the convergence speed may be slightly slower.

[0117] Scenario 2: When trans>ll When the high-potential inlier region is outside the reach of the tentacle, the global search phase begins. Once the global exploration phase is initiated, the OOA algorithm takes full control. This is because the OOA's "head-centric diffusion + Levy flight" mechanism excels at covering large areas of unknown space. However, transmission line point clouds often have large spans and sparse point clouds in some areas, requiring the algorithm to possess strong spontaneous exploration capabilities to avoid missing scattered inlier points within target components. The specific update formula strictly follows the OOA design:

[0118] in, Hunters i .head The core position of the hunter is the optimal area that the hunter has explored in previous iterations. Expanding outwards from this center can ensure the continuity of exploration and avoid meaningless random jumps. (Huntersi.head Huntersi.Tgroupj) of The vector difference defines the propagation direction of the tentacles, ensuring that the tentacles explore the surrounding high-quality area corresponding to the head from their current position, rather than jumping around aimlessly; and rand Random numbers ∈ [0, 1] add a degree of randomness to the diffusion step size, preventing multiple tentacle exploration paths from completely overlapping. Most importantly, this relates to the Levi flight item. LF (dim), its parameters b =1.5 is the optimal configuration for OOA, which produces the characteristic of "mostly small-step exploration + occasional large-step leaps". This characteristic is well-suited to the power transmission line scenario. Small-step exploration can finely examine the adjacent areas around the head, while large-step leaps can cross background areas such as trees and sky to quickly reach the point cloud area of ​​the target component that may exist in the distance, effectively preventing the algorithm from getting stuck in local optima.

[0119] During the head update process, the position with the highest adaptability among the hunter's eight new tentacle positions is selected and updated as the hunter's head. Hunters i .head This is because the core objective of the global exploration phase is to discover new high-potential areas. Preserving the optimal tentacle position can quickly solidify valuable exploration results into the hunter's core positioning, providing a better starting point for the next iteration. At the same time, it can also increase the probability of discovering sparse interior point areas through parallel exploration of multiple tentacles.

[0120] (3) Reconnaissance soldier update Scouts play a crucial role in the algorithm by exploring overlooked areas and maintaining population diversity. Therefore, they must always maintain the purity of global exploration and not be restricted by the elite guidance of α, β, and δ in GWO. If GWO guidance is introduced, scouts may be attracted to high-potential areas that have already been discovered, thus losing their essential role of exploring the unknown.

[0121] Based on this, the scout's update strictly adopts the global exploration logic of OOA, selecting the total from the head positions of all hunters. Ns This set of locations must include one optimal hunter head, corresponding to the optimal high-potential in-line area already discovered within the transmission line; and one worst hunter head, corresponding to a low-potential area with poor exploration results. Ns - Two random hunter heads, corresponding to unexplored general areas. The optimal head reference allows the scout to explore the sparse point cloud at the edge around the core area of ​​the tower or conductor; the worst head reference guides the scout away from background noise areas; and the random head reference ensures that the exploration range covers easily overlooked areas such as the gaps between conductor spans and the sides of the tower.

[0122] Based on the above selection Ns The specific formula for updating the scout's position based on the hunter's head position is as follows:

[0123] in, Hunters z .head The selected head is the z-th hunter, which could be the best, worst, or a random hunter. If the best hunter head is selected, the scout will explore the edge of the area; if the worst hunter head is selected, the scout will explore in a direction away from the area. UB / LB The upper and lower bounds of the spatial bounding box of the point cloud of the transmission line. (4) Scouts turned hunters ① Triggering conditions Only when a certain reconnaissance soldier Scouts z Its fitness value is better than that of its reference hunter head position. Hunters z .head The conversion is only triggered when the scout discovers a high-potential in-point area that is of higher quality than the reference area, such as a denser conductor point cloud or a more complete tower structure point cloud, before it will be included in the hunter population to ensure the effectiveness of population iteration.

[0124] ② New Hunter Head Update After the conversion is triggered, the scout's position is directly assigned to the new hunter's head, using the following formula:

[0125] The new hunter's core positioning completely inherits the exploration results of the scout, without any additional adjustments. In the power transmission line scenario, this means that high-quality areas such as sparse point cloud areas of conductor joints and point cloud areas at the edges of tower crossarms discovered by the scout can directly become the new hunter's core exploration targets, quickly solidifying the exploration results.

[0126] ③ New Hunter Tentacle Regeneration Based on the new hunter's head position, eight new tentacles are generated, using the following formula:

[0127] in, Hunters z .Tgroup j Indicates the first z The hunter's first j The tentacles are positioned within a certain range around the head, allowing new hunters to quickly enter a localized, detailed phase without needing to re-explore the entire area, thus rapidly adapting to the species' optimized rhythm.

[0128] (5) Population iteration termination ①Global optimization After each round of tentacle and scout update, a population iteration cleanup is needed to lay the foundation for the next round of iteration. This step is crucial to ensuring the algorithm's continued convergence. First, a global elite update is performed: the positions of all hunter heads and scouts are collected, their fitness values ​​are calculated one by one, and then the top three individuals with the highest fitness are selected and updated to GWO. α , β , δ Guiding individuals. The core significance of this step is to ensure that the next round of local utilization phases can be guided based on the optimal exploration results of the current population, thus ensuring the accuracy of the guidance source and preventing the algorithm from backtracking.

[0129] ② Constraint boundary verification Secondly, boundary constraint verification is performed: the positions of all hunter tentacles, heads, and scouts are compared one by one with the spatial bounding box of the power transmission line point cloud. If the coordinates of a certain dimension exceed […], the boundary constraint verification is performed. x min , x max ]、[ y min , y max ]or[ z min , z max If the range is limited, the algorithm is truncated to the boundary value. This operation is to prevent the algorithm from exploring invalid areas beyond the actual inspection range, such as spaces where power transmission line components are unlikely to exist, such as the sky or below the ground, ensuring that all calculations are focused on the effective point cloud area and improving algorithm efficiency.

[0130] ③ Iteration count update and termination condition judgment Set the current iteration number t Increment by 1, return to the tentacle state determination stage, and simultaneously reset or accumulate the stability counter. If the optimal fitness in this round has not improved, stable count The value is incremented by one if necessary, otherwise reset to 0. Then the next iteration begins, forming a closed loop until the iteration termination condition is met, i.e., the maximum number of iterations is reached. T max The optimal fitness changes less than a set threshold multiple times consecutively. ε The final output is a set of high-potential interior point regions after multiple rounds of optimization.

[0131] 4. Output and Sampling Weight Calculation for High-Potential In-Point Regions (1) Screening of high-potential inland point areas After the iteration terminates, extract the population whose fitness values ​​satisfy... F ( X iPoint cloud regions corresponding to all individuals with a value ≥ 0.7 (an empirical threshold, which can be adjusted according to the scenario) are considered as a set of high-potential inlier regions. S ={ S 1, S 2, …, S k}, k This refers to the number of regions. Two additional conditions must be met during the filtering process: ①Region deduplication: If the spatial overlap rate of two regions is ≥80%, retain the region with higher fitness to avoid duplicate sampling; ② Semantic verification: The point cloud ratio of the target component in the final output area is ≥60%, ensuring no background noise interference.

[0132] (2) Calculation of sampling weights To improve the efficiency and accuracy of subsequent RANSAC fitting, sampling weights are calculated based on the fitness values ​​of each high-potential region. Higher weights indicate a greater probability of interior points in that region, and therefore require more sampling resources. The formula is as follows:

[0133] in, The sampling weight for the m-th high-potential region is... The fitness value for this region satisfies:

[0134] (3) Output format The final output consists of the following two parts: ① Set of high-potential in-point regions S: Each region is labeled with its 3D center coordinates, search radius, and target component type; ②Sampling weight table g Sort by weight from high to low to clarify the sampling proportion of each region, providing a basis for subsequent RANSAC stratified sampling.

[0135] Step 4: Intelligent Guided RANSAC Sampling and Point Cloud Fitting (GWO-OOA-RANSAC) This step uses the high-potential inlier region set S and the sampling weight table g output from step three as the core inputs, replacing the pure random sampling logic of traditional RANSAC. It combines the structural characteristics of transmission line components with a dedicated geometric model, and achieves high-precision fitting of point clouds through dynamic inlier selection and multi-dimensional verification, thus solving the pain points of blind sampling, poor model adaptability, and weak robustness of traditional RANSAC.

[0136] 1. Weighted Non-Uniform Intelligent Sampling (1) Sampling frequency allocation First, based on the sampling weights of high-potential areas w (S m ), allocate the number of samples to each region, and the total number of samples. K Set according to the point cloud scale of the transmission line, default. K =200, which can be adjusted according to the point cloud density. The formula for calculating the number of samples in each region is:

[0137] in, Let be the number of samplings for the m-th high-potential region, and [·] be the floor function to ensure that each high-potential region receives reasonable sampling resources.

[0138] (2) Point cloud sampling within the region In each high-potential area S m Within the sample, a method of "spatial grid division + random sampling" is used to avoid excessive concentration of sampling points.

[0139] Search radius by region r Divide the space into 1 / 5 grids to ensure that the point cloud covered by each grid has local geometric consistency; Count the number of target point clouds within each grid, i.e., those semantically labeled as target parts with geometric consistency. G For point clouds with a value ≥0.4, if there are ≥3 target points within the grid, then one point is randomly selected as the sampling point. If there are not enough grid target points in a certain area, supplement the sampling with the points with the highest geometric consistency in that area to ensure the effectiveness of the sampling points.

[0140] (3) Secondary screening of sampling points The initial sampled point cloud is then screened a second time to remove potential outliers.

[0141] Semantic validation: Sampling points with semantic labels of target components such as conductors, towers, and insulators are retained, while points without target labels are directly discarded; Geometric verification: Preserving local curvature c k Sampling points with a curvature threshold of ≤0.2 (curvature threshold for regular structures such as conductors and towers) and a normal vector variance of ≤0.1 are discarded for those with abnormal geometric features. Final sample set: Composed of the selected sample points. Psample ={ p 1, p 2, …, p K The proportion of interior points is increased by more than 40% compared to traditional random sampling.

[0142] 2. Transmission line component model adaptation For the structural characteristics of core components such as transmission line conductors, towers, and insulators, a dedicated geometric model is adapted to avoid the fitting deviation of traditional single linear and cylindrical models, and improve the matching degree between the model and the point cloud.

[0143] First, the component type is automatically determined based on the geometric features and multimodal semantic tags of high-potential areas. The semantic tag for a wire is "wire," the area is elongated with an aspect ratio ≥ 5:1, and exhibits geometric consistency. G ≥0.6; The semantic tag corresponding to poles and towers is "pole and tower", the area is columnar or frame-shaped, and the spatial density is... D ≥0.5, point cloud is dense; the semantic label corresponding to the insulator is "insulator", the region is short columnar or disk-shaped, the local curvature fluctuates greatly, and the disk-shaped structure features are present.

[0144] (1) Conductor model: Catenary model Traditional straight-line fitting cannot reflect the sag characteristics of a conductor; therefore, a catenary model is used, with the following formula:

[0145] in,( x 0, z 0) represents the center coordinates of the high-potential area. y The direction is the direction of the wire extension, and it is fixed during fitting; a These are parameters of the catenary, which are related to conductor tension and self-weight. b For offset; parameter initialization is based on the longitudinal span of high-potential areas. L With verticality h (Difference between upper and lower boundaries of the region), initially set as follows:

[0146]

[0147] (2) Tower model: Combined cylindrical model The tower consists of columns (circular cross-section) and crossarms (rectangular cross-section), each adapted to a specific model. The columns use a cylindrical model:

[0148] in, The coordinates of the cylinder center are... For the column radius, parameter initialization As the regional center, The radius is 1 / 2 of the region radius; the crossarm is a long beam structure, using a multi-segment straight line model, and the parameter initialization is based on the main direction vector of the point cloud within the region, setting the endpoints of the straight lines as high geometric consistency points on the edge of the region.

[0149] (3) Insulator model: Multi-cylinder combination model An insulator consists of multiple porcelain discs and metal connectors, using a multi-cylinder series model, where each cylinder corresponds to one porcelain disc or connector. The formula is as follows:

[0150] in, The number of insulator porcelain discs is determined based on semantic tags and point cloud density. Let be the center coordinates of the i-th cylinder, distributed along the insulator axis; r i The radius of the ceramic plate is initialized to 0.8 times the average radius of the region.

[0151] 3. Interior point selection and model validation (1) Interior point screening The core of interior point screening is to accurately distinguish target component points from noise and background points, avoid the misjudgment problem of traditional single distance judgment, and ensure that the screened interior point set can accurately support model fitting.

[0152] First, calculate the spatial distance from all points within the high-potential area to the "continuous geometric model with initial parameters." The distance calculation method for different component models must be strictly adapted to their structural characteristics. ① Conductor (catenment model): Because the conductor follows... y directional extension, point p =( x p , y p , z p The distance to the model is z Vertical distance, ignored y The small noise deviation in direction, i.e.: |

[0153] ② Tower (cylindrical model): The distance from the point to the model is the radial distance deviation, and the cylinder along... z Directional extension, only calculation x - y The difference between the distance from the center of the cylinder to the radius in the plane, i.e.:

[0154] ③ Insulator (multi-cylinder combination model): Calculate the radial distance from each point to each cylinder sequentially, and take the minimum value as the final distance. This is suitable for a multi-disc series structure, i.e.:

[0155] Secondly, calculate the dynamic threshold. τ First, collect distance data from all points within the high-potential region to the initial model, and then calculate the standard deviation of this dataset.σ This reflects the degree of dispersion of distance, i.e., the noise level, and the threshold is then calculated using the following formula:

[0156] σ The larger the value, the more noise there is; the threshold is automatically widened to avoid false rejection of inliers. σ A smaller threshold indicates a cleaner point cloud; tightening the threshold ensures the purity of the inlier points.

[0157] Next, perform a double interior point determination: ①Basic judgment: distance d ≤ τ Points that exceed the distance threshold are added to the candidate internal point set, and points that exceed the distance threshold are directly removed. ② Semantic validation: Candidate interior points must match the semantic tags of the component type; background points with mismatched semantic tags are removed. ③ Geometric feature enhancement verification: Adjust verification rules for different components; guide points must satisfy local curvature. c k ≤0.15 (curvature threshold of regular curve), normal vector variance ≤0.08; tower points must meet spatial density requirements. D ≥0.4 (dense characteristics of column point cloud), number of neighboring points ≥8; insulator points must meet the local curvature fluctuation range [0.3, 0.7] (porcelain disk concave-convex structure characteristics), and noise points with abnormal geometric features should be removed; ④ Secondary filtering of outliers: If there are "isolated points" in the candidate inlier set, they are identified as outliers and removed to avoid individual noise points affecting model fitting.

[0158] Finally, after multiple rounds of screening, a pure set of interior points was obtained. P in It also records the distance value, semantic label, and geometric feature parameters of each inlier, providing data support for subsequent model validation and parameter optimization.

[0159] (2) Model validation ①Verification of the proportion of interior points Calculate the percentage of interior points :

[0160] in, P region The total point cloud count for high-potential areas directly reflects the model's coverage capability of the target components. A differentiated threshold is set based on the point cloud characteristics of transmission line components.

[0161] wire: η ≥60%, the point cloud of the conductor is continuous and dense, and the proportion of interior points needs to be even higher; Tower: η≥55%, the point cloud on the pole and column is dense, but the point cloud on the edge of the crossarm may be sparse; insulator: η ≥45%, the insulator point cloud is relatively sparse, and there are a few blank spaces between the porcelain discs, so the threshold should be appropriately relaxed.

[0162] If the percentage of inliers is lower than the threshold, it indicates that the initial parameters of the model are too biased or the high-potential regions are not accurately selected. It is necessary to go back and readjust the initial parameters of the model or add high-potential regions.

[0163] ② Fitting error verification Average distance error This reflects the degree to which the interior points fit the model as a whole; the guide wire requirements... ≤0.02m, tower requirements ≤0.03m, insulator requirements ≤0.025m.

[0164]

[0165] Maximum distance error d max The limit value reflecting local deviation is required for all components. d max If there are in-line points exceeding the limit (≤0.05m), it is necessary to check whether they are noise points that have not been removed or poor local adaptability of model parameters.

[0166] ③ RANSAC parameter iteration If any of the above dimensions fails to meet the requirements, RANSAC parameter iterative optimization is initiated, with a maximum of 3 iterations. Each iteration executes the following complete closed loop: Model parameter update: using the currently selected pure interior point set P in Using the least squares method as input, the geometric model is refitted to obtain the updated optimal model parameters, such as those for a catenary. a , x 0、 b , tower cylinder c x , c y , r Replace the original or old parameters; Distance and Interior Point Recalculation: Based on the updated model, recalculate the spatial distances from all points within the high-potential region to the new model. d ( p , model new The dynamic threshold is recalculated according to the same rules. τ Perform interior point determination to obtain a new set of pure interior points. Pin ´ ; Model validation retest: using P in ´ Recalculate the percentage of interior points η Average error Maximum error d max Verify whether it meets the standards; Iteration termination criteria: If the verification is successful, terminate the iteration and output the current optimal model parameters; if the verification is unsuccessful but the number of iterations has not reached 3, repeat the RANSAC iteration.

[0167] (3) Dynamic adjustment of RANSAC sampling strategy If the validation fails after three parameter iterations, a full-dimensional adjustment of the RANSAC sampling strategy will be performed to optimize sampling quality from three aspects: sampling space distribution, point cloud selection rules, and sampling resource allocation. The specific adjustment details are as follows: ① Reduce the size of the spatial sampling grid: for the average error In scenarios where the limits are exceeded, the sampling grid within the region is reduced from 1 / 5 of the original region radius r to 1 / 6 of the region radius r, allowing the sampling points to focus more on sub-regions with stable local geometric features, thereby improving the matching degree between the sampling points and the model. ② Tighten the geometric feature screening threshold: based on the proportion of interior points η For scenarios with low values, the screening rules are tightened according to component type. The threshold for local curvature of conductors is tightened from ≤0.2 to ≤0.15, the threshold for spatial density of towers is increased from ≥0.4 to ≥0.45, and the threshold for curvature fluctuation of insulators is increased from ≥0.3 to ≥0.35. This removes more noise points with abnormal geometric features and improves the purity of sampling points. ③ Fine-tune the sampling weight allocation in the region: For scenarios where the fitting of a local region continues to fail, reduce the sampling weight of that region by 20%, and at the same time allocate the released sampling resources to the adjacent high-potential regions that have a good fit, thereby reducing the consumption of invalid sampling. If the verification still fails after adjustment, it is determined that the high-potential area itself has too much noise or missing target point cloud. Return to the high-potential area screening step in step three, re-optimize the OOA-GWO sampling weights or supplement the target component point cloud data to ensure that the fitting basis is reliable.

[0168] Step 5: Closed-loop feedback optimization (1) Feedback of interior point results After completing step four, convert the interior point set from step four. P in and interior point ratio η Average error Feedback is given to the OOA-GWO algorithm, if a certain high-potential region η≥60% and ≤0.02m indicates high point purity within this area, and it is marked as a "key exploration area"; if η <40% or A value >0.04m indicates that the area has high noise levels or poor model adaptation, and is marked as a "restricted exploration area".

[0169] (2) Iterative correction of OOA-GWO sampling weights Based on the feedback results, the sampling weights were adjusted, with only two core corrections made. The weight of the key exploration area was increased by 20%, and more sampling attempts were allocated in the next sampling; the weight of the restricted exploration area was reduced by 50% to reduce invalid sampling; the weights of other areas remained unchanged to ensure focused optimization and avoid redundancy.

[0170] Step Six: Outputting Fitting Results and Evaluating Performance (1) Output of optimal model parameters Conductor (catenline model): Output key parameters a , b , x 0 and the direction of wire extension ( y (axis range); Tower (cylindrical model): Output the coordinates of the tower center. c =( c x , c y , c z ),radius r and column height range; Insulator (multi-cylinder combination model): Outputs the center of each ceramic disc. c i ,radius r i and the total length of the insulator; Additional output: Pure interior point set P in It includes coordinates, semantic labels, and geometric features, which facilitates subsequent secondary processing.

[0171] (2) Performance index evaluation Key coverage metric: Percentage of in-point coverage η ≥50% is the basic qualification line, and different components are judged according to differentiated thresholds; Core accuracy indicator: Average distance error ≤0.03m is the general acceptable line; different components are judged according to different thresholds. Model fit: Calculation 1 / τ , τThe threshold is dynamic. A value ≥0.7 indicates that the model has a high degree of fit with the point cloud and can be directly used for subsequent scenarios such as line shape reconstruction and defect detection.

[0172] To verify the improvement effect of the OOA-GWO composite optimization algorithm over the traditional OOA algorithm, a standard test dataset (containing optimization objective functions of different dimensions and complexities) was selected. The traditional OOA algorithm and the single GWO algorithm were compared, with the core evaluation metrics including the accuracy of optimization objective mining. Number of iterations to reach a stable solution T Optimal solution fluctuation amplitude Each indicator focuses on the performance improvement of the algorithm itself, without being related to specific application scenarios. The calculation method is as follows:

[0173] To accurately identify the number of optimal solution regions; This represents the total number of candidate optimal solution regions output by the OOA-GWO algorithm.

[0174]

[0175] The number of independent tests is set to 20 here; This represents the optimal fitness value in the m-th test. The average of the optimal fitness from 20 tests; The smaller the value, the lower the fluctuation range and the stronger the algorithm stability.

[0176] The comparison results are shown in Table 1 below:

[0177] Table 1 As shown in the table above, the OOA-GWO composite algorithm significantly improves upon the traditional OOA algorithm. This optimizes the target discovery accuracy. The accuracy rate reached 89.47%, an improvement of 23.58 percentage points compared to traditional OOA and 14.23 percentage points compared to the single GWO algorithm. This was achieved by using GWO's... α , β , δThe hierarchical guidance mechanism integrated into OOA solves the problems of single guidance source and inaccurate optimal solution region mining in traditional OOA. The convergence speed requires only 42 iterations, which is 44 iterations (51.16%) less than traditional OOA and 21 iterations (33.33%) less than the single GWO algorithm. The dynamic switching mechanism of "global exploration + local utilization" greatly improves the algorithm's convergence efficiency. The fluctuation range of the optimal solution is only 2.13%, which is 6.63 percentage points lower than traditional OOA and 3.21 percentage points lower than the single GWO algorithm. The algorithm's stability is significantly enhanced, achieving a comprehensive improvement in the core performance of the OOA algorithm.

[0178] To verify the performance advantages of the OOA-GWO-RANSAC intelligent fitting algorithm in the point cloud fitting scenario of transmission lines, point cloud data of three core components—conductors, towers, and insulators—were selected. The algorithm was compared with traditional RANSAC, OOA-RANSAC, PSO (Particle Swarm Optimization)-RANSAC, and SA (Simulated Annealing)-RANSAC algorithms. The core evaluation index focused on the point cloud fitting effect, including the percentage of interior points. Average distance error Fitting time t The calculation method and scheme remain consistent, and the comparison results are shown in the table below:

[0179] Table 2 As shown in the table above, the OOA-GWO-RANSAC algorithm exhibits the best performance among various improved RANSAC algorithms. (Interior point percentage) The average accuracy reached 78.32%, an improvement of 41.57 percentage points compared to traditional RANSAC, 26.17 percentage points compared to OOA-RANSAC, 18.04 percentage points compared to PSO-RANSAC, and 19.61 percentage points compared to SA-RANSAC. The improved OOA-GWO algorithm provides more accurate sampling weights for RANSAC, significantly reducing the probability of outliers being included; the average distance error... The accuracy was controlled at 0.018m, which is 0.033m (64.71%) lower than traditional RANSAC, 0.017m (48.57%) lower than OOA-RANSAC, 0.011m (37.93%) lower than PSO-RANSAC, and 0.013m (41.94%) lower than SA-RANSAC. Combined with the dedicated geometric model of transmission line components, the fitting accuracy was significantly improved. The average fitting time was 0.87s, which is 1.21s (58.17%) lower than traditional RANSAC, 0.69s (44.23%) lower than OOA-RANSAC, 0.45s (34.09%) lower than PSO-RANSAC, and 0.58s (40.00%) lower than SA-RANSAC. The closed-loop feedback optimization reduced invalid iterations and greatly improved the fitting efficiency, fully meeting the fitting requirements of inspection scenarios. In summary, this technical solution first improves the core performance of traditional OOA through the OOA-GWO composite algorithm, and then integrates the improved OOA-GWO into the RANSAC algorithm to form the OOA-GWO-RANSAC algorithm. Combined with multimodal fusion technology, it achieves the core objectives of high accuracy, high efficiency, and strong robustness in point cloud fitting of transmission lines. The accuracy of interior point identification, fitting efficiency, and anti-interference ability all meet the engineering practical standards, and can directly support core applications such as line morphology reconstruction and defect detection in smart grid inspection.

[0180] This technical solution has strong versatility and flexibility. Its core logic is not limited to specific algorithms or fixed parameters and can be flexibly adjusted according to actual application scenarios, data characteristics and hardware conditions.

[0181] Regarding algorithm selection, the Octopus Optimization (OOA) and Grey Wolf Optimization (GWO) algorithms involved in the scheme can be replaced with other swarm intelligence optimization algorithms with better performance. The feature extraction methods used in the multimodal feature fusion process (including ORB feature extraction and semantic segmentation network) can also be replaced with more suitable technical solutions according to the data accuracy requirements. For the improvement logic of the RANSAC algorithm, the OOA-GWO composite optimization can also be replaced with other intelligent optimization strategies such as particle swarm optimization and simulated annealing, all of which can achieve optimization and upgrading of the sampling logic.

[0182] Regarding parameter settings, the key parameters (including population size) mentioned in the paper include population size, number of iterations, screening threshold, and feature weights. N Maximum number of iterations T max Fitness convergence threshold ε (etc.), is not a fixed value, and can be dynamically adjusted according to the actual point cloud data density, noise intensity, equipment computing power, etc., in order to balance processing efficiency and fitting accuracy.

[0183] It should be noted that the sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0184] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such process, method, product, or apparatus. The terms "first," "second," and "third," etc., are used to distinguish different objects, etc., and do not indicate a sequence, nor do they limit "first," "second," and "third" to different types.

[0185] In the description of the embodiments of this application, terms such as "exemplary," "for example," or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "exemplary," "for example," or "for instance" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary," "for example," or "for instance" is intended to present the relevant concepts in a concrete manner.

[0186] In the description of the embodiments of this application, unless otherwise stated, " / " means "or". For example, A / B can mean A or B. The "and / or" in the text is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist simultaneously, and B exists alone. In addition, in the description of the embodiments of this application, "multiple" means two or more.

[0187] In some processes described in the embodiments of this application, multiple operations or steps are included in a specific order. However, it should be understood that these operations or steps may not be executed in the order they appear in the embodiments of this application, or they may be executed in parallel. The sequence number of the operation is only used to distinguish different operations, and the sequence number itself does not represent any execution order. In addition, these processes may include more or fewer operations, and these operations or steps may be executed sequentially or in parallel, and these operations or steps may be combined.

[0188] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a terminal device to execute the methods described in the various embodiments of this application.

[0189] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for fitting point clouds of transmission lines, characterized in that, The point cloud fitting method for transmission lines includes: Geometric features corresponding to the point cloud data of the transmission line and semantic features corresponding to the image data of the transmission line are extracted. Feature fusion is performed on the geometric features and semantic features to obtain the multimodal fusion features corresponding to each point cloud. Based on the multimodal fusion features, a composite swarm intelligent optimization algorithm is used to mine high-confidence in-point regions in the point cloud of transmission lines, and to generate sampling weights corresponding to each high-confidence in-point region. Guided by the aforementioned sampling weights, weighted non-uniform sampling is performed on the sampling process of the random sampling consensus algorithm to obtain the target sampling point set; Based on the target sampling point set, the geometric model corresponding to the transmission line component is fitted to complete the point cloud fitting of the transmission line.

2. The point cloud fitting method for transmission lines as described in claim 1, characterized in that, The step of performing weighted non-uniform sampling on the sampling process of the random sampling consensus algorithm, guided by the sampling weights, to obtain the target sampling point set includes: The number of sampling times for each high-confidence in-point region is allocated according to its sampling weight. Within each high-confidence inner point region, sampling is performed within the region based on spatial grid division to obtain a preliminary set of sampling points; Semantic and geometric feature verifications are performed on the initial sample point set to eliminate potential outliers and obtain the final target sample point set.

3. The point cloud fitting method for transmission lines as described in claim 1, characterized in that, The step of fitting the geometric model corresponding to the transmission line component based on the target sampling point set to complete the transmission line point cloud fitting includes: Based on the multimodal features of the high-confidence in-point region, the corresponding transmission line component type is determined, and a geometric model that matches the component's structural characteristics is matched. Based on the initial parameters of the corresponding geometric model fitted to the target sampling point set, the spatial distance from each point in the point cloud to the initial geometric model is calculated. The dynamic inlier determination threshold is calculated based on the point cloud distance distribution characteristics. Inlier filtering is performed by combining semantic verification and geometric feature verification to obtain a pure inlier set. The parameters of the geometric model are optimized based on the pure interior point set. After model validation, the optimal geometric model and fitting results are output.

4. The point cloud fitting method for transmission lines as described in claim 3, characterized in that, Also includes: When the proportion of interior points and the fitting error in model validation do not meet the preset requirements, iterative optimization of geometric model parameters is initiated. If the requirements are still not met after iterative optimization, the sampling strategy is dynamically adjusted, the sampling point set is optimized, and the model fitting and parameter optimization are re-executed.

5. The point cloud fitting method for transmission lines as described in claim 1, characterized in that, Based on the multimodal fusion features, a composite swarm intelligence optimization algorithm is used to mine high-confidence in-point regions in the transmission line point cloud, and sampling weights are generated corresponding to each high-confidence in-point region, including: The population parameters and iteration parameters of the composite swarm intelligent optimization algorithm are initialized. The composite swarm intelligent optimization algorithm integrates the global exploration capability of the octopus optimization algorithm and the hierarchical guided convergence capability of the gray wolf optimization algorithm to complete the generation of the initial population and the group division of hunters and scouts. Construct a fitness function for point mining within the adaptive point cloud, calculate the fitness value of each initial individual in the population based on the multimodal fusion features, and determine the initial hierarchical guiding individuals; Entering the iterative process, for each individual hunter, the relative distance of the tentacles is calculated and compared with a preset proximity threshold. Based on the comparison results, the global exploration stage and the local utilization stage are dynamically switched, and the corresponding optimization algorithm is used to update the tentacle position, while simultaneously updating the hunter's head position. Perform global exploration updates for the scout group, and based on the fitness comparison results between the scouts and the corresponding reference hunters, complete the population conversion of scouts to hunters and the initialization of new hunter individuals; After each iteration, update the global hierarchical guiding individual and iteration parameters, and check the iteration termination condition. If the termination condition is not met, proceed to the next iteration; if the termination condition is met, stop the iteration. After the iteration terminates, a set of high-confidence in-point regions is obtained by filtering based on the fitness values ​​of individuals in the population. Based on the fitness values ​​corresponding to each high-confidence in-point region, the sampling weights corresponding to each region are calculated and generated.

6. The point cloud fitting method for transmission lines as described in claim 1, characterized in that, The fitness function for constructing the adaptive point cloud intra-point mining includes: We construct a single-objective fitness function that integrates spatial density, geometric consistency, and historical reward dimensions, set differentiated weighting coefficients for each dimension, and incorporate semantic mask hard constraint rules.

7. The point cloud fitting method for transmission lines as described in claim 1, characterized in that, The geometric features corresponding to the point cloud data of the transmission line and the semantic features corresponding to the image data of the transmission line are extracted. Feature fusion is performed on the geometric features and semantic features to obtain the multimodal fusion features corresponding to each point cloud, including: Geometric features are extracted point by point from the preprocessed point cloud data, and normalization is performed on the geometric features. Semantic segmentation is performed on the preprocessed image data to generate corresponding semantic masks, semantic features are extracted, and after normalization processing is performed on the semantic features, the semantic features are mapped to the corresponding point cloud based on the registration relationship. Simultaneously, visual features of the image data are extracted, and after dimensionality reduction and normalization processing of the visual features, they are mapped to the corresponding point cloud. Differential fusion weights are set according to the contribution of various features to the fitting. Weighted splicing and fusion are performed on the normalized geometric features, semantic features and visual features to generate multimodal fusion features corresponding to each point cloud.

8. The point cloud fitting method for transmission lines as described in claim 7, characterized in that, After mapping the semantic features to the corresponding point cloud, the process further includes: The neighborhood majority voting method is used to perform spatial filtering and smoothing on the semantic labels mapped to the point cloud, thereby correcting the semantic label noise caused by registration error.

9. The point cloud fitting method for transmission lines as described in claim 1, characterized in that, Before extracting the geometric features corresponding to the transmission line point cloud data and the semantic features corresponding to the transmission line image data, the method further includes: Acquire 3D point cloud data and 2D image data corresponding to power transmission line inspection, complete spatial registration and preprocessing of the two types of data, and obtain registered and aligned point cloud data and image data.

10. The point cloud fitting method for transmission lines as described in claim 1, characterized in that, After completing the initial fitting of the transmission line point cloud, the following steps are also included: The fitted set of interior points, the proportion of interior points, and the fitting error are fed back to the composite swarm intelligent optimization algorithm. Based on the feedback results, the high-confidence in-point regions are classified and labeled, and the sampling weights of each region are adjusted accordingly. Based on the adjusted sampling weights, a new round of sampling and fitting optimization is performed to complete the closed-loop feedback iteration.

Images