A power line laser radar point cloud extraction method based on hierarchical feature perception

Abstract

This invention proposes a method for extracting power line lidar point clouds based on hierarchical feature perception, belonging to the field of point cloud data processing. The method includes: acquiring raw 3D point cloud data of power lines from lidar, generating a digital elevation model, and constructing a 3D density field; normalizing the 3D point cloud density field, calculating dynamic elevation offset, and filtering based on a dynamic elevation threshold surface to obtain potential power line point clouds; acquiring local point sets within the neighborhood, calculating local linearity features, and thresholding the local point sets to obtain refined power line point clouds; performing density-based spatial clustering on the refined power line point clouds, and obtaining independent power line point clouds through graph segmentation and aggregation optimization; and iteratively optimizing the independent power line point clouds based on spatial continuity constraints to obtain power line point sets. This invention achieves high-precision and robust extraction of power line point clouds in complex scenarios.

Description

Technical Field

[0001] This invention relates to the field of point cloud data processing technology, and in particular to a method for extracting point clouds from power line lidar based on hierarchical feature perception. Background Technology

[0002] Safety inspection of power lines is a crucial aspect of power grid operation and maintenance. LiDAR technology can rapidly acquire high-precision three-dimensional spatial data, providing an effective means for digital modeling of power lines. Automatically and accurately extracting power line point clouds from complex LiDAR point clouds is fundamental for subsequent sag analysis, foreign object detection, and safety distance assessment.

[0003] Existing power line point cloud extraction techniques suffer from the following main problems: First, many methods rely on a globally fixed elevation threshold for initial screening. This method is poorly adaptable to complex terrain conditions, and in undulating mountainous areas, a fixed threshold can easily lead to the misclassification of ground points or the misfiltering of power line points. Second, the limited penetration of lidar in dense vegetation areas results in sparse power line point cloud sampling, and a fixed threshold makes it difficult to distinguish ground features at different elevation levels. Third, while machine learning or deep learning methods offer better extraction results, they require a large amount of labeled data for training, the model's generalization ability is affected by the training set, and computational resources are consumed significantly. Furthermore, traditional rule-based methods often use a single feature for discrimination, which is insufficient for distinguishing power lines with complex shapes or those that are close together.

[0004] Chinese invention patent CN112465765B discloses a method for constructing a road surface depth information model based on vehicle-mounted mobile laser point clouds. This method employs a single-lane overlapping segment division strategy to control the road surface model. First, the RANSAC algorithm is used to obtain a consistent estimation plane of a certain length of single-lane road surface point cloud, and the elevation of the original point cloud is normalized using this plane as a reference. Then, a combined filtering method of median filtering and Butterworth low-pass filtering is used to smooth local high-frequency noise and global vibration noise on the road surface. Finally, a quadratic surface model is constructed, and using the filtered and smoothed vehicle-mounted mobile laser road surface point cloud as observations, an iterative least squares fitting method that suppresses low-frequency features is used to obtain unbiased estimates of the quadratic surface parameters. However, its density field construction is only performed on the horizontal plane, which cannot reflect the differences in point cloud distribution at different heights in the vertical direction and makes it difficult to accurately capture the density changes at different heights caused by vegetation shading. Summary of the Invention

[0005] In view of this, the present invention proposes a method for extracting power line lidar point clouds based on hierarchical feature perception, which solves the problem that existing technologies cannot extract power line point clouds in complex terrain and vegetation interference situations. Through scene-adaptive initial screening and multi-dimensional feature fusion refinement, it achieves high-precision and high-robust extraction of power line point clouds in complex scenes.

[0006] The technical solution of this invention is implemented as follows: This invention provides a method for extracting point clouds from power line lidar based on hierarchical feature perception, comprising the following steps:

[0007] S1. Obtain the original lidar 3D point cloud data of the power line, separate the ground points from the original lidar 3D point cloud data to generate a digital elevation model, divide the original lidar 3D point cloud data into grid units on the horizontal plane, divide each grid unit into multiple elevation layers along the vertical direction based on the digital elevation model, and calculate the point cloud density of each grid unit in each elevation layer to construct a three-dimensional density field.

[0008] S2. Normalize the 3D point cloud density field to generate a normalized density field. Substitute the normalized density field into the dynamic offset function that is negatively correlated with the point cloud density to calculate the dynamic elevation offset of each grid cell at each elevation layer. Superimpose the digital elevation model and the dynamic elevation offset grid by grid and elevation layer by elevation and generate a dynamic elevation threshold surface through spatial smoothing. Filter the original point cloud according to the dynamic elevation threshold surface to obtain the potential power line point cloud.

[0009] S3. Construct a spatial spherical neighborhood for each point in the potential power line point cloud and obtain the local point set within the neighborhood. Perform principal component analysis on the local point set to calculate the local linearity feature and calculate the standard deviation of the reflection intensity of the local point set. Based on the local linearity feature and the standard deviation of the reflection intensity, perform threshold determination on the local point set to obtain the refined power line point cloud.

[0010] S4. Density-based spatial clustering is performed on the refined power line point cloud to obtain multiple trajectory segments. Geometric and morphological features of the trajectory segments are extracted, and similarity relationships are constructed based on the geometric and morphological features. Independent power line point clouds are obtained through graph segmentation and aggregation optimization.

[0011] S5. For independent power line point clouds, the main direction is calculated and the point cloud projection is sorted. Key control points are extracted, and a three-dimensional spline curve is constructed based on the key control points. The arc length change rate and curvature of the three-dimensional spline curve are calculated. Iterative optimization is performed according to spatial continuity constraints until the preset continuity constraint conditions are met, and the power line point set is obtained.

[0012] Based on the above technical solutions, preferably, step S2 specifically includes:

[0013] The density values ​​in the three-dimensional density field are normalized to obtain the normalized density field.

[0014] Set basic safety height and density sensitivity coefficient The dynamic elevation offset is calculated by substituting the normalized density field into the dynamic offset function, which is:

[0015] ;

[0016] in, This represents the dynamic elevation offset of a grid cell at a horizontal position (x, y) and a height h above the ground. This represents the x-coordinate of the point cloud on the horizontal plane. This represents the ordinate of the point cloud on the horizontal plane. Indicates the current point's altitude above the ground. Indicates the basic safety height. Represents the density sensitivity coefficient. The coupling weight coefficients represent the plane density. This represents the normalized density value of the grid cell at the horizontal position (x, y) and elevation layer h. The coupling weight coefficient represents the height above the ground. This represents the estimated maximum ground clearance in the scenario. Indicates a nonlinear exponent;

[0017] The ground elevation values ​​of the digital elevation model The initial dynamic elevation threshold matrix is ​​obtained by superimposing it grid by grid with the dynamic elevation offset;

[0018] Spatial smoothing is applied to the initial dynamic elevation threshold to obtain the dynamic elevation threshold surface;

[0019] Traverse each point P in the 3D point cloud data of the LiDAR, determine the grid cell to which point P belongs based on the horizontal coordinate of point P, and determine the absolute elevation of point P. The threshold of the corresponding grid in the dynamic elevation threshold surface of the grid cell to which point P belongs. Compare;

[0020] Extract all These points form a cluster of potential power line points.

[0021] Based on the above technical solutions, preferably, the step of spatially smoothing the initial dynamic elevation threshold to obtain the dynamic elevation threshold surface specifically includes:

[0022] The ground elevation values ​​of each grid cell in the digital elevation model. The dynamic elevation threshold is obtained by superimposing the dynamic elevation offset of the corresponding grid cell at the corresponding elevation layer on a grid-by-grid and elevation-by-elevation-layer basis.

[0023] The initial dynamic elevation threshold matrix is ​​convolved using a 3x3 three-dimensional Gaussian convolution kernel to obtain a spatially smoothed dynamic elevation threshold surface.

[0024] Based on the above technical solutions, preferably, step S3 specifically includes:

[0025] S31. Construct a spatial index structure for potential power line point clouds based on kd-tree, and determine the neighborhood search radius. ;

[0026] S32. For each point X in the potential power line point cloud, construct a system with radius X centered at point X. Given a spherical neighborhood, obtain all points within the neighborhood as a local point set. Calculate the local point set The three-dimensional centroid coordinates;

[0027] S33, local point set Principal component analysis is performed on any point Q in the matrix to obtain the eigenvalues ​​of the three principal components. Based on the eigenvalues ​​of the three principal components, the local linearity feature of point Q is calculated. ;

[0028] S34. Calculate local point sets Standard deviation of reflection intensity at all points ;

[0029] S35, Set linearity threshold and the standard deviation threshold of reflection intensity Threshold determination is performed on point Q:

[0030] when At that time, point Q is a candidate point for refining the electric power line;

[0031] S36. Repeat steps S33-S35 until a local point set is reached. All points are traversed, and all candidate points for refined power lines form a set of refined power line points.

[0032] Clustering based on connected component analysis is performed on the refined power line point set to obtain the refined power line point cloud.

[0033] Based on the above technical solutions, preferably, the geometric and morphological features include a dominant direction vector. Local elevation model and the centroid of the fragment space ,in,

[0034] The dominant direction vector By analyzing each trajectory segment The first principal component vector obtained by PCA calculation Characterizing trajectory segments The direction of extension;

[0035] The local elevation model It is a trajectory fragment Projecting points within the trajectory segment The result is obtained by quadratic polynomial fitting along the dominant direction;

[0036] Fragment Space Center of Mass It is a trajectory fragment The average three-dimensional coordinates of all points within the area.

[0037] Based on the above technical solutions, the preferred approach is to construct similarity relationships based on geometric and morphological features, specifically including:

[0038] Each trajectory segment Defined as a graph a vertex ;

[0039] For any two vertices and Calculate the comprehensive similarity weight of its edges. The formula for calculating the comprehensive similarity weight is as follows:

[0040] ;

[0041] ;

[0042] ;

[0043] ;

[0044] in, Indicates directional similarity, Indicates spatial connectivity, Indicating morphological similarity, Representing trajectory segments The dominant direction vector, Representing trajectory segments The dominant direction vector, Indicates the center of mass of two segments and Distance in the plane perpendicular to its average direction Represents the distance scale parameter. Representing trajectory segments The coefficients of the quadratic term in the quadratic polynomial fitting of the local elevation model. Representing trajectory segments The coefficients of the quadratic term in the quadratic polynomial fitting of the local elevation model. This represents the curvature tolerance parameter.

[0045] Based on the above technical solutions, preferably, the graph segmentation specifically includes:

[0046] The k-nearest neighbor strategy is used for each vertex. Only the edges between the k most similar vertices are retained to form a sparse similarity matrix. ;

[0047] Based on sparse similarity matrix Calculate the normalized Laplacian matrix and perform eigenvalue decomposition on it, selecting the smallest eigenvalue. The eigenvectors corresponding to each eigenvalue ;

[0048] Each vertex use The components corresponding to each feature vector are represented as points in the feature space. ;

[0049] The K-means clustering algorithm is used to cluster points in the feature space. Clustering into K clusters yields the primary trajectory clusters.

[0050] Based on the above technical solutions, preferably, the aggregation optimization specifically includes:

[0051] Define the global energy function The formula is:

[0052] ;

[0053] in, Represents the data fitting term. The weighting coefficients of the smoothing constraint term are represented. Represents the smoothing constraint term. This represents the weighting coefficients that incorporate prior knowledge. This indicates the incorporation of prior knowledge. Represents the data fitting term The power coefficient, The exponential growth factor representing the data fit term. Represents the smoothing constraint term The power coefficient, The exponential growth factor representing the smoothing constraint term. Represents physical constraint terms The power coefficient, The exponential growth factor representing the physical constraint term;

[0054] The aggregation problem of K clusters is formalized as a global energy minimization problem;

[0055] An optimization algorithm is used to search for an aggregation scheme that minimizes the global energy function. According to the aggregation scheme The primary trajectory clusters belonging to the same power line and their corresponding original LiDAR 3D point cloud data are merged to obtain the 3D point cloud sets of each independent power line. .

[0056] Based on the above technical solutions, preferably, step S5 specifically includes:

[0057] For each independent power line, the main direction of the overall point set in the three-dimensional point cloud is used to project all points onto an auxiliary spatial line parallel to the main direction and sort them along the projection points to obtain an ordered point sequence. ;

[0058] Using the Douglas-Puk algorithm on ordered point sequences Key points were extracted from the spline as control points. ;

[0059] Based on spline control points Constructing piecewise 3D Catmull-Rom spline curves ;

[0060] In three-dimensional Catmull-Rom spline curves Dense sampling is performed on the sampled area, and the rate of change of arc length between adjacent sampled segments is calculated. and the curvature of each sampling point ;

[0061] Set the threshold for the rate of change of arc length and curvature threshold Identify Arc length abrupt change region and curvature Far exceeding the average curvature Or, if there is a sudden change in curvature with a sharp peak, the corresponding control point or data point will be marked as an abnormal candidate point;

[0062] Spatial continuity constraints are applied to the candidate points of anomalies to obtain the final spline curve. High-density uniform sampling is performed on the final spline curve to generate a point set representing the three-dimensional centerline of the power line. The total length of the three-dimensional centerline of the power line, the elevation of each point, and the sag are calculated.

[0063] Furthermore, spatial continuity constraints are applied to the outlier candidate points, specifically including:

[0064] Step a: Remove the point that most severely violates the spatial continuity constraint from the abnormal candidate points;

[0065] Step b: Refit the Catmull-Rom spline using the remaining control points to obtain a new spline curve;

[0066] Step c: Calculate the rate of change of the arc length of the new spline curve. and the curvature of each sampling point ;

[0067] Step d, repeat step ac, until all and This yields the final spline curve.

[0068] The point cloud extraction method for electric power line lidar based on hierarchical feature perception of the present invention has the following advantages over the prior art:

[0069] (1) This invention generates an adaptively adjustable dynamic threshold surface by deeply fusing point cloud density field and digital elevation model. This enables the algorithm to automatically “raise the line of sight” in dense vegetation areas to strongly suppress tree canopy, and at the same time “lower the line of sight” in high-altitude sparse areas where power lines are located to accurately capture targets, thereby achieving high-precision and high-robust extraction of power line point clouds in complex scenarios.

[0070] (2) Based on the universal physical and geometric characteristics of power lines, this invention does not rely on data training for specific scenarios. Therefore, it has good generalization ability for power line extraction tasks of different voltage levels, different regions and different environments. Attached Figure Description

[0071] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0072] Figure 1 This is a flowchart of a power line lidar point cloud extraction method based on hierarchical feature perception according to the present invention.

[0073] Figure 2 This is a block diagram of a power line lidar point cloud extraction method based on hierarchical feature perception according to the present invention. Detailed Implementation

[0074] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0075] like Figure 1 and Figure 2 As shown, this invention provides a method for extracting point clouds from power line lidar based on hierarchical feature perception, comprising the following steps:

[0076] S1. Obtain the original lidar 3D point cloud data of the power line, separate the ground points from the original lidar 3D point cloud data to generate a digital elevation model, divide the original lidar 3D point cloud data into grid units on the horizontal plane, divide each grid unit into multiple elevation layers along the vertical direction based on the digital elevation model, and calculate the point cloud density of each grid unit in each elevation layer to construct a three-dimensional density field.

[0077] Specifically, step S1 includes:

[0078] The original lidar 3D point cloud data of the power line is acquired, and the ground points are separated from the original point cloud using the progressive triangulation encryption filtering algorithm. A continuous digital elevation model is then generated by interpolation.

[0079] The original LiDAR 3D point cloud data is divided into regular 2D grid cells on the horizontal plane (XY plane), and different elevation layers are divided at fixed intervals along the vertical direction based on the digital elevation model.

[0080] For each grid single cloud in each layer The point cloud density within each horizontal grid is statistically analyzed to construct a three-dimensional density field. ,in This refers to relative elevation (height above the ground).

[0081] The reflection intensity values ​​in the point cloud are normalized.

[0082] Understandably, digital elevation models accurately describe terrain undulations, that is, each coordinate... Corresponding elevation value You can get the grid. elevation value The density values ​​of all grids together constitute a point cloud density field covering the entire power corridor, which quantifies the sampling integrity of the lidar in different areas.

[0083] In one embodiment of the present invention, the size of the two-dimensional grid cell is 1m×1m.

[0084] In one embodiment of the present invention, the elevation layers are spaced 2 meters apart in the vertical direction.

[0085] S2. Normalize the 3D point cloud density field to generate a normalized density field. Substitute the normalized density field into the dynamic offset function that is negatively correlated with the point cloud density to calculate the dynamic elevation offset of each grid cell at each elevation layer. Superimpose the digital elevation model and the dynamic elevation offset grid by grid and elevation layer by elevation and generate a dynamic elevation threshold surface through spatial smoothing. Filter the original point cloud according to the dynamic elevation threshold surface to obtain the potential power line point cloud.

[0086] Specifically, step S2 includes:

[0087] The density values ​​in the three-dimensional density field are normalized to obtain the normalized density field.

[0088] Set basic safety height and density sensitivity coefficient The dynamic elevation offset is calculated by substituting the normalized density field into the dynamic offset function, which is:

[0089] ;

[0090] in, This represents the dynamic elevation offset of a grid cell at a horizontal position (x, y) and a height h above the ground. This represents the x-coordinate of the point cloud on the horizontal plane. This represents the ordinate of the point cloud on the horizontal plane. Indicates the current point's altitude above the ground. Indicates the basic safety height. Represents the density sensitivity coefficient. The coupling weight coefficients represent the plane density. This represents the normalized density value of the grid cell at the horizontal position (x, y) and elevation layer h. The coupling weight coefficient represents the height above the ground. This represents the estimated maximum ground clearance in the scenario. Indicates a nonlinear exponent;

[0091] The ground elevation values ​​of the digital elevation model The initial dynamic elevation threshold matrix is ​​obtained by superimposing it grid by grid with the dynamic elevation offset;

[0092] Spatial smoothing is applied to the initial dynamic elevation threshold to obtain the dynamic elevation threshold surface;

[0093] Traverse each point P in the 3D point cloud data of the LiDAR, determine the grid cell to which point P belongs based on the horizontal coordinate of point P, and determine the absolute elevation of point P. The threshold of the corresponding grid in the dynamic elevation threshold surface of the grid cell to which point P belongs. Compare;

[0094] Extract all These points form a cluster of potential power line points.

[0095] Understandable. , These are the coupling weight coefficients, and they satisfy... These determine the coupling weighting coefficients of the planar density. Coupling weight coefficient with ground clearance The relative importance of decision-making.

[0096] This invention effectively overcomes the interference of complex terrain and vegetation through density-aware dynamic elevation normalization filtering, achieving a qualitative leap in the initial screening process. It also integrates geometric and intensity features, significantly improving the ability to distinguish between electric lines and interference objects. Spline fitting under spatial continuity constraints further eliminates noise, achieving high-precision individualization.

[0097] In one embodiment of the present invention, ; It is a non-linear exponent, when When, function For accelerated growth, this means that the threshold growth slows down when density or height begins to increase; while when Very high or At very high levels, the threshold increases dramatically. This is particularly suitable for scenarios requiring strict protection of sparsely populated power lines at high altitudes. When, function It then transforms into a slow-growth type. Initial growth is rapid, but later growth is slow. This is suitable for scenarios that are sensitive to low-lying vegetation but do not require extreme differentiation from tall tree canopies.

[0098] In one embodiment of the present invention, .

[0099] Furthermore, the step of spatially smoothing the initial dynamic elevation threshold to obtain the dynamic elevation threshold surface specifically includes:

[0100] The ground elevation values ​​of each grid cell in the digital elevation model. The dynamic elevation threshold is obtained by superimposing the dynamic elevation offset of the corresponding grid cell at the corresponding elevation layer on a grid-by-grid and elevation-by-elevation-layer basis.

[0101] A 3x3 Gaussian convolution kernel is used to convolve the initial dynamic elevation threshold matrix to obtain a spatially smoothed dynamic elevation threshold surface. The calculation formula is as follows:

[0102] ;

[0103] in, This represents the initial dynamic elevation threshold of the i-th grid cell in the corresponding elevation layer. Represents the i-th grid cell. Represents the i-th grid cell The ground elevation value at the corresponding location, This represents the original point cloud density value of the i-th grid cell at the corresponding elevation layer. This represents the dynamic elevation offset of the i-th grid cell in the corresponding elevation layer.

[0104] In one embodiment of the present invention, the basic safety height It is 2.5 meters, and the density sensitivity coefficient is... It is 5 meters.

[0105] In one embodiment of the present invention, the density value in a three-dimensional density field By using linear or quantile normalization methods, a uniform mapping is performed to the interval [0, 1] to generate a normalized density field, where the mapping function is:

[0106] ;

[0107] in, This represents the original point cloud density value of the i-th grid cell at elevation layer h at location (x, y). This indicates the maximum grid density.

[0108] S3. Construct a spatial spherical neighborhood for each point in the potential power line point cloud and obtain the local point set within the neighborhood. Perform principal component analysis on the local point set to calculate the local linearity feature and calculate the standard deviation of the reflection intensity of the local point set. Based on the local linearity feature and the standard deviation of the reflection intensity, perform threshold determination on the local point set to obtain the refined power line point cloud.

[0109] Specifically, step S3 includes:

[0110] S31. Construct a spatial index structure for potential power line point clouds based on kd-tree, and determine the neighborhood search radius. ;

[0111] S32. For each point X in the potential power line point cloud, construct a system with radius X centered at point X. Given a spherical neighborhood, obtain all points within the neighborhood as a local point set. Calculate the local point set The three-dimensional centroid coordinates;

[0112] S33, local point set Principal component analysis is performed on any point Q in the matrix to obtain the eigenvalues ​​of the three principal components. Based on the eigenvalues ​​of the three principal components, the local linearity feature of point Q is calculated. ;

[0113] S34. Calculate local point sets Standard deviation of reflection intensity at all points ;

[0114] S35, Set linearity threshold and the standard deviation threshold of reflection intensity Threshold determination is performed on point Q:

[0115] when At that time, point Q is a candidate point for refining the electric power line;

[0116] S36. Repeat steps S33-S35 until a local point set is reached. All points are traversed, and all candidate points for refined power lines form a set of refined power line points.

[0117] Clustering based on connected component analysis is performed on the refined power line point set to obtain the refined power line point cloud.

[0118] Understandable, the standard deviation of the reflection intensity value The smaller the value, the more uniform the material in that local area, which conforms to the physical characteristics of electric field lines. Points on the same electric field line, due to their uniform material, exhibit small fluctuations in reflection intensity. The value is low; while the reflectance intensity of objects such as vegetation is unstable. High value.

[0119] In one embodiment of the present invention, the neighborhood radius is... It is 0.3 meters.

[0120] In one embodiment of the present invention, the eigenvalues ​​of the three principal components are respectively ,in Then the linearity feature Linearity characteristics The closer the value is to 1, the closer the local structure is to a straight line.

[0121] In one embodiment of the present invention, the linearity threshold The standard deviation threshold for reflectance is 0.95. It is 5% of the overall average strength.

[0122] S4. Density-based spatial clustering is performed on the refined power line point cloud to obtain multiple trajectory segments. Geometric and morphological features of the trajectory segments are extracted, and similarity relationships are constructed based on these features. Independent power line point clouds are obtained through graph segmentation and aggregation optimization. Among them, the geometric and morphological features include the dominant direction vector. Local elevation model and the centroid of the fragment space ,in,

[0123] The dominant direction vector By analyzing each trajectory segment The first principal component vector obtained by PCA calculation Characterizing trajectory segments The direction of extension;

[0124] The local elevation model It is a trajectory fragment Projecting points within the trajectory segment The result is obtained by performing a quadratic polynomial fitting along the dominant direction, where the fitting formula is:

[0125] ;

[0126] in, Represents a local elevation model. Representing trajectory segments The coefficients of the quadratic term in the quadratic polynomial fitting of the local elevation model reflect the magnitude of the sag curvature of the trajectory segment; a larger coefficient indicates a larger sag curvature. Representing trajectory segments The arc length parameter of the upper point is the cumulative distance along the dominant direction calculated after projecting the points within the segment onto the dominant direction. Representing trajectory segments The coefficients of the first term in the quadratic polynomial fitting of the local elevation model reflect the linear variation trend of elevation along the arc length direction. Representing trajectory segments The constant term in the quadratic polynomial fitting of the local elevation model represents the elevation reference value of the trajectory segment at the arc length starting point;

[0127] Fragment Space Center of Mass It is a trajectory fragment The average three-dimensional coordinates of all points within the area.

[0128] By segmenting a continuous point cloud of power lines into a series of localized clusters with consistent internal orientation, each cluster is called a "trajectory segment". This avoids the problem of long-distance power lines breaking due to sag or noise.

[0129] In one embodiment of the present invention, a density-based spatial clustering with relaxed parameters (such as DBSCAN, with a large eps (neighborhood radius)) is used.

[0130] Specifically, similarity relationships are constructed based on geometric and morphological features, including:

[0131] Each trajectory segment Defined as a graph a vertex Where V represents the set of vertices and E represents the set of edges;

[0132] For any two vertices and Calculate the comprehensive similarity weight of its edges. The formula for calculating the comprehensive similarity weight is as follows:

[0133] ;

[0134] ;

[0135] ;

[0136] ;

[0137] in, Indicates directional similarity, Indicates spatial connectivity, Indicating morphological similarity, Representing trajectory segments The dominant direction vector, obtained by performing principal component analysis on the trajectory segment, is a three-dimensional unit vector that indicates the direction of extension of the trajectory segment. Representing trajectory segments The dominant direction vector, also obtained through principal component analysis, is a trajectory segment. The extension direction vector, Indicates the center of mass of two segments and Distance in the plane perpendicular to its average direction Represents the distance scale parameter. Representing trajectory segments The coefficients of the quadratic term in the quadratic polynomial fitting of the local elevation model reflect the sag curvature characteristics of the trajectory segment. Representing trajectory segments The coefficients of the quadratic term in the quadratic polynomial fitting of the local elevation model reflect the sag curvature characteristics of the trajectory segment. This represents the curvature tolerance parameter.

[0138] Understandable, spatial connectivity Fragments that deviate too far from the penalty space, A larger value results in a smaller penalty, maintaining a higher weight even between distant segments. Curvature tolerance parameter. The larger the value, the smaller the penalty, allowing segments with large differences in sag shape to still have high connection weights.

[0139] In one embodiment of the present invention, the segment distance in the spatial connectivity calculation is 2 meters.

[0140] In one embodiment of the present invention, the curvature tolerance parameter The value is 0.0001, ensuring that only segments with similar drooping shapes have high connection weights.

[0141] In one embodiment of the present invention, the graph segmentation specifically includes:

[0142] The k-nearest neighbor strategy is used for each vertex. Only the edges between the k most similar vertices are retained to form a sparse similarity matrix. ,Right now ,in, Let represent the element in the i-th row and j-th column of the sparse similarity matrix A. Represents vertices and vertex The overall similarity weight between them;

[0143] Based on sparse similarity matrix Calculate the normalized Laplacian matrix and perform eigenvalue decomposition on it, selecting the smallest eigenvalue. The eigenvectors corresponding to each of the following eigenvalues ​​(usually starting from the second smallest, ignoring the first smallest eigenvalue of 0). The formula for calculating the normalized Laplace matrix is ​​as follows:

[0144] ;

[0145] in, Represents the normalized Laplace matrix. Represents the identity matrix. This represents a diagonal matrix, and the number of edges connected to that node is determined by... Summing each row yields the following: Represents a sparse similarity matrix;

[0146] Each vertex use The components corresponding to each feature vector are represented as points in the feature space. ,Right now , Represents an m-dimensional real vector space;

[0147] The K-means clustering algorithm is used to cluster points in the feature space. Clustering into K clusters yields the primary trajectory clusters.

[0148] In one embodiment of the present invention, k=5.

[0149] In one embodiment of the present invention, the aggregation optimization specifically includes:

[0150] Define the global energy function The formula is:

[0151] ;

[0152] in, Represents the data fitting term. The weighting coefficients of the smoothing constraint term are represented. Represents the smoothing constraint term. This represents the weighting coefficients that incorporate prior knowledge. This indicates the incorporation of prior knowledge. Represents the data fitting term The power coefficient, The exponential growth factor representing the data fit term. Represents the smoothing constraint term The power coefficient, The exponential growth factor representing the smoothing constraint term. Represents physical constraint terms The power coefficient, The exponential growth factor representing the physical constraint term;

[0153] The aggregation problem of K clusters is formalized as a global energy minimization problem;

[0154] An optimization algorithm is used to search for an aggregation scheme that minimizes the global energy function. According to the aggregation scheme The primary trajectory clusters belonging to the same power line and their corresponding original LiDAR 3D point cloud data are merged to obtain the 3D point cloud sets of each independent power line. .

[0155] Understandable, data fitting term Used to measure the fitting error between the final aggregated power line model (such as Catmull-Rom splines) and its corresponding point cloud under a certain aggregation scheme; smoothing constraint term. It penalizes cases where the curvature of the electric field lines changes drastically or their directions are inconsistent after aggregation, and encourages the generation of smooth curvature; it incorporates prior knowledge. This is to maintain a reasonable spacing between power line outgoing lines and avoid unnecessary merging. All values ​​are ≥1, primarily used to control the initial penalty growth rate of the error term; All values ​​are greater than 0, used to control the subsequent penalty acceleration rate of the error term.

[0156] This invention formalizes the aggregation process as a global energy minimization problem, defining an energy function that includes a data fitting term, measuring the fitting error between the final aggregated power line model and its corresponding point cloud under a given aggregation scheme. An optimization algorithm is used to search for feasible aggregation schemes, aiming to find the scheme that minimizes global energy. This process automatically determines the optimal power line aggregation result and the final number of points. Based on the optimal aggregation scheme obtained, all over-segmented clusters belonging to the same power line and their original point clouds are merged to obtain a set of three-dimensional point clouds for each independent power line.

[0157] In one embodiment of the present invention, the optimization algorithm may be one of a greedy algorithm, simulated annealing, or graph cut algorithm.

[0158] S5. For independent power line point clouds, the main direction is calculated and the point cloud projection is sorted. Key control points are extracted, and a three-dimensional spline curve is constructed based on the key control points. The arc length change rate and curvature of the three-dimensional spline curve are calculated. Iterative optimization is performed according to spatial continuity constraints until the preset continuity constraint conditions are met, and the power line point set is obtained.

[0159] This invention generates an adaptively adjustable dynamic threshold surface by deeply fusing point cloud density fields and digital elevation models. This surface enables the algorithm to automatically "raise the line of sight" in densely vegetated areas to strongly suppress tree canopies, while "lowering the line of sight" in sparsely vegetated areas containing power lines to accurately capture targets. This achieves high-precision and robust extraction of power line point clouds in complex scenarios. In complex scenarios, this invention avoids the common practice of using globally fixed elevation thresholds, representing a fundamental paradigm shift: from static, rigid "one-size-fits-all" filtering to dynamic, intelligent "context-aware" filtering. This solves the core contradiction of existing technologies, which struggle to adapt to terrain undulations and vegetation density variations, making it difficult to achieve both fidelity and noise reduction in mountainous and forested areas.

[0160] Specifically, step S5 includes:

[0161] For each independent power line, the main direction of the overall point set in the three-dimensional point cloud is used to project all points onto an auxiliary spatial line parallel to the main direction and sort them along the projection points to obtain an ordered point sequence. ;

[0162] Using the Douglas-Puk algorithm on ordered point sequences Key points were extracted from the spline as control points. ;

[0163] Based on spline control points Constructing piecewise 3D Catmull-Rom spline curves ;

[0164] In three-dimensional Catmull-Rom spline curves Dense sampling is performed on the sampled area, and the rate of change of arc length between adjacent sampled segments is calculated. and the curvature of each sampling point The calculation formula is:

[0165] ;

[0166] in, This represents the rate of change of arc length between adjacent sampling segments. This represents the arc length of the i-th sampling segment on the spline curve. This represents the arc length of the (i+1)th sampling segment on the spline curve;

[0167] Set the threshold for the rate of change of arc length and curvature threshold Identify Arc length abrupt change region and curvature Far exceeding the average curvature Or, if there is a sudden change in curvature with a sharp peak, the corresponding control point or data point will be marked as an abnormal candidate point;

[0168] Spatial continuity constraints are applied to the candidate points of anomalies to obtain the final spline curve. High-density uniform sampling is performed on the final spline curve to generate a point set representing the three-dimensional centerline of the power line. The total length of the three-dimensional centerline of the power line, the elevation of each point, and the sag are calculated.

[0169] Understandably, an ideally smooth curve, its It should be close to 1; curvature The curvature of the curve was measured; the curvature of the same electric field line should change gradually.

[0170] In one embodiment of the present invention, the arc length change rate threshold The curvature threshold is 1.5. It is 0.003.

[0171] Furthermore, spatial continuity constraints are applied to the anomalous candidate points, specifically including:

[0172] Step a: Remove the point that most severely violates the spatial continuity constraint from the abnormal candidate points;

[0173] Step b: Refit the Catmull-Rom spline using the remaining control points to obtain a new spline curve;

[0174] Step c: Calculate the rate of change of the arc length of the new spline curve. and the curvature of each sampling point ;

[0175] Step d, repeat step ac, until all and This yields the final spline curve.

[0176] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A hierarchical feature-aware based power line laser radar point cloud extraction method, characterized in that: The method comprises the following steps: S1, obtaining original laser radar three-dimensional point cloud data of a power line, separating ground points from the original laser radar three-dimensional point cloud data to generate a digital elevation model, dividing the original laser radar three-dimensional point cloud data into grid units in a horizontal plane, dividing each grid unit into multiple elevation layers along a vertical direction based on the digital elevation model, respectively counting point cloud densities of each grid unit in each elevation layer, and constructing a three-dimensional stereoscopic density field; S2, performing normalization processing on the three-dimensional point cloud density field to generate a normalized density field, substituting the normalized density field into a dynamic offset function negatively correlated with the point cloud density to calculate dynamic elevation offsets of each grid unit in each elevation layer, superimposing the digital elevation model and the dynamic elevation offsets grid by grid and elevation layer by elevation layer, and performing spatial smoothing processing to generate a dynamic elevation threshold surface, and filtering the original point cloud based on the dynamic elevation threshold surface to obtain potential power line point cloud; S3, constructing a spatial spherical neighborhood for each point in the potential power line point cloud and obtaining a local point set in the neighborhood, performing principal component analysis on the local point set to calculate a local linearity feature, calculating a reflection intensity standard deviation of the local point set, performing threshold judgment on the local point set based on the local linearity feature and the reflection intensity standard deviation, and obtaining refined power line point cloud; S4, performing density-based spatial clustering on the refined power line point cloud to obtain multiple trajectory segments, extracting geometric and morphological features of the trajectory segments, constructing a similarity relationship based on the geometric and morphological features, and obtaining independent power line point cloud through graph segmentation and aggregation optimization; S5, calculating a main direction of the independent power line point cloud, projecting and sorting the point cloud to extract key control points, constructing a three-dimensional spline curve based on the key control points, calculating an arc length variation rate and a curvature of the three-dimensional spline curve, and performing iterative optimization according to a spatial continuity constraint until a preset continuity constraint condition is met to obtain a power line point set.

2. The hierarchical feature-aware power line laser radar point cloud extraction method of claim 1, wherein: Step S2 specifically comprises: performing normalization processing on the density values in the three-dimensional stereoscopic density field to obtain a normalized density field; Setting base safety height and amplitude coefficients The normalized density field is substituted into a dynamic migration function to calculate a dynamic elevation migration amount, and the dynamic migration function is: ； wherein, represents a dynamic elevation offset of a grid cell at a horizontal position (x, y) with a height above ground h, represents a horizontal coordinate of a point cloud on a horizontal plane, represents a vertical coordinate of a point cloud on a horizontal plane, represents a height above ground of a current point, represents a base safety height, represents a density sensitivity coefficient, represents a coupling weight coefficient of a planar density, represents a normalized density value of a grid cell at a horizontal position (x, y) at an elevation layer h, represents a coupling weight coefficient of a height above ground, represents a maximum height above ground estimate in a scene, represents a non-linear exponent; ground elevation values of a digital elevation model are overlaid with dynamic elevation offsets on a grid-by-grid basis to obtain an initial dynamic elevation threshold matrix; performing spatial smoothing processing on the initial dynamic elevation threshold to obtain a dynamic elevation threshold surface; Traverse each point P in the laser radar three-dimensional point cloud data, determine the grid unit to which the point P belongs according to the horizontal coordinate of the point P, and determine the threshold value of the corresponding grid in the dynamic height threshold surface of the grid unit to which the point P belongs according to the absolute height of the point P Compare with the threshold value of the corresponding grid in the dynamic height threshold surface of the grid unit to which the point P belongs Perform extracting all points, forming a potential power line point cloud.

3. The hierarchical feature-aware power line laser radar point cloud extraction method of claim 2, wherein: the spatial smoothing processing on the initial dynamic elevation threshold to obtain the dynamic elevation threshold surface specifically comprises: The ground elevation value of each grid cell in the digital elevation model is obtained The dynamic elevation offset of the corresponding grid cell in the corresponding elevation layer is superimposed grid by grid and elevation layer by elevation layer to obtain an initial dynamic elevation threshold; performing convolution operation on the initial dynamic elevation threshold matrix using a 3x3 three-dimensional Gaussian convolution kernel to obtain the dynamic elevation threshold surface subjected to spatial smoothing processing.

4. The hierarchical feature-aware power line laser radar point cloud extraction method of claim 1, wherein: Step S3 specifically comprises: S31, constructing a spatial index structure of the potential power line point cloud based on the kd-tree, and determining a neighborhood search radius ; S32, for each point X in the potential power line point cloud, a spatial spherical neighborhood with a radius of is constructed around the point X, all points within the neighborhood are obtained as a local point set , and the three-dimensional centroid coordinates of the local point set are calculated; S33, performing principal component analysis on any point Q in the local point set , to obtain eigenvalues of three principal components, and calculating a local linearity feature of the point Q based on the eigenvalues of the three principal components ; S34, calculating the local point set standard deviation of the reflection intensity of all points in the middle ; S35, set linearity threshold and a reflection intensity standard deviation threshold threshold determination is performed on the point Q: When point Q is a refined power line candidate point; S36, repeat steps S33-S35 until all points in the local point set are traversed, and all refined power line candidates form a refined power line point set; performing clustering on the refined power line point set based on connected component analysis to obtain refined power line point cloud.

5. The hierarchical feature-aware power line laser radar point cloud extraction method of claim 4, wherein: The geometry and morphology features include dominant direction vectors , local elevation models , and segment spatial centroids wherein, the principal direction vector by performing PCA computation on each trajectory segment the first principal component vector , representing the extension direction of the trajectory segment ​ the local elevation model is obtained by projecting points within a trajectory segment to a second order polynomial fit in the dominant direction of the trajectory segment ​ fragment space centroid is the trajectory fragment average of the three-dimensional coordinates of all points within.

6. The hierarchical feature-aware power line laser radar point cloud extraction method of claim 5, wherein: constructing a similarity relationship based on the geometric and morphological features specifically comprises: Each trajectory segment Defined as a graph a vertex ; For any two vertices and , the integrated similarity weight of their edge is calculated , and the formula for calculating the integrated similarity weight is ； ； ； ； wherein, denotes a direction similarity, denotes a spatial connectivity, denotes a shape similarity, denotes a dominant direction vector of a trajectory segment , denotes a dominant direction vector of a trajectory segment , denotes a distance between two segment centroids and in a plane perpendicular to their average direction, denotes a distance scale parameter, denotes a quadratic term coefficient of a quadratic polynomial fit in a local elevation model of a trajectory segment , denotes a quadratic term coefficient of a quadratic polynomial fit in a local elevation model of a trajectory segment , denotes a curvature tolerance parameter.

7. The hierarchical feature-aware power line laser radar point cloud extraction method of claim 6, wherein: the graph segmentation specifically comprises: Adopt k-nearest neighbor strategy for each vertex Only keep edges between the k most similar vertices, form a sparse similarity matrix ; Based on sparse similarity matrix A normalized Laplacian matrix is calculated and eigen decomposition is performed on the normalized Laplacian matrix, and the eigenvector corresponding to the smallest eigenvalue is selected ​ Each vertex with The components corresponding to each eigenvector are represented as points in the eigenspace ; using a K-means clustering algorithm to cluster points in the feature space into K clusters, resulting in preliminary trajectory clusters.

8. The hierarchical feature-aware power line laser radar point cloud extraction method of claim 7, wherein: the aggregation optimization specifically comprises: Defining the global energy function is given by ； in, Represents the data fitting term. The weighting coefficients of the smoothing constraint term are represented. Represents the smoothing constraint term. This represents the weighting coefficients that incorporate prior knowledge. This indicates the incorporation of prior knowledge. Represents the data fitting term The power coefficient, The exponential growth factor representing the data fit term. Represents the smoothing constraint term The power coefficient, The exponential growth factor representing the smoothing constraint term. Represents physical constraint terms The power coefficient, The exponential growth factor representing the physical constraint term; formalizing the aggregation problem of the K clusters into a global energy minimization problem; An aggregation scheme that uses an optimization algorithm to search for a global energy function to achieve a minimum value , according to the aggregation scheme The primary track cluster belonging to the same power line and its corresponding raw lidar three-dimensional point cloud data are merged to obtain a three-dimensional point cloud set of each independent power line .

9. The hierarchical feature-aware power line laser radar point cloud extraction method of claim 8, wherein: Step S5 specifically comprises: A main direction of an overall point set is calculated for a three-dimensional point cloud of each independent power line, all points are projected onto an auxiliary space line parallel to the main direction, and an ordered point sequence is obtained by sorting the projected points ; Key points are extracted as spline control points in a sequence of ordered points using the Douglas-Peucker algorithm ;​ Spline control points Constructing a piecewise three-dimensional Catmull-Rom spline curve ; In three-dimensional Catmull-Rom spline curves are densely sampled and the rate of change of arc length between adjacent sample segments is calculated and the curvature of each sample point ; Setting an arc length change rate threshold and a curvature threshold , identifying an arc length mutation region and a curvature far exceeding the average curvature or a curvature mutation point with a sharp peak, marking the corresponding control point or data point as an abnormal candidate point; performing spatial continuity constraint on the abnormal candidate points to obtain a final spline curve, performing high-density uniform sampling on the final spline curve to generate a point set representing a three-dimensional center line of the power line, and calculating a total length, elevations and sag of each point of the three-dimensional center line of the power line.

10. The hierarchical feature-aware power line laser radar point cloud extraction method of claim 9, wherein: performing spatial continuity constraint on the abnormal candidate points specifically comprises: Step a, eliminate the most serious point which violates the space continuity constraint from the abnormal candidate points; Step b, use the remaining control points to perform Catmull-Rom spline fitting again to obtain a new spline curve; Step c, calculating the new rate of change of arc length of the spline curve and the curvature of each sample point ; Step d, repeat steps a-c until all and , resulting in the final spline curve.

Images