A method for catenary contact line segmentation based on 3D point cloud data

Abstract

The present invention relates to a contact line segmentation method for a contact network based on three-dimensional point cloud data, and belongs to the technical field of point cloud segmentation. In order to overcome the defects existing in the prior art, the present invention aims to provide a contact line segmentation method for a contact network based on three-dimensional point cloud data, including: collecting three-dimensional point cloud data of the contact network by laser radar; performing point cloud preprocessing on the three-dimensional point cloud data of the contact network; since there are linear structures such as droppers and load-bearing cables near the contact line, the principal component analysis method is used to calculate the eigenvalues ​​to distinguish these interfering components; using the linear characteristics of the contact line, the contact line three-dimensional point cloud data after PCA processing is subjected to contact line detection and data extraction by the RANSAC straight line detection algorithm based on the three-dimensional point cloud, so as to obtain the three-dimensional point cloud data of the contact line. By using PCA, the present invention can reduce the influence of droppers and load-bearing cables on the contact line, so as to focus more on the characteristics of the contact line.

Description

Technical Field

[0001] The invention relates to a contact line segmentation method for a contact network based on three-dimensional point cloud data, and belongs to the technical field of point cloud segmentation. Background Art

[0002] As one of the main modes of modern transportation, railway transportation has the advantages of high speed, high safety, good comfort and high cost performance. Intercity railway transportation is becoming increasingly important, and the contact network is an important system for electric locomotives to obtain electricity. As the core part of electrified railways, the contact network plays a key role in ensuring safety, improving efficiency and reducing energy consumption. However, due to bad weather and train operation, the contact network may have problems such as loose structure, aging and broken wires, which may affect the operation of the entire railway line. In order to improve the safety and reliability of the traction power supply system, it is necessary to regularly detect the status of the contact network and timely understand the line geometry to avoid failures during operation.

[0003] Applying LiDAR technology to contact network inspection can improve the efficiency, accuracy and safety of detection. However, due to the wide distribution range of the contact network and the large amount of point cloud data collected by LiDAR, it is very important for contact network detection to efficiently and accurately segment the point cloud data to identify the contact network components of the point cloud.

[0004] At present, 3D point cloud segmentation methods can be divided into two categories: statistical analysis-based and deep learning-based. The statistical analysis-based method uses the feature differences such as distance, shape, echo intensity, etc. between different parts of the point cloud data to achieve the recognition and segmentation of specific targets or planes through fitting and analysis. This method usually uses machine learning algorithms such as random sample consensus (RANSAC), clustering or region growing, but has the following disadvantages: it is sensitive to parameter selection, and the number of clusters or other parameters need to be determined in advance. It is sensitive to the selection of the initial seed point, which may affect the final segmentation result.

[0005] The semantic segmentation method based on deep learning does not rely on manual extraction of target features, but automatically learns to extract multi-dimensional features through convolutional neural networks, and uses activation functions to perform semantic segmentation on high-dimensional features. Although this method can fully understand and identify targets in point clouds and performs well in dealing with complex scenes, it has the following disadvantages: deep learning models require a lot of computing resources during training and reasoning, especially for processing large-scale three-dimensional point cloud data, which requires more powerful computing power. Due to the diversity and complexity of point cloud data, the model is prone to overfitting the training data, especially when the amount of training data is insufficient or the labels are inaccurate. Summary of the invention

[0006] In order to overcome the defects in the prior art, the present invention aims to provide a contact line segmentation method for a contact network based on three-dimensional point cloud data.

[0007] The present invention provides a technical solution to solve the above technical problems: a method for segmenting a contact line of a contact network based on three-dimensional point cloud data, comprising the following steps:

[0008] Step S10, collecting contact network three-dimensional point cloud data through laser radar;

[0009] Step S20, performing point cloud preprocessing on the contact network three-dimensional point cloud data;

[0010] Step S30, using principal component analysis to filter out the contact network three-dimensional point cloud data;

[0011] Step S40: using the linear features of the contact line to perform contact line detection and data extraction on the filtered contact network three-dimensional point cloud data through a RANSAC line detection algorithm based on three-dimensional point cloud, to obtain contact line three-dimensional point cloud data.

[0012] A further technical solution is that the laser radar adopts Livox HAP.

[0013] A further technical solution is that in step 20, noise points and some background points in the contact network point cloud data are removed by point cloud preprocessing, including data denoising, filtering and discarding abnormal points; a straight-through filter is used to filter out point clouds outside the contact line range on the z-axis distance and retain point clouds within the contact line range. Since the dropper part is vertically connected to the contact line, the filter's filtering of the dropper part is not ideal and some parts will remain; for my country's electrified railways, the distance from the contact line to the track is generally 5.3 to 6.5 meters, and based on this data, the point cloud within this range of the z-axis is retained; the processing of the dropper part point cloud is completed in subsequent steps.

[0014] A further technical solution is that the filtering is one of statistical filtering, through filtering and conditional filtering.

[0015] A further technical solution is that the specific process of step S30 is: randomly select a point from the contact network three-dimensional point cloud data as the starting point, search for point clouds within a radius of 0.2m from the point, and if there is a point cloud within the range, use PCA to calculate the eigenvalue and linearity of the point. If the linearity is greater than 1.0, extract it, otherwise discard it, and repeat this step until all points are searched.

[0016] A further technical solution is that the calculation formula of the characteristic value and linearity is:

[0017]

[0018] Cv j =λ j v j

[0019]

[0020] l=(λ1-λ2) / λ1

[0021] Where: j is the eigenvalue; v j is the eigenvector; C is the covariance matrix; k is the number of domain points; is the three-dimensional centroid of the neighborhood point set; l is the linearity.

[0022] A further technical solution is that the specific process of step S40 is:

[0023] Step S41, determining the number of iterations and the distance threshold according to the continuous line features of the contact line;

[0024] Step S42, randomly selecting two points in the filtered contact network three-dimensional point cloud data to establish a spatial straight line model;

[0025] Step S43, calculating the distance from each remaining sampling point to the spatial straight line;

[0026] Step S44, judging whether each remaining sampling point is an inlier according to the distance threshold, and determining the number of inliers;

[0027] Step S45, repeat steps S42-S44 until the number of iterations is completed;

[0028] Step S46: After the iteration is completed, the set with the largest number of internal points is selected as the contact line three-dimensional point cloud.

[0029] A further technical solution is that the calculation formula for the number of iterations is:

[0030]

[0031] Where: t is the probability of an interior point in the data set; m is the number of iterations; N is the minimum amount of data required to calculate the model parameters; and P is the confidence level.

[0032] The present invention has the following beneficial effects:

[0033] 1. LiDAR measures the scene non-contactly by continuously emitting laser beams to obtain the distance information of the target points in the scene. It has the characteristics of fast speed, high accuracy and high resolution. It has been widely used in the fields of autonomous driving, robot navigation, environmental monitoring, etc.

[0034] 2. PCA can help convert high-dimensional contact line feature data into lower-dimensional data while retaining most of the information. At the same time, there are many other contact network components around the contact line, which may interfere with the analysis and identification of contact line features. By using PCA, the influence of these components can be reduced, so that more focus can be placed on the characteristics of the contact line. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is the overall flow chart of the present invention;

[0036] Figure 2 It is a flow chart of the RANSAC algorithm of the present invention. DETAILED DESCRIPTION

[0037] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] like Figure 1 As shown, a contact line segmentation method for a contact network based on three-dimensional point cloud data specifically includes the following steps:

[0039] Step S10, collecting contact network three-dimensional point cloud data through laser radar;

[0040] Specifically, the laser radar is installed at the front of the vehicle. The laser radar uses Livox HAP. Considering that the vertical angle of the selected laser radar is 25° and the horizontal angle is 120°, the laser radar is installed 8m away from the contact network support device and 3.8m away from the roof to ensure good contact line scanning effect.

[0041] Step S20, performing point cloud preprocessing on the contact network three-dimensional point cloud data to remove noise points and some background points in the contact network point cloud data, including data denoising, filtering (using one of statistical filters, pass filters, and conditional filters for filtering) and discarding abnormal points;

[0042] Step S30, using principal component analysis to filter out the contact network three-dimensional point cloud data;

[0043] Principal component analysis is an algorithm commonly used in the field of data analysis. It can use the linear combination of a few indicators to express most of the information in the data.

[0044] For any point p in the contact network point cloud, we can query the points near the domain. Taking the k domain as an example, assuming M = P i (x i ,y i,z i )| i＝1,2,…,k is a set of k points in the domain of point p, and the covariance matrix C is constructed through these k domain points. j | j＝1,2,3 is the eigenvalue of matrix C, which represents the degree of change of point p in the three main directions. Since the covariance matrix is ​​a semi-positive symmetric matrix, it can be known that the eigenvalue λ j Both are greater than 0.

[0045] In the following formula, v j is the corresponding eigenvector, is the three-dimensional centroid of the neighborhood point set.

[0046]

[0047] Cv j =λ j v j (2)

[0048]

[0049] l=(λ1-λ2) / λ1 (4)

[0050] Where: j is the eigenvalue; v j is the eigenvector; C is the covariance matrix; k is the number of domain points; is the three-dimensional centroid of the neighborhood point set; l is the linearity;

[0051] According to the size of the eigenvalue (λ1>λ2>λ3), λ j When the three eigenvalues ​​are approximately equal, the current point set can be considered to have spherical characteristics; when two of the three eigenvalues ​​are approximately equal and much larger than the third eigenvalue, the current point set can be considered to have surface characteristics; when only one of the three eigenvalues ​​is larger and the other two are smaller and approximately equal, the current point set can be considered to have linear characteristics.

[0052] Since the contact line and the catenary cables are all straight components in the contact network, the principal component analysis method is used to calculate the point set in the neighborhood of the contact line of the contact network. It can be obtained that the eigenvalue of the covariance matrix C constructed by the k-domain point set is only one large, and the other two eigenvalues ​​are small and approximately equal, so the current point set has a linear characteristic. The eigenvector of the dropper part is always perpendicular to the eigenvector of the catenary cable and the contact line. Even if there is a certain error, the inner product of the two vectors is approximately 0. Since the spatial coordinate system of the point cloud data is based on the laser radar as the origin, the coordinate value of the z direction of the eigenvector of the catenary cable calculated in the scanned point cloud data must be greater than the contact line, so the point cloud data of the contact line is preliminarily obtained;

[0053] Specifically, a point is randomly selected from the contact network 3D point cloud data as the starting point, and the point cloud within the radius of 0.2m of the point is searched. If there is a point cloud within the range, PCA is used to calculate the eigenvalue and linearity of the point. If the linearity is greater than 1.0, it is extracted, otherwise it is discarded, and this step is repeated until all points are searched.

[0054] Step S40, using the linear features of the contact line to perform contact line detection and data extraction on the filtered contact network three-dimensional point cloud data through a RANSAC line detection algorithm based on three-dimensional point cloud, to obtain contact line three-dimensional point cloud data;

[0055] Step S41, determining the number of iterations and the distance threshold according to the continuous line features of the contact line;

[0056] The objective function of the RANSAC algorithm is selected according to the data structure characteristics of the contact network point cloud: that is, through m iterations, the process of maximizing the number of internal points n under model parameters in different iteration processes.

[0057] The objective function of the RANSAC algorithm applied to spatial straight line fitting is to find the parameter model containing the most inliers based on the distance threshold K of the known spatial straight line;

[0058] The number of RANSAC iteration terminations m can be obtained through theoretical calculation.

[0059] Under the condition of confidence level P, during the loop, there is at least one sampling, so that the N points of the sampling subset are all internal points, thus ensuring that during the iteration, at least one sampling can obtain the maximum value of the objective function. The RANSAC algorithm requires that under a certain confidence probability, the minimum number of samples N of its basic subset and the probability P (P>t) of obtaining at least one benign sampling subset satisfy the relationship of formula (5).

[0060] P=1-(1-t N ) m (5)

[0061] Where: t is the probability of an inlier in the data set, that is, t = number of inliers / (number of inliers + number of outliers); N is the minimum amount of data required to calculate the model parameters.

[0062] Therefore, from formula (5), it can be obtained that the loop termination condition m should satisfy condition (6):

[0063]

[0064] The confidence level is generally set to the range of [0.95, 0.99]. In general, t is the probability of an inlier in the data set and is an unknown parameter. Therefore, t can be taken as the proportion of inliers under the worst conditions, and the current maximum proportion of inliers is continuously updated as the number of iterations increases. The confidence level is generally set to 0.99.

[0065] Step S42, randomly selecting two points in the filtered contact network three-dimensional point cloud data to establish a spatial straight line model;

[0066] There are three ways to express the equation of a line: general form, point form, and parametric form. PCL uniformly uses the point form. The point form equation of a line is:

[0067]

[0068] Among them, the direction vector s'=(m,n,p) and the known point (x0,y0,z0) are used to determine the line model.

[0069] Step S43, calculating the distance from each remaining sampling point to the spatial straight line;

[0070] Step S44, judging whether each remaining sampling point is an inlier according to the distance threshold, and determining the number of inliers;

[0071] Step S45, repeat steps S42-S44 until the number of iterations is completed;

[0072] Step S46: After the iteration is completed, the set with the largest number of internal points is selected as the contact line three-dimensional point cloud.

[0073] The above description is not intended to limit the present invention in any form. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are within the scope of the technical solution of the present invention.

Claims

1. A contact line segmentation method for a contact network based on three-dimensional point cloud data, characterized in that: The following steps are involved: Step S10, collecting contact network three-dimensional point cloud data through laser radar; Step S20, performing point cloud preprocessing on the contact network three-dimensional point cloud data; Step S30, using principal component analysis to filter out the contact network three-dimensional point cloud data; A point is randomly selected from the contact network 3D point cloud data as the starting point, and the point cloud within the radius of 0.2m of the point is searched. If there is a point cloud within the range, PCA is used to calculate the eigenvalue and linearity of the point. If the linearity is greater than 1.0, it is extracted, otherwise it is discarded. This step is repeated until all points are searched. The calculation formula of the characteristic value and linearity is: Cv j =λ j v j l=(λ1-λ2) / λ1 Where: j is the eigenvalue; v j is the eigenvector; C is the covariance matrix; k is the number of neighborhood points; is the three-dimensional centroid of the neighborhood point set; l is the linearity; Step S40, using the linear features of the contact line to perform contact line detection and data extraction on the filtered contact network three-dimensional point cloud data through a RANSAC line detection algorithm based on three-dimensional point cloud, to obtain contact line three-dimensional point cloud data; Step S41, determining the number of iterations and the distance threshold according to the continuous line features of the contact line; Step S42, randomly selecting two points in the filtered contact network three-dimensional point cloud data to establish a spatial straight line model; Step S43, calculating the distance from each remaining sampling point to the spatial straight line; Step S44, judging whether each remaining sampling point is an inlier according to the distance threshold, and determining the number of inliers; Step S45, repeat steps S42-S44 until the number of iterations is completed; Step S46: After the iteration is completed, the set with the largest number of internal points is selected as the contact line three-dimensional point cloud.

2. The contact line segmentation method of a contact network based on three-dimensional point cloud data according to claim 1, characterized in that: The laser radar adopts Livox HAP.

3. The contact line segmentation method of a contact network based on three-dimensional point cloud data according to claim 1, characterized in that: In the step S20, noise points and some background points in the contact network point cloud data are removed by point cloud preprocessing, including data denoising, filtering and discarding abnormal points; a straight-through filter is used to filter out point clouds outside the contact line range on the z-axis distance and retain point clouds within the contact line range. Since the dropper part is vertically connected to the contact line, the filter is not ideal for filtering the dropper part and some parts will remain; for my country's electrified railways, the distance from the contact line to the track is 5.3 to 6.5 meters, and the point cloud within this range of the z-axis is reserved according to this data; the processing of the dropper part point cloud is completed in subsequent steps.

4. The contact line segmentation method of a contact network based on three-dimensional point cloud data according to claim 3, characterized in that: The filtering is one of statistical filtering, through filtering and conditional filtering.

5. The contact line segmentation method of a contact network based on three-dimensional point cloud data according to claim 1, characterized in that: The calculation formula for the number of iterations is: Where: t is the probability of an interior point in the data set; m is the number of iterations; N is the minimum amount of data required to calculate the model parameters; and P is the confidence level.

Images