A power transmission channel hidden danger real-time three-dimensional reconstruction and ranging method, medium and system

Abstract

This invention provides a method, medium, and system for real-time 3D reconstruction and ranging of hidden dangers in power transmission channels. It belongs to the technical field of real-time 3D reconstruction and ranging of hidden dangers in power transmission channels. This invention continuously acquires point clouds of power transmission channels using mobile lidar and unifies their coordinates. Based on a risk fidelity factor-driven adaptive voxel fidelity algorithm, computing resources are concentrated on key areas such as conductors, towers, and insulators. Then, an artificial intelligence model performs streaming segmentation and geometric reliability re-scoring on the point cloud, establishing associations between adjacent window objects to suppress duplicate alarms. An envelope quantile robust limit distance algorithm is used to calculate the robust clearance between conductors and obstacles. Finally, a comprehensive risk score is output by combining measurement quality scores and safety margins, forming a closed-loop supplementary acquisition system. This invention solves the technical problem in power transmission channel point cloud processing where using uniform processing intensity for key target areas and background areas makes it difficult to balance the accuracy and real-time performance of reconstructing slender structures under limited computing power.

Description

Technical Field

[0001] This invention belongs to the field of real-time three-dimensional reconstruction and ranging technology of hidden dangers in power transmission channels. Specifically, it relates to a method, medium and system for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels. Background Technology

[0002] Safety inspection of power transmission channels is a crucial aspect of power system operation and maintenance. Mobile LiDAR, due to its efficient 3D perception capabilities, has been widely applied to point cloud acquisition and hazard detection for power transmission lines. Current technologies typically process large-scale point clouds using uniform voxel downsampling combined with deep learning segmentation networks. This involves setting a uniform voxel scale for the entire scene, then feeding the downsampled point cloud into a segmentation model to extract targets such as conductors, towers, and insulators, and calculating safe distances. This approach meets basic requirements in scenarios with uniform point cloud density and similar target scales. However, in power transmission channel scenarios, key targets such as conductors and insulator attachment points often exhibit elongated structures or strong spatial locality, coexisting with extensive vegetation and structural backgrounds. When using a uniform voxel scale, excessively large voxel settings result in significant loss of geometric details of key targets, while excessively small voxel settings lead to a surge in global computational load, making real-time processing difficult. Simultaneously, the segmentation network exhibits low confidence in identifying key targets in occluded or sparse areas, and uniform resource allocation methods cannot provide targeted enhancement for these areas. In other words, existing technologies have a technical problem: using the same processing intensity for key target areas and background areas in power transmission channel point cloud processing makes it difficult to balance the accuracy and real-time performance of slender structure reconstruction under limited computing power. Summary of the Invention

[0003] In view of this, the present invention provides a method, medium and system for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels, which can solve the technical problem in the prior art that the key target area and the background area are processed with the same intensity in the point cloud processing of power transmission channels, which makes it difficult to balance the reconstruction accuracy and real-time performance of slender structures under limited computing power.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides a method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels, comprising the following steps: Mobile lidar continuously collects point clouds of power transmission channels, combines pose data to achieve coordinate unification, maps the original point cloud to a unified line coordinate frame, and forms a continuous data stream. Divide the sliding window according to the route direction and set the overlapping area. Calculate the risk fidelity factor for the spatial voxels in the window. Based on the risk fidelity factor, call the risk-driven adaptive voxel fidelity algorithm to adaptively determine the voxel scale and the retention rate. The risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network is invoked to perform initial semantic judgment on the preprocessed point cloud, and to calculate the linearity, segment continuity and catenary residual of the conductor candidate object, and to calculate the hanging point proximity and string direction consistency of the insulator candidate object. Local bounce verification is performed on low confidence objects to form a geometric confidence re-scoring. Establish association costs for adjacent window objects, determine the object identity inheritance relationship based on the center position offset, the main direction angle and the degree of overlap of the spatial envelope, suppress repeated alarms, extract a stable skeleton for the tower object, and calculate the tilt angle, tilt orientation and shape and position difference index. A spatial curve representation is established for the guide object, a local spatial envelope is established for the obstacle object, the envelope quantile robust limit distance algorithm is called to calculate the minimum distance, vertical clearance, horizontal clearance and robust distance, and output the distance measurement stability label; The measurement quality score is calculated based on point cloud density, occlusion degree, cross-frame overlap, curve fitting residual and geometric consistency. The safety margin and comprehensive risk score are calculated by combining robust distance and safe distance threshold, and the risk level is divided. For low-confidence and high-risk results, suggestions for supplementary sampling areas and viewpoints are generated, and the backflow forms a closed loop.

[0005] The risk fidelity factor is specifically a value obtained by weighted summation of the proximity of voxels to candidate regions for conductors, candidate regions for tower frames, and candidate regions for insulator attachments, and the local classification uncertainty.

[0006] The voxel scale is controlled inversely by the risk fidelity factor. The higher the risk fidelity factor, the smaller the voxel scale. The voxel scale is limited between the minimum voxel scale threshold and the maximum voxel scale threshold.

[0007] The minimum voxel scale threshold ranges from 3 to 8. The maximum voxel scale threshold ranges from 50 to 120. .

[0008] The risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network receives the coordinates of continuous window point clouds, reflection intensity, timestamps, pose matrices, historical object identifiers, and initial risk fidelity factors at the input layer. Local geometric features are extracted by a sparse voxel encoder, and adjacent window object-level features are transmitted through a line direction temporal memory module. Decoding channels are allocated by a risk-gated decoder, and low-confidence objects are re-encoded through a low-confidence region back-loop branch. The output layer outputs point-level categories, object-level confidence, risk fidelity heatmaps, and reconstruction priorities.

[0009] The risk-gated decoder converts the proximity of voxels to conductor candidate areas, tower skeleton candidate areas, and insulator hanging point candidate areas into dynamic sampling weights, allocating more decoding channels to high-weight areas.

[0010] In the low-confidence region bounce branch, when the probability of the conductor or insulator category is lower than the first probability threshold but the linearity is higher than the linearity threshold, the corresponding region features are sent back to the small-scale neighborhood coding layer to re-extract the features and then input into the decoder again.

[0011] The geometric credibility re-scoring is calculated by weighting semantic initial judgment probability, linearity, cross-segment continuity, vertical skeleton consistency, string credibility, noise ratio and boundary confusion. Objects with a re-scoring below the first threshold are triggered for backtracking and verification, while objects with a re-scoring above the second threshold are directly confirmed.

[0012] The first threshold ranges from 0.35 to 0.50, and the second threshold ranges from 0.65 to 0.80.

[0013] The calculation of the associated cost comprehensively considers three indicators: center position offset, main direction angle and spatial envelope overlap. The weight coefficients of the three indicators are determined by establishing a labeled sample set of historical inspection data and conducting multiple sets of comparative experiments.

[0014] The envelope quantile robust limit distance algorithm establishes a ranging neighborhood near the initial nearest point, extracts multiple sets of candidate distances, and performs density filtering, outlier removal, and quantile statistics to obtain robust distance and nearest point stability.

[0015] The quantile value ranges from 85 to 95, and the neighborhood radius ranges from 0.3 to 0.8. .

[0016] The ranging stability marker is marked as low stability when the difference between the minimum distance and the robust distance is greater than the difference threshold or the neighborhood density is lower than the density threshold.

[0017] The training of the risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network involves first jointly pre-training the encoder and decoder, then sequentially training the line direction temporal memory module, and finally fine-tuning the low-confidence region back-jump branches using verification samples.

[0018] The memory allocation of the sparse voxel encoder is determined by the mean and variance of the risk fidelity factor distribution within the window, and the number of cyclic memory units in the line direction timing memory module is increased or decreased accordingly with the number of overlapping areas of the sliding window.

[0019] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for real-time three-dimensional reconstruction and ranging of potential hazards in power transmission channels.

[0020] A third aspect of the present invention provides a real-time three-dimensional reconstruction and ranging system for potential hazards in power transmission channels, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0021] This invention quantifies the proximity and local classification uncertainty of each voxel to candidate areas for conductors, tower frames, and insulator attachment points by constructing a risk fidelity factor. This drives adaptive voxel scale allocation and centralized scheduling of computing resources to key areas, enabling slender structures to achieve higher point cloud density and more refined reconstruction processing under limited computing power. Simultaneously, the risk fidelity-guided streaming point cloud segmentation-reconstruction collaborative network allocates more decoding channels to high-risk areas and performs local resampling encoding on low-confidence objects through a risk-gated decoder and a low-confidence region bounce branch, fundamentally compensating for the shortcomings of uniform resource allocation in handling occlusion and sparse areas. The envelope quantile robust limiting distance algorithm suppresses the interference of isolated branches and leaves on the net distance results through quantile statistics and outlier removal, ensuring that the distance measurement results reflect the true spatial relationship between the overall outline of the obstacle and the conductor. In summary, this invention solves the technical problem mentioned in the background art where using the same processing intensity for key target areas and background areas in transmission channel point cloud processing makes it difficult to balance the reconstruction accuracy and real-time performance of slender structures under limited computing power. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a schematic diagram of the risk fidelity factor distribution and voxel scale adaptation results in this embodiment.

[0024] Figure 3 This is a graph showing the net distance distribution between the conductor and the obstacle in this embodiment.

[0025] Figure 4 This is a distribution diagram of the comprehensive risk score and safety margin along the route in this embodiment. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0027] like Figure 1 The diagram shown is a flowchart of a method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels provided by the first aspect of the present invention. This method includes the following steps: S01. The mobile lidar continuously collects point clouds of the power transmission channel, combines them with pose data to achieve coordinate unification, and maps the original point cloud to a unified line coordinate frame to form a continuous data stream. S02. Divide the sliding window according to the route direction and set the overlapping area. Calculate the risk fidelity factor for the spatial voxels in the window. Based on the risk fidelity factor, call the risk-driven adaptive voxel fidelity algorithm to adaptively determine the voxel scale and the retention rate. S03. Call the risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network to perform semantic preliminary judgment on the preprocessed point cloud, calculate the linearity, segment continuity and catenary residual of the conductor candidate object, calculate the hanging point proximity and string direction consistency of the insulator candidate object, and perform local bounce verification on the low confidence object to form a geometric confidence re-scoring. S04. Establish association costs for adjacent window objects, determine the object identity inheritance relationship based on the center position offset, the main direction angle and the degree of overlap of the spatial envelope, suppress repeated alarms, extract the stable skeleton for the tower object, and calculate the tilt angle, tilt orientation and shape and position difference index. S05. Establish a spatial curve representation for the guide object, establish a local spatial envelope for the obstacle object, call the envelope quantile robust limit distance algorithm, calculate the minimum distance, vertical clearance, horizontal clearance and robust distance, and output the distance measurement stability flag. S06. Calculate the measurement quality score based on point cloud density, occlusion degree, cross-frame overlap, curve fitting residual, and geometric consistency. Combine robust distance and safe distance thresholds to calculate safety margin and comprehensive risk score, classify risk levels, and generate supplementary sampling areas and viewpoint suggestions for low-confidence, high-risk results. Then, feed back to S02 to S06 to form a closed loop.

[0028] The following is an explanation of the non-public concepts, model structure, algorithm principles, and parameter sources in the above steps.

[0029] The risk fidelity factor is a weighted value obtained by weighting the proximity of the grid to candidate regions for conductors, tower frames, and insulator attachment points with local classification uncertainty, denoted as . ,in The weighting coefficient is determined by conducting multiple experiments on historical inspection data and comparing the accuracy of hazard identification with the false negative rate. The voxel scale is inversely controlled by the risk fidelity factor, denoted as... and limited to and Between, the The value range is 3 to 8. The The value range is 50 to 120. The range of values ​​was determined by comparing the reconstruction accuracy and calculation time of line samples at different voltage levels.

[0030] Geometric credibility re-scoring refers to a weighted calculation of the initial semantic judgment result, combining linearity, cross-segment continuity, vertical skeleton consistency, serial credibility, noise ratio, and boundary confusion. It is denoted as [insert denoting value here]. The The weights are determined through iterative optimization using cross-validation on the labeled sample set. The first threshold... The value range is 0.35 to 0.50, the second threshold. The value range is 0.65 to 0.80, and the range is determined by the statistical distribution of false positive and false negative rates in multiple verification experiments.

[0031] Risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network is a deep learning-based model. Its structure is as follows: The input layer receives continuous window point cloud coordinates, reflection intensity, timestamps, pose matrices, historical object identifiers, and initial risk fidelity factors. These are then processed by a sparse voxel encoder through voxelization and sparse convolution to extract local geometric features of towers, conductors, vegetation, and structures. The encoded features are then fed into a line-direction temporal memory module. This module establishes recurrent memory units along the line direction, transferring and updating object-level features and identifiers between adjacent windows to ensure that the same object is present in adjacent windows. The features in the mouth are continuous; the encoded features simultaneously enter the risk-gated decoder, which is equipped with a risk-gated unit to convert the proximity of the raster to the candidate areas of conductors, tower skeletons and insulator hanging points into dynamic sampling weights, and allocate more decoding channels to high-weight areas; the network is equipped with a low-confidence region back-loop branch, when the probability of the conductor or insulator category is lower than the set value but the linearity is higher than the set value, the corresponding region features are sent back to the small-scale neighborhood coding layer to re-extract features and then input into the decoder again; the output layer outputs point-level categories, object-level confidence, risk-fidelity heatmap and reconstruction priority.

[0032] The specific steps for establishing the training dataset for the model include: collecting mobile lidar point cloud data of transmission channels with different terrains, vegetation densities, and voltage levels; manually labeling conductors, towers, insulators, vegetation, structures, ground, and foreign objects in the point cloud; separately labeling partially obscured, broken, and sparse areas as verification samples; dividing windows according to the line direction to generate training sample pairs; and attaching an initial value of the risk fidelity factor and an object identification label to each window.

[0033] The model training steps specifically include: first, jointly pre-training the encoder and decoder, using a weighted sum of point-level classification loss and object-level confidence regression loss as the total loss function; then, sequentially training the line direction temporal memory module, using adjacent window object identity consistency loss to constrain the memory unit parameters; finally, fine-tuning the low-confidence region bounce branches separately, using the difference in classification probability before and after the bounce calculated from the verification samples as a constraint term, iteratively adjusting the network parameters until the total loss converges.

[0034] Regarding resource allocation, the model establishes a functional relationship between the number of sparse voxel encoder layers and the distribution range of risk fidelity factors. When the proportion of high-risk fidelity factor voxels within the window is high, the memory allocation and CUDA stream allocation for convolutional layers in the region corresponding to the sparse voxel encoder are increased, while the memory allocation for convolutional layers in the background region is reduced, so that the total memory allocation is proportional to the distribution range within the window. The mean and variance of the distribution are jointly determined; a correspondence is established between the number of circular memory units in the line direction timing memory module and the number of overlapping areas in the sliding window. When the number of overlapping areas increases, the memory allocation of memory units is increased proportionally to ensure that cross-window feature transmission is not lost.

[0035] The principle of the risk-driven adaptive voxel-fidelity algorithm lies in transforming the relevance of hazard assessment into a basis for allocating processing intensity at the voxel level. This allows key areas such as the lowest point of the conductor, the tower frame, and the insulator attachment point to obtain smaller voxel scales and higher update frequencies under limited computing power, while background areas use larger voxels and lower update frequencies. Furthermore, object identification and local statistics are preserved in overlapping window areas to avoid redundant calculations in adjacent windows. The algorithm's technical effect is that by concentrating computing resources on areas relevant to hazard assessment, it improves the point cloud density and reconstruction accuracy of key targets while ensuring overall real-time processing, and reduces the loss of information about slender structures caused by uniform downsampling.

[0036] The principle of the envelope quantile robust limiting distance algorithm is that it does not directly use the absolute nearest point between the traverse point and the obstacle point as the ranging result. Instead, it first establishes the traverse space curve and the obstacle space envelope, builds a ranging neighborhood near the initial nearest point, extracts multiple candidate distances, and performs density filtering, outlier removal, and quantile statistics to obtain the robust distance and the stability of the nearest point. When the minimum distance differs significantly from the robust distance or the neighborhood density is insufficient, it is marked as low stability. The quantile value... The value range is 85 to 95, and the neighborhood radius is... The value range is 0.3 to 0.8. The range of values ​​was determined through multiple comparative experiments of distance measurement results and manual verification results under different vegetation density scenarios. The technical effect of this algorithm is that, by suppressing the interference of isolated branches and noise points on the nearest distance through quantile statistics and outlier removal, the output net distance result can reflect the true spatial relationship between the overall outline of the obstacle and the guide wire, reducing the probability of false high-risk alarms.

[0037] The weighting coefficients and thresholds involved in cross-frame association cost, shape and position difference index, measurement quality score, safety margin and comprehensive risk score are all determined by establishing a labeled sample set of historical inspection data, conducting multiple sets of comparative experiments and combining statistical analysis of false detection rate, false negative rate and review time, and periodically iterating and updating based on feedback data in actual operation.

[0038] The specific implementation of step S01 is as follows: the mobile lidar continuously scans along the power transmission channel, outputting three-dimensional point coordinates and reflection intensity information at high frequency, and simultaneously recording the pose matrix fused from inertial navigation and global positioning, including translation vectors and rotation matrices. The pose matrix is ​​used to transform each frame of the original point cloud from the sensor's local coordinate system to the global line coordinate system, ensuring that point clouds collected at different times and locations have a unified spatial reference and eliminating coordinate deviations introduced by the carrier's motion. The coordinate-unified point clouds are then stitched together in chronological order to form a continuous data stream extending along the line direction, providing a data foundation for subsequent window segmentation and sliding processing.

[0039] The specific implementation of step S02 is as follows: The continuous point cloud data stream is divided into sliding windows along the line direction. A certain proportion of overlap area is set between adjacent windows. The overlap area is used to retain object identification and local statistics, avoiding repeated complete calculations at the boundaries of adjacent windows and improving processing continuity. A risk fidelity factor is calculated for the spatial voxels within each window. ,in , , Representing voxels The degree of similarity between the candidate regions for conductors, tower frames, and insulator attachment points. Due to local classification uncertainty, to The weighting coefficients were determined through multiple experiments using historical inspection data. The risk-driven adaptive voxel fidelity algorithm is based on... Inverse control voxel scale , Higher voxel counts result in smaller voxel scales; the minimum voxel scale threshold is typically set between 3 and 8. The reference range for the maximum voxel scale threshold is 50–120. The voxel scale is limited between these two values, and the point preservation rate increases as the voxel scale decreases. The memory allocation and CUDA stream allocation of the sparse voxel encoder are based on the window size. The mean and variance of the distribution are dynamically adjusted, and high-risk fidelity factor areas receive more computing resources.

[0040] The specific implementation of step S03 is as follows: the preprocessed window point cloud is input into a risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network. The sparse voxel encoder of this network performs sparse convolution operations on the input voxels to extract local geometric features of towers, conductors, vegetation, and structures; the line direction temporal memory module establishes a circular memory unit along the line direction, transferring and updating object-level features and identity identifiers of adjacent windows to ensure feature continuity of the same object in cross-window processing; the risk-gated decoder converts the risk fidelity factor into dynamic sampling weights, allocating more decoding channels to high-risk areas; in low-confidence regions, when the class probability is lower than the first probability threshold and the linearity is higher than the linearity threshold, the back-jump branch sends the corresponding features back to the small-scale neighborhood encoding layer for re-extraction and decoding. After the network outputs point-level categories and object-level confidence, it calculates linearity, segment continuity, and catenary residuals for conductor candidate objects, and calculates hanging point proximity and string direction consistency for insulator candidate objects, according to the formula... The geometric confidence score is calculated, with a first threshold reference range of 0.35 to 0.50 and a second threshold reference range of 0.65 to 0.80. Objects below the first threshold trigger a local bounce-back review, while objects above the second threshold are directly confirmed.

[0041] The specific implementation of step S04 is as follows: For objects already identified in adjacent windows, an association cost matrix is ​​established based on three indicators: center position offset, main direction angle, and spatial envelope overlap. The Hungarian algorithm is then used for cross-window object matching to determine object identity inheritance relationships. Detection results for the same object in different windows are merged to suppress duplicate alarms. For tower objects, the tilt angle and tilt direction of the tower's main axis are calculated by extracting the point cloud's principal axis and the vertical skeleton direction. This is then compared with historical baseline skeletons to calculate a shape and position difference index, used to determine if the tower has potential tilting or deformation risks.

[0042] The specific implementation of step S05 involves fitting a catenary model to the point cloud of the guide object to establish a three-dimensional spatial curve representation; and establishing a local convex hull or minimum bounding volume as the spatial envelope for the point cloud of the obstacle object. The envelope quantile robust limiting distance algorithm uses the initial nearest point as the center, within a radius of 0.3 to 0.8. Multiple candidate point pairs are extracted from the neighborhood, and outlier removal based on local density is performed. Candidate distances that meet the density requirements are retained, and the 85th to 95th percentile values ​​of the remaining candidate distances are taken as robust distances. The minimum distance, vertical net distance, and horizontal net distance are output simultaneously. When the difference between the minimum distance and the robust distance exceeds the difference threshold, or when the effective point density in the neighborhood is lower than the density threshold, a low-stability ranging marker is output, indicating that the reliability of the result is low.

[0043] The specific implementation of step S06 involves comprehensively calculating a measurement quality score for each ranging object. The score inputs include five indicators: local point cloud density, target occlusion degree, cross-frame overlap, curve fitting residual, and geometric consistency. The weights of each indicator are determined through statistical analysis of historical data. The difference between the robust distance and the preset safe distance threshold is normalized to obtain a safety margin. This safety margin is then weighted with the measurement quality score to obtain a comprehensive risk score, which is used to classify risk levels. For ranging results with low confidence and high risk, the system automatically generates the coordinate range of the supplementary sampling area and a suggested inspection viewpoint based on the current point cloud density distribution and occlusion direction. This information is fed back to step S02 as an instruction to drive the next round of processing, forming an adaptive closed loop.

[0044] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described method for real-time three-dimensional reconstruction and ranging of potential hazards in power transmission channels.

[0045] A third aspect of the present invention provides a real-time three-dimensional reconstruction and ranging system for potential hazards in power transmission channels, comprising the aforementioned computer-readable storage medium. The system can be any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0046] Specifically, the principle of this invention is: The fundamental reason why this invention can solve the above-mentioned technical problems is that it transforms the correlation of potential risks into a basis for allocating processing intensity at the voxel level, thereby breaking the rigid constraint of uniform processing intensity on computing power allocation. Specifically, risk fidelity factor. This comprehensively characterizes the spatial proximity of each voxel to the candidate region of key targets and its classification uncertainty; the numerical value directly reflects the importance of that voxel in hazard assessment. Voxel Scale Depend on Inverse control automatically assigns smaller voxel scales to high-risk fidelity factor regions, resulting in higher point cloud density and more complete preservation of geometric details; while background regions use larger voxels to reduce redundant computation. This mechanism allows slender or local structures such as the lowest point of the conductor, tower frame, and insulator attachment points to be preserved with near-original accuracy, without the loss of critical geometric information due to global downsampling, all while maintaining overall computational resources.

[0047] At the segmentation and reconstruction level, the risk-gated decoder of the risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network transforms the risk fidelity factor into dynamic sampling weights, enabling high-risk areas to obtain more channel resources during the decoding stage. This fundamentally solves the problem of insufficient identification of small targets due to uniform decoding resource allocation. The low-confidence region bounce-back branch, when detecting a candidate object for a conductor or insulator with a classification probability below the threshold but a linearity above the threshold, sends the corresponding region features back to the small-scale neighborhood encoding layer for re-extraction and re-enters the decoder. This allows for targeted feature enhancement of targets in occluded or sparse areas, rather than simply discarding or maintaining low-confidence outputs. The line direction temporal memory module ensures continuous feature transmission of the same object between adjacent windows, avoiding target fragmentation and duplicate alarms caused by window segmentation.

[0048] At the ranging level, the envelope quantile robust limit distance algorithm does not use the absolute nearest point as the ranging basis. Instead, it extracts multiple candidate distances within the neighborhood of the initial nearest point, and outputs a robust distance after density filtering, outlier removal, and quantile statistics. The principle of quantile statistics is that isolated branch points or noise points usually constitute the minimum outliers in the distance distribution. Their density in the neighborhood is much lower than that of the main obstacle points. By removing low-density outliers and taking high quantiles, the obtained distance can truly reflect the spatial relationship between the overall outline of the obstacle and the guide wire, thereby suppressing false high-risk alarms caused by noise points. The above three-level mechanisms work together to ensure both the reconstruction accuracy of key areas and the real-time performance of the overall processing under limited computing power, logically forming a complete closed loop from resource allocation to feature extraction to ranging output.

[0049] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0050] The specific implementation of step S01 is as follows: the mobile lidar continuously acquires three-dimensional point clouds of the power transmission channel during its movement, with each frame of the point cloud carrying a timestamp, reflection intensity, and local sensor coordinates. The synchronous inertial navigation and global positioning system outputs a pose matrix, transforming the point cloud in the sensor coordinate system to a unified line coordinate frame, forming a continuous data stream, which is then used as the basis for subsequent windowing processing.

[0051] The specific implementation of step S02 is to divide the sliding window along the line direction with a fixed step size and set an overlapping area, and to process each spatial voxel within the window. The risk fidelity factor is calculated using the following formula: ; In the formula, voxels The risk fidelity factor, with a dimension of 1. voxels The normalized approximation to the candidate region of the conductor, with a dimension of 1. voxels The normalized approximation of the candidate regions for the tower frame, with dimensions of 1. voxels The normalized approximation to the candidate region of the insulator hanging point, with a dimension of 1. voxels The local classification uncertainty has a dimension of 1. , , , These are weighting coefficients, all greater than 0 and satisfying... After conducting multiple comparative experiments on historical inspection data and comparing the accuracy and false negative rates of hazard identification, it was determined that all terms are dimensionless normalized quantities, thus ensuring dimensional consistency on both sides of the equation. Based on the risk fidelity factor controlling the voxel scale inversely, the formula is expressed as follows: ; In the formula, voxels The voxel scale, in units of , This is the lower limit of the voxel scale, in units of The value ranges from 3 to 8. , This represents the upper limit of the voxel scale, in units of The value range is 50 to 120. , It is a dimensionless normalized quantity. The unit is The units on both sides of the equation are The value range was determined through experiments comparing the reconstruction accuracy and computation time of line samples at different voltage levels. Limiting The risk-driven adaptive voxel-fidelity algorithm allocates smaller voxels and higher update frequencies to key regions based on the aforementioned voxel scale, while using larger voxels and lower update frequencies for background regions. It also preserves object identifiers and local statistics in overlapping window areas, avoiding redundant calculations in adjacent windows. This improves the density and reconstruction accuracy of key target point clouds while ensuring real-time performance.

[0052] The specific implementation of step S03 is as follows: A risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network is invoked to perform initial semantic judgment on the preprocessed point cloud. The input layer of this network receives the coordinates of continuous window point clouds, reflection intensity, timestamps, pose matrices, historical object identifiers, and initial risk fidelity factors. These are then processed by a sparse voxel encoder through voxelization and sparse convolution to extract local geometric features. The encoded features enter the line-direction temporal memory module, establishing a circular memory unit along the line direction to transfer and update the object-level features and identifiers of adjacent windows, ensuring feature continuity for the same object across windows. Simultaneously, the encoded features enter the risk-gated decoder... , , This is converted into dynamic sampling weights, allocating more decoding channels to high-weight regions. The network sets up low-confidence region bounce-back branches when the probability of the conductor or insulator category falls below a first threshold. When the linearity exceeds the set value, the corresponding region features are sent back to the small-scale neighborhood coding layer for re-extraction and then input into the decoder again. The value range is 0.35 to 0.50. The output layer outputs point-level categories, object-level credibility, risk fidelity heatmaps, and reconstruction priorities. During the training phase, the total loss function is composed of a weighted sum of point-level classification loss and object-level credibility regression loss. Subsequently, a constraint of adjacent window object identity consistency loss is applied to the temporal memory module. Finally, the back-jump branch is fine-tuned using the difference in classification probabilities before and after the back-jump as a constraint term. In terms of resource allocation, the sparse voxel encoder increases the convolutional layer memory and computation flow allocation for windows with a high proportion of high-risk fidelity factor voxels. The total memory is determined by the window size. The mean and variance of the distribution are used to determine the number of circular memory units; the number of overlapping regions in the sliding window corresponds to the number of circular memory units, and the memory allocation increases proportionally as the number of overlapping regions increases. After the initial semantic judgment is completed, the linearity of the candidate wires is calculated. Segment continuity Compared with the catenary residual, the proximity of the hanging point and the consistency of the string direction are calculated for candidate insulators. For objects with low confidence, a local bounce-back review is performed to generate a geometric confidence rescore, expressed by the following formula: ; In the formula, For object The geometric credibility re-rating is dimensionless (1). For object In observation Under what conditions does it belong to a category? The posterior probability of , with a dimension of 1. For object The observed feature vector, For object semantic category labels, For object The linearity, with dimensions of 1. For object The segmental continuity, dimensionless, is 1. For object The vertical skeletal structure is consistent, with a dimension of 1. For object The serial reliability is dimensionless, with a value of 1. For object The noise ratio, dimensionless (1). For object The boundary confusion, with dimensions of 1. to The weight coefficients are dimensionless (1), and all terms are dimensionless normalized quantities. Therefore, both sides of the equation have the same dimensions. The weight coefficients are determined through cross-validation and iterative optimization on the labeled sample set. When a bounce-back verification is triggered, It is confirmed as a highly trusted object at that time. The second threshold is set at 0.65 to 0.80, and the range is determined by the statistical distribution of the false positive rate and the false negative rate in multiple verification experiments.

[0053] The specific implementation of step S04 involves establishing association costs for adjacent window objects, determining the object identity inheritance relationship based on the center position offset, the angle of the main direction, and the degree of overlap with the spatial envelope, and suppressing duplicate alarms. A stable skeleton is extracted from the tower object, and the tilt angle, tilt azimuth, and shape-position difference index are calculated. The weight coefficients and thresholds are determined by establishing a labeled sample set based on historical inspection data, using multiple sets of comparative experiments combined with statistical analysis of false detection rate, false negative rate, and review time, and are periodically updated based on feedback data during actual operation.

[0054] The specific implementation of step S05 involves establishing a spatial curve representation for the traverse object, establishing a local spatial envelope for the obstacle object, and calling the envelope quantile robust limit distance algorithm to calculate the minimum distance, vertical clearance, horizontal clearance, and robust distance. This algorithm does not directly use the absolute nearest point between the traverse point and the obstacle point as the ranging result. Instead, it first establishes the traverse spatial curve and the obstacle spatial envelope, and then establishes a neighborhood with a radius of [missing information] near the initial nearest point. The ranging neighborhood is used to extract multiple candidate distances and perform density filtering and outlier removal. The remaining candidate distance set is then selected. Percentiles, as a robust distance, are expressed by the following formula: ; In the formula, For robust distances, the unit is 1. , This is the set of candidate distances after density filtering and outlier removal, in units of... , Indicates taking the first value from the set. Percentiles are a sorting statistic; the output has the same units as the input set. This is a quantile parameter, with a dimension of 1, and a value range of 85 to 95. The distance to the neighborhood is the radius, in units of The value ranges from 0.3 to 0.8. The value range was determined through multiple comparative experiments on distance measurement results and manual verification results under different vegetation density scenarios. When the difference between the minimum distance and the robust distance is large or the neighborhood density is insufficient, a low stability marker is output to suppress the interference of isolated branch and leaf points and noise points on the net distance result.

[0055] The specific implementation of step S06 is to calculate the measurement quality score based on point cloud density, occlusion degree, cross-frame overlap, curve fitting residual, and geometric consistency, and combine it with robust distance. The safety margin and comprehensive risk score are calculated with the safety distance threshold, and the risk level is divided. For low-confidence and high-risk results, suggestions for supplementary sampling areas and perspectives are generated and returned to steps S02 to S06 to form a closed loop. Each weight coefficient and threshold is determined by multiple sets of comparative experiments on historical inspection and annotation sample sets and combined with statistical analysis. In actual operation, it is periodically updated based on feedback data.

[0056] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2: This embodiment takes the inspection of potential hazards on an overhead transmission line as the application background. The transmission channel extends along the mountainous and hilly terrain, with dense vegetation on both sides of the line. There are potential hazards of excessively tall trees in some sections. At the same time, there is a tower with a slight tilt record in the historical records. It is necessary to perform real-time three-dimensional reconstruction and distance measurement of the net distance between the conductor and the vegetation and the tilt status of the tower.

[0057] The inspection vehicle, equipped with a mobile lidar, travels at a constant speed along the route. The lidar collects approximately 2 million points per second and simultaneously outputs a pose matrix fused from inertial navigation and global positioning. In step S01, the system transforms each frame of point cloud data to a unified route coordinate system using the pose matrix, and then stitches them together to form a continuous point cloud data stream along the route. The original point cloud size in a single window is approximately 1.5 million points.

[0058] The system presses 40 The line direction length is divided into sliding windows, and the overlap length of adjacent windows is set to 8. In step S02, the risk fidelity factor is calculated for each voxel within the window, and the voxels near the candidate regions for conductors and tower frames are analyzed. The value is relatively high, and the background vegetation area is relatively high. The value is low. Based on... Distribution, the voxel scale near the conductor adaptively shrinks to approximately 5 The background vegetation area is expanded to a voxel scale of approximately 80 The fidelity retention rate is significantly higher in the conductor region than in the background region. Within this window, the proportion of high-risk fidelity factor voxels is approximately 18%. Memory allocation is increased in the sparse voxel encoder region, while memory allocation is reduced in the background region. The total memory allocation is adjusted based on the window size. The mean and variance of the distribution are jointly determined. After the risk-driven adaptive voxel fidelity algorithm is completed, the point density of key regions within the window is significantly improved compared with the uniform downsampling scheme, and the geometric integrity of slender structures is maintained.

[0059] In step S03, the preprocessed point cloud is input into a risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network. A sparse voxel encoder extracts local geometric features of various targets. The line direction temporal memory module inputs the conductor and tower object features and identifiers from the previous window into the current window, ensuring the continuity of cross-window object features. The risk-gated decoder allocates more decoding channels to the conductor and insulator hanging point regions, enhancing the segmentation accuracy of these regions. For several points at the top of a tree on the left side of the line, the initial semantic probability of the conductor candidate object is 0.42, lower than the first threshold of 0.45, but the linearity calculation result is 0.81, higher than the linearity threshold, triggering a low-confidence region jumpback branch. The corresponding features are sent back to the small-scale neighborhood encoding layer for re-extraction and decoding. After verification, the class probability increases to 0.71, and the final geometric confidence re-score is 0.73, higher than the second threshold of 0.70, confirming it as a conductor object. For the insulator candidate object, the hanging point proximity and string direction consistency calculation results are normal, and the geometric confidence re-score is 0.76, directly confirming it. like Figure 2 As shown, this figure illustrates the distribution of risk fidelity factors and the adaptive results of voxel scale in this embodiment. The figure shows the distribution pattern of voxel scale in each region along the line direction, with the voxel scale near the conductor and tower being significantly smaller than that in the background region.

[0060] In step S04, the system establishes association costs for the identified conductors, towers, and insulators in adjacent windows, and uses the Hungarian algorithm to complete cross-window matching, suppressing duplicate alarms caused by window segmentation. A stable frame is extracted for the target tower, with a calculated tilt angle of 1.8°, a tilt azimuth of east-northeast, and a shape-position difference index of 0.12. Compared with the historical baseline frame, the tilt condition is determined not to have worsened further, and the risk level remains at the concern level. The cross-window matching results for each object within this window are shown in Table 1.

[0061] Table 1. Results of Cross-Window Object Matching and Identity Inheritance

[0062] In step S05, a catenary model is used to fit the spatial curve of the point cloud of the conductor object, and a local convex hull spatial envelope is established for the trees on the left side of the line. The envelope quantile robust limiting distance algorithm establishes a neighborhood with a radius of 0.5 near the initial nearest point. The ranging neighborhood was calculated, and a total of 184 candidate point distance pairs were extracted. After performing local density outlier removal, 162 pairs were retained. The 90th quantile value was taken as the robust distance, resulting in a value of 3.21. The minimum distance is 2.87. The vertical clearance is 2.93. The horizontal clearance is 3.15. The difference between the minimum distance and the robust distance is 0.34. Below the difference threshold of 0.5 The neighborhood density meets the requirements, and the ranging stability is marked as normal. For example... Figure 3 As shown, this is the net distance distribution curve between the conductor and the obstacle in this embodiment. The figure shows the variation trend of robust distance and minimum distance along the line direction, as well as the distribution of distance measurement stability markers.

[0063] In step S06, a measurement quality score is calculated for the above-mentioned ranging object, and the input indicators are shown in Table 2.

[0064] Table 2 Input Indicators for Measurement Quality Scoring

[0065] Stable distance 3.21 With a safe distance threshold of 4.00 The normalized safety margin of the difference is 0.79. The weighted comprehensive risk score is 0.67, classifying it as a general risk level. The system automatically generates a suggestion for re-scanning the area, recommending an additional scan from the top-down view on the right side of the route. The re-scanning area is then returned to step S02, driving the next round of adaptive processing. Figure 4As shown, this figure illustrates the distribution of the comprehensive risk score and safety margin along the route in this embodiment. The risk level division boundary and the trigger position for supplementary sampling suggestions are marked in the figure.

[0066] Compared to traditional uniform downsampling with fixed threshold ranging, the technological advancements of this invention are reflected in the following aspects: Traditional uniform downsampling methods apply the same voxel scale to the global point cloud, resulting in the geometric details of slender wires and insulator attachment points being diluted to the same extent as background vegetation, and the point density of key targets cannot be increased as needed. This invention, through a risk fidelity factor-driven adaptive voxel fidelity algorithm, automatically shrinks the voxel scale in key areas, ensuring the integrity of the geometric information of slender structures from a fundamental mechanism. Traditional segmentation networks with fixed resource allocation have low confidence in target recognition in occluded and sparse regions, and cannot automatically resample and encode low-confidence regions. The risk-gated decoder and low-confidence region bounce-back branch of this invention achieve targeted enhancement processing for high-risk regions and low-confidence objects at the network structure level, thus ensuring the accuracy of recognition in occluded scenarios from a theoretical perspective. Traditional nearest-point ranging methods are easily affected by isolated branches and noise points, and the output net distance results cannot distinguish between the real obstacle contour and occasional noise points. The envelope quantile robust limit distance algorithm of this invention, through quantile statistics and density filtering, enables the output net distance to truly reflect the spatial relationship between the overall contour of the obstacle and the guide wire, thereby reducing the probability of false high-risk alarms from a statistical perspective.

[0067] It should be noted that the variables involved in this invention are explained in detail in Table 3.

[0068] Table 3. Variable Explanation Table

[0069] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for real-time three-dimensional reconstruction and ranging of potential hazards in power transmission channels, characterized in that, Includes the following steps: Mobile lidar continuously collects point clouds of power transmission channels, combines pose data to achieve coordinate unification, maps the original point cloud to a unified line coordinate frame, and forms a continuous data stream. Divide the sliding window according to the route direction and set the overlapping area. Calculate the risk fidelity factor for the spatial voxels in the window. Based on the risk fidelity factor, call the risk-driven adaptive voxel fidelity algorithm to adaptively determine the voxel scale and the retention rate. The risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network is invoked to perform initial semantic judgment on the preprocessed point cloud, and to calculate the linearity, segment continuity and catenary residual of the conductor candidate object, and to calculate the hanging point proximity and string direction consistency of the insulator candidate object. Local bounce verification is performed on low confidence objects to form a geometric confidence re-scoring. Establish association costs for adjacent window objects, determine the object identity inheritance relationship based on the center position offset, the main direction angle and the degree of overlap of the spatial envelope, suppress repeated alarms, extract a stable skeleton for the tower object, and calculate the tilt angle, tilt orientation and shape and position difference index. A spatial curve representation is established for the guide object, a local spatial envelope is established for the obstacle object, the envelope quantile robust limit distance algorithm is called to calculate the minimum distance, vertical clearance, horizontal clearance and robust distance, and output the distance measurement stability label.

2. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels according to claim 1, characterized in that, The risk fidelity factor is specifically a value obtained by weighted summation of the proximity of voxels to candidate regions of conductors, candidate regions of tower frames, and candidate regions of insulator attachment points, and local classification uncertainty.

3. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels according to claim 2, characterized in that, The voxel scale is inversely controlled by the risk fidelity factor. The higher the risk fidelity factor, the smaller the voxel scale. The voxel scale is limited between the minimum voxel scale threshold and the maximum voxel scale threshold.

4. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels according to claim 3, characterized in that, Also includes: The measurement quality score is calculated based on point cloud density, occlusion degree, cross-frame overlap, curve fitting residual and geometric consistency. The safety margin and comprehensive risk score are calculated by combining robust distance and safe distance threshold, and the risk level is divided. For low-confidence and high-risk results, suggestions for supplementary sampling areas and viewpoints are generated, and the backflow forms a closed loop.

5. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels according to claim 4, characterized in that, The risk-fidelity-guided streaming point cloud segmentation-reconstruction collaborative network receives the coordinates of continuous window point clouds, reflection intensity, timestamps, pose matrices, historical object identifiers, and initial risk fidelity factors at the input layer. Local geometric features are extracted by a sparse voxel encoder, and adjacent window object-level features are transmitted through a line direction temporal memory module. Decoding channels are allocated by a risk-gated decoder, and low-confidence objects are re-encoded through a low-confidence region back-loop branch. The output layer outputs point-level categories, object-level confidence, risk fidelity heatmaps, and reconstruction priorities.

6. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels according to claim 5, characterized in that, The risk-gated decoder converts the proximity of voxels to conductor candidate areas, tower skeleton candidate areas, and insulator hanging point candidate areas into dynamic sampling weights, allocating more decoding channels to high-weight areas.

7. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels according to claim 6, characterized in that, When the probability of a conductor or insulator category is lower than the first probability threshold but the linearity is higher than the linearity threshold, the low-confidence region bounce branch sends the corresponding region features back to the small-scale neighborhood coding layer to re-extract the features and then inputs them back into the decoder.

8. The method for real-time three-dimensional reconstruction and ranging of hidden dangers in power transmission channels according to claim 7, characterized in that, The geometric credibility re-scoring is calculated by weighting semantic initial judgment probability, linearity, cross-segment continuity, vertical skeleton consistency, serial credibility, noise ratio and boundary confusion. Objects with a re-scoring below the first threshold are triggered for backtracking review, while objects with a re-scoring above the second threshold are directly confirmed.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform a real-time three-dimensional reconstruction and ranging method for potential hazards in power transmission channels as described in any one of claims 1-8.

10. A real-time three-dimensional reconstruction and ranging system for potential hazards in power transmission channels, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images