A catenary target recognition algorithm based on laser radar

Abstract

The application discloses a catenary target recognition algorithm based on a laser radar and belongs to the technical field of laser radars. The catenary target recognition algorithm comprises the following steps: decoupling the adaptive voxel filtering of a vehicle speed and radar parameters to downsample original point clouds; using a vehicle body vibration motion state solution to intercept core point clouds of an envelope boundary; distinguishing catenary wires and load-bearing cables, constructing a trajectory optimization model, and independently adjusting a soft threshold value; in a catenary search area, using density inverse variable adaptive radius hierarchical clustering, and checking through spatial manifold eigenvalue algebra; using spatial compactness matrix eigenvalue ratio and projection information entropy to determine a positioner rigid body; based on a lightweight representation state vector, performing multi-label scene classification and dynamically correcting a global hard boundary; and uniformly modeling camera perspective geometry and kinematics feedforward to accurately control exposure. The application eliminates parameter cross-coupling through global control and local layer single-source adjustment, and improves the component recognition accuracy and linkage accuracy of high dynamic inspection under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of lidar technology, specifically a lidar-based contact network target recognition algorithm. Background Technology

[0002] As a core component of electrified railways, the railway catenary's operational status directly affects the safety and stability of current collection by the train's pantograph and catenary. Erected along the railway line, the catenary is constantly exposed to a complex natural environment, enduring the continuous effects of train operation impacts, environmental corrosion, and material aging. Its geometric parameters and structural integrity gradually change over time. Abnormal conditions such as changes in contact wire height, broken droppers, misaligned positioners, and damaged section insulators, if not detected and addressed promptly, can lead to pantograph-catenary failures and even serious railway accidents. Therefore, regular, comprehensive, and accurate inspection and condition assessment of the catenary facilities has always been a crucial aspect of railway power supply equipment maintenance, and the reliability of the inspection results directly contributes to the safety and efficiency of railway transportation.

[0003] In existing technologies, some solutions have attempted to perform parameter detection through point cloud analysis. For example, Chinese invention patent application publication number CN103852011A discloses a method for analyzing the geometric parameters of railway catenary based on lidar. This method mainly utilizes the characteristics of the spatial distribution of the contact line point cloud, such as finding the minimum guide height value in the data or using prior information about the spatial position of the columns, to analyze the geometric parameters of the catenary. Furthermore, Chinese invention patent authorization announcement number CN108564575B also discloses a non-contact catenary parameter detection method based on three-dimensional point cloud data. This method preprocesses the collected three-dimensional point cloud data of the catenary in stages, successively removing noise points and background interference, and then uses predetermined geometric space filtering and standard one-dimensional straight line model fitting to extract the contact line and calculate relevant parameters.

[0004] However, the aforementioned existing solutions have significant limitations in adaptability to complex real-world applications. Specifically, in actual operation, the contact wire and catenary inevitably experience changes in flexible suspension sag or localized wave deformation under the continuous effects of complex dynamic impacts, drastic changes in ambient temperature, and fluctuations in cable self-weight tension. The ideal rigid linear fitting model commonly relied upon in the aforementioned prior patents deviates significantly from the actual flexible suspension physical form of each cable entity in the contact network, easily leading to distortion in cable spatial trajectory extraction. This, in turn, causes a chain reaction of serious errors in the subsequent positioning and identification of related components based on this spatial reference. Furthermore, due to the dynamic operation of the detection vehicle, laser beam divergence, non-uniform reflectivity of component surfaces, and physical environmental interference such as special cross-overlapping obstructions, slender components like droppers often exhibit severe spatial density non-uniformity or manifold fracture morphology in discrete point cloud data. Traditional fixed threshold filtering or conventional standard clustering methods are prone to oversegmentation in sparse areas or undersegmentation in areas where cables intersect and adhere, resulting in frequent component omissions and misidentifications.

[0005] Furthermore, existing technologies typically employ a uniform, fixed processing threshold across the entire process, failing to adapt to the heterogeneous variations in complex scenarios such as confined spaces like tunnels, high-altitude obstacles, and sudden changes in pull-out values ​​during curves. Simultaneously, during high-speed inspections, the system lacks a feedforward linkage mechanism between the high-dynamic camera's perspective geometry and the vehicle's one-dimensional kinematic state, making it difficult to accurately predict the appropriate moment for image capture under asynchronous spatiotemporal conditions. This results in frequent issues such as targets deviating from the center of the field of view or blurred edges in the reviewed visible light images.

[0006] Therefore, developing a contact network target recognition method that can maintain high extraction accuracy and high scene adaptability under various complex working conditions has become a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0007] The purpose of this invention is to provide a contact wire target recognition algorithm based on lidar, in order to solve the problems of poor adaptability of contact wire morphology, low reliability of small component extraction, insufficient adaptability in complex scene processing, and poor accuracy of detection and inspection linkage in the prior art.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A contact wire target recognition algorithm based on lidar includes the following steps: Step 1: Acquire high-density, high-precision spatial point cloud data based on 3D LiDAR and perform preprocessing. The preprocessing includes downsampling the point cloud using an adaptive stepless variable resolution voxel filtering model that decouples the instantaneous driving speed of the inspection vehicle from the inherent optical parameters of the radar, and extracting features from the downsampled point cloud based on a multi-scale sparse feature aggregation method to obtain an initial sparse feature map containing multi-scale geometric topology information. Step 2: Based on the preset spatial region of interest, the preprocessed point cloud data is filtered and truncated to obtain the core region point cloud data. The spatial region of interest is constructed as a time-varying cross-sectional dynamic envelope boundary dynamically corrected by the vertical and lateral vibration displacement response of the vehicle body. Step 3: Process the point cloud data of the core area based on the adaptive morphology fitting algorithm, distinguish and extract the spatial trajectory lines of the contact line and the catenary. The adaptive morphology fitting algorithm uses a trajectory deformation fusion optimization model for deformation compensation. Step 4: Delineate the search area based on the spatial trajectory lines of the contact line and the catenary, and extract the dropper point cloud data and the locator point cloud data respectively within the search area; Step 5: The initial sparse feature map obtained in Step 1 is dimensionality reduced based on the lightweight representation algorithm to generate a lightweight representation state vector; the lightweight representation state vector is input into a multi-label scene classifier for scene recognition, and the global baseline parameters in the algorithm are dynamically corrected according to the output scene recognition results to realize the single-source hierarchical adjustment of the global scene control layer and the local dynamic solution layer.

[0009] According to the above technical solution, in step one, the adaptive stepless variable resolution voxel filtering model defines the three-dimensional side length vector of the dynamic adaptive voxel grid; wherein, the voxel step size along the railway line extension direction is the maximum value of the ratio of the instantaneous travel speed of the inspection vehicle to the fixed scanning frequency of the lidar, and the product of the maximum ranging distance from the current radar to the target area of ​​the contact network and the fixed angular resolution of the lidar; the voxel step size perpendicular to the line direction and the height direction are equal, and both are taken as the algebraic product of the radial distance of the current scanned transient target point, the fixed angular resolution of the lidar, and the preset point cloud density adjustment factor.

[0010] According to the above technical solution, in step one, the implementation process of the multi-scale sparse feature aggregation method is as follows: a three-dimensional sparse tensor is constructed based on the voxel grid after adaptive voxel filtering, and multiple local sparse receptive fields with different expansion rates are set; within the scale space of each receptive field, sparse three-dimensional convolution operation is performed only on non-empty voxels, and the local curvature and normal vector gradient features extracted at each scale are concatenated along the channel dimension to complete the multi-scale sparse feature aggregation.

[0011] According to the above technical solution, in step two, the algebraic solution control rule for the time-varying cross-sectional dynamic envelope boundary is as follows: The upper and lower bounds of the time-varying vertical height of the region of interest are determined by adding and subtracting the vertical correction increment at the current moment from the highest and lowest installation heights of the overhead contact line standard design. The vertical correction increment is obtained by summing the instantaneous vertical vibration displacement generated by the integration of the vehicle-mounted inertial measurement unit and the redundant term of the radar ranging constant. The time-varying lateral left and right boundaries of the region of interest are determined by algebraically superimposing the lateral vibration displacement derived from the inertial measurement unit based on the lateral reference half-width of the track centerline, and further superimposing the roll compensation term determined by multiplying the sine of the vehicle roll angle by the highest installation height of the overhead contact line.

[0012] According to the above technical solution, in step three, the objective function of the trajectory deformation fusion optimization model integrates the physical curvature penalty term based on elastic mechanical boundary constraints and the local error penalty term based on robust loss piecewise function; The adaptive morphology fitting algorithm also includes a soft threshold independent dynamic adjustment module. This module calculates the statistical variance of the point cloud height residual within the current moving fitting window, combines it with the discrete continuity index along the railway line extension direction, and independently scales and locks the implicit soft threshold in the robust loss piecewise function through piecewise linear mapping rules.

[0013] According to the above technical solution, in step four, the process of extracting dropper point cloud data includes: within the dropper search bounding box constructed by the contact wire and the catenary cable, hierarchical clustering is implemented using an adaptive clustering radius mapping model with local density inverse characteristics of the point cloud. The adaptive clustering radius of each sampling point is dynamically calculated based on the exponential power of the ratio factor between the global baseline clustering radius and the estimated maximum value of the local point cloud Gaussian kernel density. Subsequently, the generated clusters are verified by dual algebraic criteria. The dual algebraic criteria include a one-dimensional linear manifold verification based on the ratio of eigenvalues ​​of the cluster space covariance matrix, and an angular boundary verification based on the dot product of the first-order principal direction eigenvector and the vertical unit vector.

[0014] According to the above technical solution, in step four, the process of extracting the locator point cloud data includes: extracting a set of oblique candidate point clouds based on the spatial angle between the locator and the contact line and the coplanarity of the normal vectors, and constructing a rigid body judgment model composed of spatial compactness features and normalized orthogonal projection Shannon information entropy for integrity verification; wherein, the spatial compactness features are obtained by solving the ratio of the maximum eigenvalue to the minimum eigenvalue of the two-dimensional centered deviation covariance matrix constructed on the orthogonal projection plane of the candidate point clouds; the normalized orthogonal projection Shannon information entropy is obtained by performing a negative summation operation on the normalized distribution probability of all non-space grids and its base-2 logarithmic operator in the full grid domain after the projection plane is gridded.

[0015] According to the above technical solution, in step five, the lightweight representation algorithm processes as follows: calculate the feature tensor covariance matrix of the multi-scale sparse feature map, remove redundant channels whose contribution weights to scene classification are lower than a preset minimum threshold through an adaptive channel pruning algorithm, and project the retained high-dimensional feature vectors onto a low-rank manifold space using a nonlinear mapping function to generate a low-dimensional lightweight representation state vector; the multi-label scene classifier outputs a multi-label state vector representing one or more coexisting states among straight line segments, curved segments, tunnel scenes, and crossing scenes based on the lightweight representation state vector.

[0016] According to the above technical solution, in step five, the implementation process of the single-source hierarchical adjustment is as follows: based on the activation flags of each scene in the output multi-label state vector, a diagonal parameter state transition correction matrix is ​​constructed by analytical mapping calculation; the parameter state transition correction matrix is ​​multiplied by the global scene control vector containing multi-level fitting hard thresholds and benchmark clustering radius, so as to dynamically and adaptively scale the global benchmark parameters without interfering with the soft threshold solution of local trajectory deformation.

[0017] According to the above technical solution, the algorithm also includes step six for feedforward control of the visible light camera: unifying the perspective geometry equation of the camera extrinsic parameters with the one-dimensional vehicle kinematic recursive equation to solve the longitudinal span of the trigger mileage constrained by the image center coordinates; calculating the instantaneous mileage residual between the current mileage of the inspection vehicle and the mileage position of the target component in real time; generating a hard trigger level signal when the instantaneous mileage residual reaches the dynamic trigger critical threshold; the dynamic trigger critical threshold compensates for the instantaneous driving speed, instantaneous one-dimensional acceleration of the inspection vehicle, and the comprehensive exposure delay time of the system hardware and software.

[0018] Compared with the prior art, the present invention has the following beneficial effects: In this invention, during the preprocessing stage, an adaptive stepless variable resolution voxel filtering model is used to decouple the voxel step size from the instantaneous vehicle speed and radar optical parameters in real time. The step size along the line direction is adaptive with the vehicle speed, and the step size in the lateral and vertical directions is adaptive with the target distance, eliminating the non-uniformity of point cloud density caused by vehicle speed fluctuations, and ensuring that the downsampled point cloud density remains normalized under different vehicle speeds and ranging conditions. The vehicle-mounted inertial measurement unit is used to calculate vehicle vibration and roll in real time, dynamically adjusting the boundary of the region of interest to avoid point cloud loss during high-speed shaking or cornering and suppress background noise.

[0019] In morphological fitting, a nonlinear optimization model combining Huber robust loss and physical curvature penalty is constructed, and local independent soft threshold dynamic adjustment is adopted to effectively suppress pantograph-catenary impact and noise interference, accurately compensate for contact line sag, and improve trajectory fitting accuracy and robustness. During dropper extraction, Gaussian kernel density estimation and density inverse adaptive clustering radius are integrated, along with dual criteria of linear manifold and vertical direction based on three-dimensional principal component analysis (PCA), to avoid over-segmentation and under-segmentation of small components.

[0020] The locator verification introduces a rigid body judgment model composed of a two-dimensional compactness matrix and projection information entropy, transforming volume and area-based geometric threshold judgments into projection distribution entropy measurements, thus enhancing recognition capabilities under fragmented occlusion. By lightweighting multi-scale sparse features and using a multi-label random forest classifier to identify complex scenes, a single-source hierarchical adjustment mechanism decouples global hard threshold correction from local soft threshold correction, achieving adaptive convergence across all operating conditions. For trigger control, the camera perspective projection and vehicle kinematics feedforward equations are combined to derive an analytical solution for trigger mileage and compensate for velocity and acceleration terms, eliminating spatiotemporal asynchronous errors and ensuring the target is located in the center of the imaging field of view at the moment of image exposure. Attached Figure Description

[0021] Figure 1 This is a flowchart of the overhead contact line target recognition algorithm of the present invention; Figure 2 This is a flowchart of the point cloud preprocessing process of the present invention; Figure 3 This is a flowchart illustrating the processing of the region of interest in space according to the present invention; Figure 4 This is a flowchart of the trajectory extraction and target recognition process of the present invention; Figure 5 This is a flowchart illustrating the lightweight characterization and scene control processes of this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1: like Figures 1 to 5As shown, in one specific implementation, the contact wire target recognition algorithm based on lidar first requires the establishment of a complete hardware detection system. This system includes a multi-line lidar, a visible light camera, an onboard inertial measurement unit, and a supporting data processing unit. Specifically, the multi-line lidar uses a lidar device with no less than 32 lines, a scanning frequency of no less than 10Hz, and a ranging accuracy better than 2cm. This lidar device is installed on the roof of the railway inspection vehicle, with the scanning direction along the extension direction of the railway line. The visible light camera is also installed on the inspection vehicle to collect two-dimensional image information of the contact wire components, providing image data support for subsequent data fusion. The onboard inertial measurement unit is installed on the inspection vehicle body to collect the three-axis acceleration and attitude angle signals of the vehicle body in real time, and output the vertical vibration displacement, lateral vibration displacement, and roll angle data of the vehicle body, providing input for the dynamic correction of the boundary of the region of interest in space. The data processing unit uses a high-performance industrial computer, configured with an independent graphics processing unit (GPU) for real-time high-concurrency processing of point cloud data. Furthermore, the algorithm is implemented using the C++ programming language, with the compilation standard being C++14. The underlying point cloud processing algorithm calls the PCL point cloud processing library, preferably version 1.12.0 or higher.

[0024] like Figure 2 As shown, under hardware system operation, the method of this invention first preprocesses the raw 3D point cloud data collected by the multi-line lidar to obtain preprocessed point cloud data. Preprocessing includes coordinate transformation, noise filtering, and point cloud downsampling. Specifically, coordinate transformation converts the raw point cloud data collected by the lidar from the local radar coordinate system to the geographic coordinate system, ensuring that all subsequent location calculations can be expressed using a unified geographic coordinate system. Noise filtering employs a statistical outlier removal method, filtering isolated noise points caused by airborne particles or equipment vibration by setting neighborhood search parameters. In a preferred embodiment, the specific parameters are set to 50 search neighborhood points and a standard deviation threshold of 2.0. Furthermore, to eliminate the point cloud density non-uniformity caused by the high dynamic fluctuations in the inspection vehicle's speed, point cloud downsampling employs an adaptive continuously variable resolution voxel filtering model. Because the inspection vehicle's speed varies significantly during actual operation... It changes dynamically, while the scanning frequency of the lidar... and angular resolution The constant density of the point cloud leads to uneven spatial density. To address this, a three-dimensional side length vector of a dynamically adaptive voxel mesh is defined based on a real-time decoupling logic between vehicle speed and inherent radar optical parameters. Its mathematical expression is:

[0025] ;

[0026] ; In the formula, The instantaneous time of the inspection vehicle's dynamic scanning. For the corresponding The 3D voxel grid step size vector at time t. The voxel step size is the distance along the direction of the railway line. and These are the voxel step sizes perpendicular to the path direction and the height direction, respectively. Let be a function of the instantaneous speed of the inspection vehicle. This refers to the fixed scanning frequency of the lidar. The fixed angular resolution for lidar, This represents the maximum ranging distance from the radar to the target area of ​​the overhead contact line. The preset point cloud density adjustment factor. The radial distance of the transient target point during the current scan, indicated by the superscript. Represents the transpose operation of a matrix or vector; preferably, under the reference straight line condition, the voxel step size is perpendicular to the line direction and the height direction. , Initialized to 0.02m, the voxel step size along the railway line extension direction. No fixed initial value is set, but rather... The model is determined by real-time vehicle speed calculation; the step size along the route is adaptive with vehicle speed and the step size in the horizontal and vertical directions is adaptive with target distance. The density of the downsampled point cloud remains normalized under different vehicle speeds and ranging conditions to eliminate morphological distortion in subsequent feature extraction.

[0027] Furthermore, after completing the preprocessing, such as Figure 3As shown, this invention filters and extracts preprocessed point cloud data based on a preset spatial region of interest (ROI) to obtain core region point cloud data. Specifically, the ROI constructs a dynamic envelope topology model based on the orbital geometric dynamics boundary. The vertical height and lateral width boundaries of this model are not static boundaries, but are dynamically controlled through algebraic calculations. Specifically, the time-varying upper bound of the ROI is determined by algebraically superimposing the instantaneous vertical vibration displacement generated by the Z-axis accelerometer in the inertial measurement unit through double time integration from the highest installation height of the contact wire standard design, and further superimposing a constant redundancy term characterizing the random error of radar ranging; the time-varying lower bound of the ROI is determined by algebraically subtracting the instantaneous vertical vibration displacement from the lowest installation height of the contact wire standard design, and further subtracting a constant redundancy term. Similarly, the time-varying lateral right boundary of the region of interest is determined by algebraically superimposing the lateral vibration displacement derived from the integration of the Y-axis accelerometer in the aforementioned inertial measurement unit, based on the lateral reference half-width of the track centerline, and further superimposing a roll compensation term determined by multiplying the sine of the vehicle roll angle by the highest installation height of the overhead contact line; the time-varying lateral left boundary of the region of interest is determined by algebraically adding the lateral vibration displacement based on the negative lateral reference half-width of the track centerline, and further algebraically subtracting the roll compensation term. In a specific reference example, its vertical height range dynamic envelope is between 4.5m and 7.0m, the total width dynamic envelope perpendicular to the track direction is 10m, and the filter intercept length benchmark along the railway line direction is set to 100m. This dynamic envelope boundary ensures that the core point cloud is not lost when the train is swaying at high speed or going through curves, and effectively removes unnecessary external environmental background noise.

[0028] Furthermore, such as Figure 4 As shown, after obtaining the core area point cloud data, this invention processes the core area point cloud data based on an adaptive morphological fitting algorithm to extract the contact wire spatial trajectory line and the catenary spatial trajectory line. Specifically, the adaptive morphological fitting algorithm includes a multi-level fitting screening module, a sag compensation module, and a dynamic threshold adjustment module. It extracts the contact wire and catenary separately by distinguishing the spatial height difference between them. In a specific example, the contact wire height range is initially set to 5.0m to 5.8m under the baseline working condition, and the catenary height range is initially set to 5.8m to 7.0m. The point cloud in the overlapping area participates in the fitting calculation of both. Specifically, the processing of the multi-level fitting screening module is divided into three levels. The first level performs a first-level hard threshold straight line fitting on the core area point cloud data to obtain an initial fitting straight line, and selects lines with a vertical distance less than the first hard threshold. The points are retained as the first set of filtered points, and the first hard threshold is set. The initial value is set to 0.15m; the second level performs a second-level hard threshold line fitting on the first set of selected points to obtain a corrected fitting line; the third level sets the distance from the corrected fitting line to the second hard threshold. The points are retained as effective contact line point clouds or effective load-bearing cable point clouds, with a second hard threshold. The initial state is set to 0.08m.

[0029] Furthermore, to address the flexible sag deformation of the contact wire cable due to its own weight tension, and the interference from high-frequency airflow and residual pantograph vibration during dynamic inspections, the sag compensation module constructs a nonlinear curve optimization model based on dynamic constraints and an error-insensitive robust loss function. The optimized nonlinear objective function is specifically expressed as follows:

[0030] ; In the formula, , , The coefficients of the quadratic curve model equation to be solved; Let i be the mileage coordinates of the i-th point cloud sample within the current moving fitting window along the railway line. For height coordinates, The total number of point clouds participating in the flexible fitting within the moving fitting window serves as the upper bound parameter for the summation; i is the discrete cyclic index variable of the point cloud samples. This is the stiffness-elasticity weighted coefficient for the cable, used to balance the physical curvature penalty term and the residual penalty term. Its dimensions are specified according to the principle that the two dimensions are homogeneous. The mileage length of the current moving fitting window is obtained by subtracting the starting mileage from the end mileage of the window. In the quadratic curve fitting form, the physical curvature penalty term is simplified to a stiffness-elasticity weighted coefficient. The square of the second derivative constant term 2A of the quadratic parabola model and the length of the window mileage. The algebraic product of the three factors is used to limit the abrupt change in the second derivative of the fitted curve to conform to the continuous deformation boundary. The independent variable of the introduced Huber robust loss piecewise function is... Characterized as the geometric residual from discrete sampling points to the fitted manifold curve, i.e., satisfying = This is used to apply piecewise penalty to the vertical height residual of the spatial point cloud within the window, and its mathematical expression is: When | |≤ ; When | |> ; In the formula, It serves as the implicit soft threshold for the Huber robust loss function, and also as the penalty switching point for outlier high-frequency vibration points within the current moving window, controlling the boundary of spatial geometric distance.

[0031] Furthermore, to ensure closed-loop self-consistency between the robust estimation system and macroscopic environmental noise fluctuations, the dynamic threshold adjustment module employs local adaptive control. It adjusts the soft threshold of the robust loss function based on the distribution characteristics of the point cloud height residuals within the current moving window. Independent dynamic adjustments are made, with specific quantitative adjustment rules based on piecewise linear mapping functions and discrete continuity checks. =0.08m, when σ 2 ≤0.015m 2 ; =0.08m+(σ 2 -0.015m 2 ) / (0.020m 2 )×0.07m, when 0.015m 2 <σ 2 ≤0.035m 2 ; =0.18m, when σ 2 >0.035m 2 or <80%; In the formula, σ 2 This is the statistical variance of the point cloud height residuals within the current moving window, with dimensions m. 2 ; The discrete continuity index along the railway line is defined as the percentage of mileage segments containing valid point clouds to the total number of segments. The quantitative logic of this model lies in: when the statistical variance σ... 2 ≤0.015m 2 When the signal-to-noise ratio of the current point cloud is good, the soft threshold is applied. The baseline value is locked at 0.08m; when the statistical variance is 0.015m. 2 up to 0.035m 2 When the threshold is between 1 and 2, it indicates that there is local noise degradation due to environmental variations, and the soft threshold is used. The variance is monotonically and positively correlated with the linear scaling between 0.08m and 0.15m to broaden the flexible acceptance corridor of the effective point cloud and avoid effective points being misclassified as outliers; however, when the statistical variance further increases beyond the catastrophic variance threshold of 0.035m... 2 or continuous indicators When the threshold drops below 80%, it indicates that the current area is experiencing severe local noise pollution. The algorithm triggers a high-noise safety protection mechanism, lowering the soft threshold. The threshold is directly raised and locked to a safe upper limit of 0.18m, thus preventing the loss function from diverging in weights over noisy regions. This soft threshold... It is solved independently by the one-dimensional piecewise function of the local dynamic threshold adjustment module, without being affected by external global scene control factors.

[0032] Furthermore, after extracting the contact wire spatial trajectory line and the catenary spatial trajectory line, this invention constructs a dropper search region based on these two lines, and extracts dropper point cloud data within the search region using an improved hierarchical clustering method. Specifically, the method for constructing the dropper search region involves using the contact wire spatial trajectory line as the lower boundary and the catenary spatial trajectory line as the upper boundary, extending a preset width to both sides in a direction perpendicular to the railway line. In a specific example, this extension is 0.3m to both sides, forming a strip-shaped region with a total width of 0.6m, and creating a spatial bounding box for the dropper search region within this strip. Specifically, the improved hierarchical clustering method includes an adaptive clustering parameter determination step based on point cloud density distribution and a clustering result verification step based on spatial constraints. It also includes merging processing for undersegmentation and splitting processing for oversegmentation. Merging is performed when the distance between the centers of two clusters is less than a merging threshold, which is preferably set to 0.1m.

[0033] Furthermore, such as Figure 4 As shown, to eliminate the uneven density of the point cloud caused by laser beam divergence and differences in surface reflectivity of components, this step first introduces a continuous spatial Gaussian kernel density estimation model within the string search region to quantitatively describe the local spatial distribution probability of the point cloud. Specifically, the algorithm quantitatively performs the following recursive calculations: First, within a set radius and a smoothing bandwidth... Within a spherical neighborhood, calculate the coordinates of the current target coordinate vector and the coordinates of the target vector within the neighborhood. The three-dimensional Euclidean distance between the spatial coordinate vectors of neighboring points is calculated, and this distance is squared to obtain the squared distance term. Next, this squared distance term is divided by twice the squared smoothing bandwidth term, and the negative algebraic quotient is used as the base of the natural logarithm. The exponent is raised to a power to obtain the Gaussian decay kernel function value at a single point; then, the Gaussian decay kernel function values ​​are summed over the entire neighborhood, and the sum is divided by the total number of neighboring points in the neighborhood. To obtain the neighborhood average kernel response, the neighborhood average kernel response is then divided by the smoothing bandwidth. cubed and multiplied by twice pi The normalization constant coefficients obtained by taking the cube root of the value and then taking its reciprocal are used to quantitatively calculate the Gaussian kernel density estimate of the discrete sampling points in a stepless and smooth manner. The estimated kernel density is directly used as the Gaussian kernel density maxima within the subsequent local search window. And the algebraic solution basis for single-point transient estimates. Furthermore, to avoid the undersegmentation and oversegmentation problems caused by traditional fixed-radius clustering, this invention uses an adaptive clustering radius... An explicit construction of a continuous inverse-proportional transformation function for the spatial density of the current sampling point, satisfying the analytical expression:

[0034] ;

[0035] ; In the formula, Let be the coordinate vector of any point to be classified within the search area. For the current sampling point Adaptive clustering radius, The baseline cluster radius is preferably set to 0.05m initially, and the dynamic shrinkage boundary is limited to between 0.02m and 0.15m, while the nonlinear adjustment index... With preset density coefficient Satisfying algebraic mapping relations =1.5-k, and in a specific preferred example, the density coefficient k is set to 0.5. Through this nonlinear inverse proportional transformation, the cluster radius is infinitely scaled with the spatial density. When the point cloud becomes sparse, the cluster radius automatically expands to avoid oversegmentation, and conversely, the cluster radius shrinks inward to avoid undersegmentation.

[0036] Furthermore, regarding the clustering result verification step, this invention introduces a spatial manifold eigenvalue constraint model based on three-dimensional principal component analysis. Specifically, the algorithm first constructs a 3×3 covariance matrix C of the spatial coordinates of each generated cluster, the mathematical definition of which is:

[0037] ; In the formula, This represents the total number of point clouds within the current cluster to be verified. For the traversal index of the points within the cluster, Let m be the three-dimensional spatial coordinate vector of the m-th point within the cluster. Let C be the three-dimensional vector representing the geometric center of this cluster. By performing eigenvalue decomposition on the covariance matrix C, we obtain three eigenvalues ​​sorted in descending order. ≥ ≥ The algorithm then extracts the corresponding feature vectors. Based on the physical property that the suspension strings have a one-dimensional linear vertical distribution, the algorithm performs a double algebraic inequality criterion verification: ,and ; In the formula, The preset linearity control threshold is preferably set to 3.0; The largest eigenvalue The corresponding normalized first-order principal direction eigenvector; The standard vertical axis vector of the geodetic coordinate system is represented as: ; The maximum allowable tilt angle for the physical installation of the overhead contact line. Only clusters that simultaneously satisfy the above one-dimensional linear manifold constraints and vertical direction constraints are determined to be valid droppers. In a specific preferred example, the spatial geometric discrimination range of the reference example is synchronously limited to a height direction between 0.8m and 2.0m, a horizontal extension length between 0.1m and 1.5m, and an aspect ratio greater than 3.0.

[0038] Furthermore, this invention extracts point cloud data of the locator based on the spatial structural characteristics of the locator and the contact line. These spatial structural characteristics include the angular and positional relationships between the locator and the contact line. Specifically, the extraction method involves first establishing a local frame based on the continuous tangent vector of the contact line trajectory space, and then estimating the spatial directional gradient field of the discrete point set within a preset range of 1.0m on each side of the contact line trajectory. Geometric mapping is then performed using the tensor product between the first-order principal direction vector of the point cloud spatial distribution covariance matrix and the horizontal transverse vector of the track. This allows for the search and extraction of oblique point cloud sets that satisfy the spatial angle and whose spatial boundary normal vector is coplanar with the contact line normal vector. In a preferred embodiment, the angular characteristics are characterized by an angle between 15° and 75°.

[0039] Furthermore, such as Figure 4 As shown, to eliminate false detections, structural integrity verification constructs a rigid body determination model based on spatial compactness features and orthogonal projection information entropy. Specifically, it first extracts the set of two-dimensional coordinates of the candidate point cloud on the current transverse orthogonal projection plane ( , ),in Let be the traversal index of the candidate point cloud, and let be the total number of points. Specifically, the two-dimensional centering bias is obtained by calculating the difference between the lateral coordinate and the mean of the lateral geometric center for each projection point, and the difference between the lateral coordinate and the mean of the vertical geometric center for each projection point. Then, the lateral centering bias of each projection point is squared and averaged over the entire cluster to calculate the lateral projection variance, which is the first element of the main diagonal. Similarly, the longitudinal centralization deviation of each projection point is squared and averaged over the entire cluster to calculate the longitudinal projection variance, which is the second element of the main diagonal. Furthermore, the lateral and longitudinal centralization deviations of each projection point are algebraically multiplied, and the product is averaged across the entire cluster to calculate the average cross product of the lateral and longitudinal coordinate centralization deviations as off-diagonal elements. and Therefore, a 2×2 spatial compactness matrix determined by the discrete spatial distribution of the candidate point cloud is constructed. :

[0040]

[0041] Each independent element of the matrix is ​​quantitatively calculated using the following formula:

[0042] ;

[0043] ;

[0044] ; In the formula, and These are the mean values ​​of the horizontal and vertical geometric centers of the current candidate point cloud on the projection plane, respectively. The two eigenvalues ​​of this matrix are then solved. ≥ The ratio of the largest to the smallest eigenvalue is used as a compactness index for one-dimensional linear rigid constraint verification, with the eigenvalue limit preferably set to be greater than or equal to 4.5. This verifies whether the candidate point cloud satisfies the one-dimensional continuous distribution characteristics of the locator fitting. Simultaneously, the candidate point cloud is orthogonally projected onto a two-dimensional plane along the main direction of the locator, and the projection plane is probabilistically mapped to a grid using the average resolution of the point cloud space as the reference grid size. During gridding, the total number of non-empty grid cells is set to... The value should be between 6 and 45 to ensure that the following feature entropy range is reachable; specifically, count the number of points falling into each two-dimensional projected grid and divide by the total number of points in that candidate cluster. The normalized distribution probability is obtained. Subsequently, all non-space grids are traversed within the full grid domain, and the normalized distribution probability of each non-space grid is calculated. The information entropy components of a single grid are obtained by algebraically multiplying the logarithmic values ​​calculated by the base-2 logarithmic operator. Finally, the information entropy components of all single grids in the entire grid domain are summed, and the sum is multiplied by negative one to perform a negative summation operation, thereby solving for the normalized orthogonal projection Shannon information entropy. :

[0045] ; In the formula, This represents the total number of two-dimensional projected grids divided on the orthogonal plane. This is the index variable for traversing the grid. This represents the information entropy calculated using the base-2 logarithmic operator. The unit is bit. Preferably, the characteristic entropy value range of the rigid rod entity is set to 2.5 to 5.5 bits. This determination logic uses volume and area measurements (the envelope hard threshold of the benchmark example corresponds to a spatial volume greater than 0.0005m) 3 And the projected area is greater than 0.005m² 2 The projection distribution entropy, which reflects the physical rigidity distribution characteristics of the component, is converted into a robust extraction under fragmented occlusion conditions.

[0046] Furthermore, this invention uses a scene classification model to identify the current detection segment and determine the current scene type, which includes straight segments, curved segments, tunnel scenes, and crossing scenes. In contrast, conventional implementations can directly employ a random forest classifier based on point cloud statistical features. This approach directly stacks multidimensional statistical features, defining the point cloud comprehensive feature state vector of the current detection segment as a six-dimensional feature vector. Specifically, this vector includes the mean Shannon entropy, variance Shannon entropy, variance of vertical spatial distribution, variance of horizontal spatial distribution, mean of normalized reflectance intensity distribution, and standard deviation of normalized reflectance intensity distribution on the height-level distribution. In this embodiment of the invention, the multi-label scene classifier does not use the six-dimensional statistical feature vector as input, but instead uses the state vector obtained through lightweight representation in step five above as input. Preferably, in order to support the physical reality of multiple overlapping scenes in complex lines, the scene classification model does not adopt a mutually exclusive single-label classification mode. Its output is defined as a multi-label state vector, which corresponds to the binary activation flags of straight sections, curved sections, tunnel scenes, and crossing scenes, respectively. When the model output shows that multiple flags are activated at the same time, the current composite scene type is determined. Specifically, the multi-label random forest classifier can contain 100 decision trees, and the maximum depth of each decision tree is 10 layers.

[0047] Furthermore, such as Figure 5 As shown, this invention adaptively adjusts the recognition algorithm parameters according to the scene type to obtain the adjusted recognition result. To eliminate the cross-coupling between the global scene control and the local estimation system, this invention implements a layered adjustment of the global scene control layer and the local dynamic solution layer. The single-source layered adjustment means that the global benchmark parameters are calculated only by the multi-label state vector through the parameter state transition correction matrix, and the local soft threshold is not connected to this path. Specifically, this invention defines a global scene control vector that is bound to the multi-level fitting screening hard threshold and the preprocessing step size. ,in The initial voxel step size is perpendicular to both the path direction and the height direction. The first hard threshold for multi-level fitting, The second hard threshold for multi-level fitting, The baseline cluster radius. Control vector. It is only related to the hard filtering boundaries of multi-level screening and the switching of external geometric scenes. This invention introduces a closed-loop parameter state transition correction matrix derived based on a large-scale environmental scene. global scene control vector The macroscopic system parameter control model for implementing dynamic multiplier correction in matrix multiplication satisfies:

[0048] ; In the formula, = This is the preset initial global control vector; This is the corrected output control vector; The closed-loop parameter state transition correction matrix is ​​constructed as a 4×4 diagonal matrix, i.e. = .in, This is the downsampling step size correction coefficient. This is the first hard threshold correction coefficient. This is the second hard threshold correction coefficient. This is the baseline clustering radius correction coefficient. Each diagonal element of this matrix is ​​derived from the multi-label state vector. The real-time mapping yields specific analytical calculation rules that satisfy: =1.0-0.5· ; =1.0+0.333· ; =1.0 + 0.5 ; =1.0. Where, and These are the activation flags for the corresponding curve segments and tunnel scenes in the multi-label state vector. Under the basic straight-line scene state ( The diagonal correction matrix is ​​the identity matrix. = (1.0, 1.0, 1.0, 1.0). When the detection section contains a tunnel scene ( When =1), the downsampling voxel step size coefficient =0.5, reducing the transverse voxel step size from 0.02m to 0.01m to preserve high-density details; when the detection segment contains a curved segment (s2=1), the first hard threshold correction coefficient is... =1.333, which adaptively increases the first hard threshold of the multi-level fitting from 0.15m to 0.2m, while the second hard threshold correction coefficient... =1.5, causing the second hard threshold to adaptively increase from 0.08m to 0.12m, in order to smooth the wide-range translation of the overall point cloud shape caused by the abrupt change in the curve pull-out value, thus combining it with the microscopic soft threshold calculated independently of the one-dimensional piecewise function. Mutual isolation. When the detection segment contains cross-scene ( When =1), the logic for hard exclusion of high-level crossing objects is activated synchronously. Regarding the composite working condition, each coefficient takes effect in synergy according to this flow-division model, effectively avoiding logical interference with the local robust loss penalty mechanism while ensuring global boundary adaptability.

[0049] Furthermore, the mileage position information of the target component is calculated based on the contact wire spatial trajectory line, and this mileage position information is matched with the real-time position information of the inspection vehicle to determine the triggering time and send a trigger signal. Specifically, the method for calculating the mileage position information is to map points on the contact wire spatial trajectory line to a mileage coordinate system to obtain the mileage value corresponding to each point, and the mileage value of the projection point of the target component on the contact wire spatial trajectory line is used as the mileage position information of the target component. To eliminate the influence of the inspection vehicle's travel speed... Instantaneous acceleration and system hardware and software response latency To address the resulting spatiotemporal asynchronous errors, this invention unifies the modeling of high-dynamic camera extrinsic perspective geometry and one-dimensional vehicle kinematics recursion, deriving a multi-objective, high-precision spatiotemporal synchronous control prediction model integrated with the optical center projection formula and kinematic feedforward compensation equation. This model specifically includes:

[0050] ;

[0051] ; In the above two formulas, The longitudinal span of the trigger mileage obtained by inverse solution. The focal length invariant is in the camera intrinsic parameter matrix. The x-coordinate of the optical geometric center corresponding to the camera's intrinsic parameters. and , Install the known calibrated rotation and translation elements from the extrinsic parameter matrix in camera space, respectively. , Let be the spatial center coordinates of the target component perpendicular to the track direction in the geodetic coordinate system, and let be the center coordinates of the target component along the track direction. Inverse solution determined; The instantaneous time of the current detection clock. For the corresponding The dynamic triggering critical threshold law equation at any given time. This is a function for the instantaneous travel speed of the inspection vehicle. For the instantaneous one-dimensional acceleration of the inspection vehicle, This refers to the overall exposure delay time for the system's hardware and software. During the high-dynamic operation of the inspection vehicle, the data processing unit dynamically calculates the instantaneous mileage residual between the current mileage value of the inspection vehicle and the target component using a high-frequency clock. When the residual reaches the value calculated by the aforementioned one-dimensional kinematic feedforward derivation model... Upon that, a hard trigger signal is immediately sent to the control system. Furthermore, based on the difference in physical longitudinal width in the spatial geometry of the target component, the signal is adjusted accordingly. The initial reference values ​​are set separately for each target component: 3m for the suspension wire target, 5m for the locator target, and 8m for the segmented insulator target. This feedforward prediction compensation ensures that the target is located in the center of the imaging field of view at the actual exposure time of the camera under high dynamic driving conditions, thus solving the technical problems of image edge blurring and target deviation from the center of the field of view under high dynamic inspection conditions.

[0052] Furthermore, the method of the present invention also includes a segmented insulator identification module, a middle anchor wire identification module, a data fusion step, and a result storage step. Specifically, the segmented insulator identification module extracts segmented insulator point cloud data based on lateral feature points on the contact wire spatial trajectory line. The lateral feature points are a set of points that are laterally distributed in the vertical direction of the contact wire spatial trajectory line and have a significantly higher reflection intensity than the surrounding point cloud. The middle anchor wire identification module extracts middle anchor wire point cloud data based on the anchoring point features on the catenary cable spatial trajectory line. The data fusion step fuses the lidar point cloud identification results with the image data acquired by the visible light camera to generate a comprehensive detection result containing three-dimensional position information and two-dimensional image information for each identified target component. The result storage step stores the identification results in a structured data format in a database.

[0053] Furthermore, in step one, to address the computational redundancy issue caused by high-density, high-precision point clouds, this invention introduces a multi-scale sparse feature aggregation method. Specifically, after completing adaptive variable-resolution voxel filtering, the center coordinates and reflection intensity of each effective voxel are extracted to construct an initial sparse tensor. Three three-dimensional spherical sparse receptive fields of different scales are set, with radii corresponding to 1, 2, and 4 times the initial voxel side length, respectively. Within each scale space, sparse three-dimensional convolution operations are performed only on non-empty voxels, and the local curvature and normal gradient features extracted at each scale are concatenated along the channel dimension. This achieves multi-scale sparse feature aggregation of local microstructures and macroscopic topologies without significantly increasing memory consumption.

[0054] Furthermore, in step five, to meet the real-time requirements of the vehicle-mounted edge computing platform, this invention employs a lightweight representation algorithm to process the aggregated high-dimensional features. Unlike traditional methods that directly stack multidimensional statistical features (such as Shannon entropy mean and spatial distribution variance), this invention maps the extracted high-dimensional multi-scale sparse feature map to a low-rank manifold space. First, the covariance matrix of the feature tensor is calculated, and redundant channels contributing to scene classification (with weights less than a preset minimum threshold) are removed using an adaptive channel pruning algorithm. Subsequently, a nonlinear mapping function is introduced. The retained feature vectors are projected into a low-dimensional space to generate a lightweight representation state vector with a final dimension of no more than 16. This lightweight representation not only retains the core discriminative power to distinguish between straight sections, curved sections, tunnel scenes, and crossing scenes, but also reduces the inference latency of the classifier, ensuring a real-time closed-loop response for subsequent dynamic parameter adjustments.

[0055] Example 2: In this embodiment, the invention was tested and verified on a typical straight section. Specifically, the test environment was a straight section of a railway passenger dedicated line. Data acquisition was performed using a 32-line lidar with a scanning frequency of 10Hz. The testing vehicle traveled at a speed of 60km / h. Under this condition, the voxel step size along the track direction was... The calculated height is 1.67m. After preprocessing, the collected point cloud data enters the adaptive morphological fitting algorithm. The first level of the multi-level fitting and filtering module performs linear fitting on the core region point cloud data. The initial fitted line has a slope of 0.003 and an intercept of 5.23m. After the first hard threshold filtering, the number of point clouds retained is 78% of the initial number. The second level performs corrective fitting on the first-filtered point set. The corrected fitted line has a slope of 0.002 and an intercept of 5.24m. After the second hard threshold filtering, the number of point clouds retained is 64% of the first-filtered point set. The dynamic threshold adjustment module determines the statistical variance of the point cloud height residual within the current sliding window. It is 0.012m 2 Because it is less than or equal to 0.015m 2The baseline variance control line was set, and the continuity index was 0.89. Therefore, the local estimation system determined it to be the baseline noise-free state, and the soft threshold δ of the Huber robust loss function remained unchanged at the initial baseline value of 0.08m. The sag value obtained by the quadratic curve fitting of the sag compensation module was 0.15m, which differed from the design sag value of 0.14m by 0.01m. In this embodiment, a total of 32 sets of dropper point cloud data were extracted, and after verification, 28 sets of effective dropper data were retained, with an effective extraction rate of 87.5%. A total of 16 sets of locator point cloud data were extracted, and after structural integrity verification, 15 sets of effective locators were retained, with an effective extraction rate of 93.75%. Since this embodiment only tests straight sections, it does not involve tunnels or curved sections.

[0056] Furthermore, comparison of test data shows that the contact line fitting error of this invention under straight-line conditions is only 0.01m, the deformation error rate is 7.1%, and the effective extraction rates of the dropper and the positioner reach 87.5% and 93.75%, respectively. The undersegmentation and oversegmentation rates of the algorithm are controlled at 6.25% and 3.125%, respectively. In contrast, Comparative Example 1, which uses a traditional fixed hard threshold straight-line fitting method (with a fixed first hard threshold of 0.15m and a second hard threshold of 0.08m, excluding multi-label global adaptive matrix and Huber local sag softening threshold feature compensation), shows an increased contact line fitting error to 0.08m and a deformation error rate of 57.1% under the same test data. The effective extraction rates of the dropper and the positioner also drop to 71.4% and 81.25%, respectively. Meanwhile, in Comparative Example 2 (with a fixed clustering radius of 0.08m, excluding radius adaptive scaling based on Gaussian kernel density and PCA manifold verification criteria), the undersegmentation and oversegmentation rates increased to 18.75% and 12.5%, respectively, due to the inability to adapt to the spatial inhomogeneity of the point cloud. This indicates that the present invention has higher geometric extraction accuracy and component isolation performance under basic operating conditions.

[0057] Example 3: This invention underwent practical testing and verification in complex curved sections combined with crossing scenarios. Specifically, the hardware detection system and underlying compilation environment relied upon in this embodiment reuse the vehicle-mounted integrated architecture from Embodiment 1. The test environment selected was a typical curved section of a railway freight line, characterized by a curve radius of 800m and a high-level interference scenario involving crossing power lines. Furthermore, the scenario classification model outputs a multi-label state vector of the current detection section. The system accurately identifies the current scene type as a curve segment combined with a crossing scene, with a confidence level of 92.5%. The parameter adaptive adjustment module initiates hierarchical control, dynamically adjusting the global parameter state transition correction matrix according to the scene type. Updated to a corrected diagonal matrix with diagonal elements of 1.0, 1.333, 1.5, and 1.0, thereby enabling the first hard threshold for multi-level fitting screening. The second hard threshold was extended from 0.15m to 0.20m. The depth was increased from 0.08m to 0.12m to absorb the macroscopic geometric translation caused by the curve. Specifically, due to the local high-frequency interference from the curve and crossing power lines, the point cloud distribution within the local window becomes discrete, and the statistical variance of its height residuals is increased. Reaching 0.025m 2 The variance is at 0.015m. 2 up to 0.035m 2 Within the segmented mapping interval, the local dynamic threshold adjustment module independently triggers one-dimensional algebraic computation to adjust the soft threshold of the Huber loss function. Linear smoothing was extended to 0.115m to avoid mechanically discarding useful microscopic sampling points at the trajectory edges according to a fixed threshold. The sag value fitted by the sag compensation module was 0.18m, which differed from the design sag value of 0.16m by 0.02m, with an error rate of 12.5%, which is within the acceptable range for high-speed railway inspection. In this embodiment, a total of 24 sets of effective droppers (efficiency 85.7%) and 13 sets of effective positioners (efficiency 92.9%) were extracted. In terms of trigger control, the established geometric projection perspective center equation was used for dynamic calculation. The positioning accuracy at the moment of taking the photo was subsequently verified to be better than 0.1m, which meets the actual inspection requirements.

[0058] Example 4: This invention was tested and verified in a tunnel scenario. Specifically, the test environment was a 1200m long tunnel on a railway passenger line, and the scene classification model identified it as a tunnel scene. Furthermore, the parameter adaptive adjustment module shrinks the adaptive voxel mesh size according to the state transition correction matrix, and its diagonal correction coefficient satisfies... =0.5, reducing the basic voxel step size perpendicular to the line direction and height direction from 0.02m to 0.01m to retain more details, while increasing the number of search neighborhood points for noise filtering to 80, and the hard threshold of multi-level fitting screening because no curve components were detected ( =0), its and The value is maintained at 1.0. Specifically, due to the smaller voxel size, the point cloud detail retention is improved, and the contact line inside the tunnel is less affected by airflow, resulting in less significant sag change. The fitting result of the sag compensation module can adaptively fit the cable deformation boundary. In this embodiment, the contact line fitting error is reduced to 0.008m, with an error rate of only 5.7%. At the same time, the effective extraction rate of the dropper is 89.3%, and the effective extraction rate of the locator is 95.2%. Furthermore, the data fusion module fuses the three-dimensional position information of the lidar with the two-dimensional image information of the visible light camera inside the tunnel, making up for the technical shortcomings of a single visible light sensor in the tunnel due to low imaging quality in weak light.

[0059] Example 5: This invention conducts comprehensive system integration testing. Specifically, the test system covers straight sections, curved sections, tunnel scenarios, and crossing scenarios. Furthermore, to verify the overall performance of the scene classification adaptive parameter adjustment module under all operating conditions, a comparative example three is introduced for comparison testing. Comparative example three uses uniform, fixed, one-size-fits-all recognition parameters under all operating conditions and does not include multi-label feature space state classification and parameter adaptive correction matrix.

[0060] Extensive testing results show that, under the adaptive parameter hierarchical flow-splitting drive mode, Example 5 achieved comprehensive identification accuracy rates of 91.5%, 89.2%, 86.8%, and 93.2% for the entire contact wire, catenary, droppers, and positioners, respectively. In complex tunnel environments, the contact wire fitting and positioner extraction accuracies remained at 88.6% and 86.3%, respectively; in curved sections with varying geometric curvature, the contact wire fitting and positioner extraction accuracies reached 90.2% and 88.7%, respectively. In open environments and complex working conditions, the image generation success rate of the data fusion step reached 97.3%, the result storage integrity rate was 100%, the triggering module executed 847 hard trigger level signal outputs, and the effective imaging synchronization rate reached 96.9%. In contrast, Comparative Example 3, which uses fixed parameters throughout the entire line, shows an overall decline in the comprehensive recognition accuracy of the contact wire, catenary, droppers, and locators, degrading to 76.4%, 74.1%, 71.5%, and 78.9%, respectively. Particularly in tunnel scenarios, due to the inability to adaptively decouple hard filtering boundaries from microscopically robust soft thresholds, the extraction accuracy of the contact wire and locators drops to 72.3% and 68.5%, respectively; in curved sections, it further decreases to 75.8% and 71.2%, respectively, failing to cope with feature degradation caused by highly dynamic multi-condition variations. The above cross-scenario test results across the entire road network demonstrate that introducing a multi-label state transition correction matrix under a hierarchical architecture has a compensatory effect on environmental heterogeneity.

[0061] As can be seen from the above test results, the present invention overcomes the problems of point cloud deformation and distortion, fragmented occlusion and image missing caused by asynchronous spatiotemporal axes under high dynamic inspection by using closed-loop feedforward solution of mathematical models of each module. Compared with the fixed parameter scheme of the whole line, it improves the geometric extraction accuracy and scene adaptability, and has engineering feasibility.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0063] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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 contact wire target recognition algorithm based on lidar, characterized in that: Includes the following steps: Step 1: Acquire high-density, high-precision spatial point cloud data based on 3D LiDAR and perform preprocessing. The preprocessing includes downsampling the point cloud using an adaptive stepless variable resolution voxel filtering model that decouples the instantaneous driving speed of the inspection vehicle from the inherent optical parameters of the radar, and extracting features from the downsampled point cloud based on a multi-scale sparse feature aggregation method to obtain an initial sparse feature map containing multi-scale geometric topology information. Step 2: Based on the preset spatial region of interest, the preprocessed point cloud data is filtered and truncated to obtain the core region point cloud data. The spatial region of interest is constructed as a time-varying cross-sectional dynamic envelope boundary dynamically corrected by the vertical and lateral vibration displacement response of the vehicle body. Step 3: Process the point cloud data of the core area based on the adaptive morphology fitting algorithm, distinguish and extract the spatial trajectory lines of the contact line and the catenary. The adaptive morphology fitting algorithm uses a trajectory deformation fusion optimization model for deformation compensation. Step 4: Delineate the search area based on the spatial trajectory lines of the contact line and the catenary, and extract the dropper point cloud data and the locator point cloud data respectively within the search area; Step 5: The initial sparse feature map obtained in Step 1 is dimensionality reduced based on the lightweight representation algorithm to generate a lightweight representation state vector; the lightweight representation state vector is input into a multi-label scene classifier for scene recognition, and the global baseline parameters in the algorithm are dynamically corrected according to the output scene recognition results to realize the single-source hierarchical adjustment of the global scene control layer and the local dynamic solution layer.

2. The contact wire target recognition algorithm based on lidar according to claim 1, characterized in that: In step one, the adaptive continuously variable resolution voxel filtering model defines the three-dimensional side length vector of the dynamic adaptive voxel grid; wherein, the voxel step size along the railway line extension direction is the maximum value of the ratio of the instantaneous travel speed of the inspection vehicle to the fixed scanning frequency of the lidar, and the product of the maximum ranging distance from the current radar to the target area of ​​the contact network and the fixed angular resolution of the lidar; the voxel step size perpendicular to the line direction and the height direction are equal, and both are taken as the algebraic product of the radial distance of the current scanned transient target point, the fixed angular resolution of the lidar, and the preset point cloud density adjustment factor.

3. The contact wire target recognition algorithm based on lidar according to claim 2, characterized in that: In step one, the implementation process of the multi-scale sparse feature aggregation method is as follows: a three-dimensional sparse tensor is constructed based on the voxel grid after adaptive voxel filtering, and multiple local sparse receptive fields with different expansion rates are set; within the scale space of each receptive field, sparse three-dimensional convolution operation is performed only on non-empty voxels, and the local curvature and normal vector gradient features extracted at each scale are concatenated along the channel dimension to complete the multi-scale sparse feature aggregation.

4. The contact wire target recognition algorithm based on lidar according to claim 3, characterized in that: In step two, the algebraic solution control rule for the time-varying cross-sectional dynamic envelope boundary is as follows: The upper and lower bounds of the time-varying vertical height of the region of interest are determined by adding and subtracting the vertical correction increment at the current moment from the highest and lowest installation heights of the overhead contact line standard design. The vertical correction increment is obtained by summing the instantaneous vertical vibration displacement generated by the integration of the vehicle-mounted inertial measurement unit and the redundant term of the radar ranging constant. The time-varying lateral left and right boundaries of the region of interest are determined by algebraically superimposing the lateral vibration displacement derived from the inertial measurement unit based on the lateral reference half-width of the track centerline, and further superimposing the roll compensation term determined by multiplying the sine of the vehicle roll angle by the highest installation height of the overhead contact line.

5. The contact wire target recognition algorithm based on lidar according to claim 4, characterized in that: In step three, the objective function of the trajectory deformation fusion optimization model integrates the physical curvature penalty term based on elastic mechanical boundary constraints and the local error penalty term based on the robust loss piecewise function; The adaptive morphology fitting algorithm also includes a soft threshold independent dynamic adjustment module. This module calculates the statistical variance of the point cloud height residual within the current moving fitting window, combines it with the discrete continuity index along the railway line extension direction, and independently scales and locks the implicit soft threshold in the robust loss piecewise function through piecewise linear mapping rules.

6. The contact wire target recognition algorithm based on lidar according to claim 5, characterized in that: In step four, the process of extracting dropper point cloud data includes: within the dropper search bounding box constructed by the contact wire and the catenary cable, hierarchical clustering is implemented using an adaptive clustering radius mapping model with local density inverse characteristics of the point cloud. The adaptive clustering radius of each sampling point is dynamically calculated based on the exponential power of the ratio factor between the global baseline clustering radius and the estimated maximum value of the local point cloud Gaussian kernel density. Subsequently, the generated clusters are verified by dual algebraic criteria. The dual algebraic criteria include a one-dimensional linear manifold verification based on the ratio of eigenvalues ​​of the cluster space covariance matrix, and an angular boundary verification based on the dot product of the first-order principal direction eigenvector and the vertical unit vector.

7. The contact wire target recognition algorithm based on lidar according to claim 6, characterized in that: In step four, the process of extracting the locator point cloud data includes: extracting a set of oblique candidate point clouds based on the spatial angle between the locator and the contact line and the coplanarity of the normal vectors, and constructing a rigid body judgment model composed of spatial compactness features and normalized orthogonal projection Shannon information entropy for integrity verification; wherein, the spatial compactness features are obtained by solving the ratio of the maximum eigenvalue to the minimum eigenvalue of the two-dimensional centered deviation covariance matrix constructed on the orthogonal projection plane of the candidate point clouds; the normalized orthogonal projection Shannon information entropy is obtained by performing a negative summation operation on the normalized distribution probability of all non-space grids and its base-2 logarithmic operator in the full grid domain after the projection plane is gridded.

8. The contact wire target recognition algorithm based on lidar according to claim 7, characterized in that: In step five, the lightweight representation algorithm processes as follows: it calculates the feature tensor covariance matrix of the multi-scale sparse feature map, removes redundant channels whose contribution weights to scene classification are lower than a preset minimum threshold using an adaptive channel pruning algorithm, and projects the retained high-dimensional feature vectors onto a low-rank manifold space using a nonlinear mapping function to generate a low-dimensional lightweight representation state vector; the multi-label scene classifier outputs a multi-label state vector representing one or more coexisting states among straight line segments, curved segments, tunnel scenes, and crossing scenes based on this lightweight representation state vector.

9. The contact wire target recognition algorithm based on lidar according to claim 8, characterized in that: In step five, the implementation process of the single-source hierarchical adjustment is as follows: based on the activation flags of each scene in the output multi-label state vector, a diagonal parameter state transition correction matrix is ​​constructed by analytical mapping calculation; the parameter state transition correction matrix is ​​multiplied by the global scene control vector containing multi-level fitting hard thresholds and benchmark clustering radii, so as to dynamically and adaptively scale the global benchmark parameters without interfering with the soft threshold calculation of local trajectory deformation.

10. The contact wire target recognition algorithm based on lidar according to claim 1, characterized in that: The algorithm also includes step six for feedforward control of visible light cameras: unifying the camera extrinsic perspective geometry equations with the one-dimensional vehicle kinematic recursive equations to solve the longitudinal span of the trigger mileage constrained by the image center coordinates. The system calculates the instantaneous mileage residual between the current mileage of the inspection vehicle and the mileage position of the target component in real time. When the instantaneous mileage residual reaches the dynamic trigger threshold, a hard trigger level signal is generated. The dynamic trigger threshold compensates for the instantaneous driving speed, instantaneous one-dimensional acceleration of the inspection vehicle, and the comprehensive exposure delay time of the system's hardware and software.

Images