Method for extracting and reconstructing railway catenary

Abstract

The present disclosure provides a railway catenary extraction and reconstruction method, comprising: obtaining the geometric structure information corresponding to the railway catenary, tower and ground point respectively to obtain initial point cloud data; inputting the initial point cloud data into a double-branch deep learning network model to extract the local feature and global feature of each point in the initial point cloud data and the neighborhood feature between the points; splicing the local feature, global feature and neighborhood feature corresponding to each point respectively to obtain the high-dimensional feature of the point; identifying the point cloud data corresponding to the catenary based on the high-dimensional feature, and reconstructing the catenary based on the random sample consensus model according to the point cloud data corresponding to the catenary. The method can efficiently and accurately realize the extraction and reconstruction of the railway catenary.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of data processing, and in particular, to a railway catenary extraction and reconstruction method. BACKGROUND

[0002] LiDAR (Light Detection and Ranging) integrates a laser scanner (LS), a global navigation satellite system (GNSS), and an inertial measurement unit (IMU), and can sample the target of interest in the form of three-dimensional point cloud data. In particular, an airborne LiDAR performs by installing a LiDAR system on an airborne platform, and combines the distance value of the laser pulse illumination point with the position and orientation obtained by GNSS / IMU, which can effectively collect accurate and general large-scale scene three-dimensional information. Airborne LiDAR has gradually become a popular method in various fields due to its fast data acquisition capability. The point cloud data collected by airborne LiDAR has been applied to geological survey, terrain extraction and modeling, three-dimensional building reconstruction, forestry survey, precision agriculture, power line inspection, etc. Since the airborne laser scanning system can accurately and efficiently obtain three-dimensional shape information on a large scale, airborne laser point cloud has become the preferred data for many related tasks such as power line classification, extraction and reconstruction.

[0003] The railway catenary is composed of a contact net and a contact line, and is one of the important components of the railway system. To ensure stable and safe railway transportation services, it is essential to regularly monitor and inspect the railway catenary. The railway catenary point cloud obtained by using airborne laser scanning can provide sufficient and effective three-dimensional information for the extraction of the catenary. However, due to the complex and delicate structure of the railway catenary, there are few studies on the extraction of airborne laser railway catenary point cloud data at present. In recent years, scholars at home and abroad have carried out in-depth research on the extraction of high-voltage power lines, and many existing methods can accurately extract high-voltage power lines. Common methods include model fitting-based methods, clustering-based methods, and learning-based methods. However, the model fitting-based method needs to select a suitable model in advance, and the extraction accuracy is easily affected by the data quality, the clustering-based method is sensitive to the setting of the threshold and is limited in use, and the learning-based method has incomplete results for tower extraction. In general, these methods are for the extraction and reconstruction of high-voltage power lines with clear structure and parallel arrangement of transmission lines, while the railway catenary is more complex in structure than the high-voltage line, and the above extraction methods for high-voltage lines cannot obtain complete extraction results for the railway catenary. SUMMARY

[0004] To solve the above technical problems, the present disclosure provides a railway catenary extraction and reconstruction method, which at least partially solves the technical problem that the existing extraction and reconstruction method cannot obtain complete extraction results for the railway catenary.

[0005] The present disclosure provides a railway catenary extraction and reconstruction method, comprising: obtaining geometric structure information corresponding to a railway catenary, a tower and a ground point respectively to obtain initial point cloud data; inputting the initial point cloud data into a double-branch deep learning network model to extract local features and global features of each point in the initial point cloud data and neighborhood features between the points; splicing the local features, the global features and the neighborhood features corresponding to each point respectively to obtain high-dimensional features of the point; identifying point cloud data corresponding to the catenary based on the high-dimensional features, and reconstructing the catenary based on the point cloud data corresponding to the catenary according to a random sample consensus model.

[0006] According to an embodiment of the present disclosure, the double-branch deep learning network model comprises a geometric feature extraction module, a PointNet module and a neighborhood information aggregation module, wherein the geometric feature extraction module and the PointNet module are connected to constitute one branch, and the neighborhood information aggregation module constitutes another branch; the point cloud data is input into the double-branch deep learning network model to extract local features and global features of each point in the point cloud data and neighborhood features between the points, specifically including: extracting local geometric features of each point in the initial point cloud data by using the geometric feature extraction module; inputting and transforming the local geometric features of each point by using the PointNet module to obtain the local features of the point.

[0007] According to an embodiment of the present disclosure, the point cloud data is input into the double-branch deep learning network model to extract global features of each point in the point cloud data, specifically including: performing multi-layer perception and maximum pooling on the local features of each point by using the PointNet module to obtain the global features of the point.

[0008] According to an embodiment of the present disclosure, the point cloud data is input into the double-branch deep learning network model to extract neighborhood features between the points in the point cloud data, specifically including: for each point in the point cloud data, taking the point as a center point, aggregating geometric shallow features of the center point with geometric shallow features of other points in the neighborhood by using the neighborhood information aggregation module to obtain neighborhood features between the center point and the other points in the neighborhood.

[0009] According to an embodiment of the present disclosure, the geometric feature extraction module is used to extract local geometric features of each point in the initial point cloud data according to geometric feature descriptors, specifically including: for each point in the point cloud data, taking the point as a center point, determining all points in the neighborhood of the center point; constructing a three-dimensional structure tensor of the points in the neighborhood according to coordinate information of the center point and all points in the neighborhood; calculating three eigenvalues of the three-dimensional structure tensor; determining the local geometric features of the center point according to the three eigenvalues.

[0010] According to an embodiment of the present disclosure, the local geometric feature of the center point is determined according to three characteristic values, specifically including: sorting the three characteristic values from large to small; when the maximum characteristic value is much larger than the intermediate characteristic value, determining that the local geometric feature of the center point is linear; when the maximum characteristic value is approximately equal to the intermediate characteristic value and the intermediate characteristic value is much larger than the minimum characteristic value, determining that the local geometric feature of the center point is planar; and when the maximum characteristic value, the intermediate characteristic value and the minimum characteristic value are approximately equal, determining that the local geometric feature of the center point is a point.

[0011] According to an embodiment of the present disclosure, the P o The multi-layer perception of the local feature of each point by the intNet module includes: multi-scale feature connection is performed on the output of each layer in the multi-layer perception.

[0012] According to an embodiment of the present disclosure, the method for extracting and reconstructing the railway catenary further includes: training a double-branch deep learning network model; wherein the training of the double-branch deep learning network model includes: obtaining a training data set; performing data enhancement on the training data set, and down-sampling ground point data in the training data set to obtain an optimized training data set; and training the double-branch deep learning network model using the optimized training data set.

[0013] According to an embodiment of the present disclosure, the data enhancement on the training data set specifically includes: dividing the training data set into multiple blocks of data; segmenting each block of data into catenary data, tower data and ground point data, and adding different labels to each type of data; and performing rotation augmentation on each block of data to obtain the optimized training data set.

[0014] According to an embodiment of the present disclosure, the down-sampling of the ground point data in the training data set specifically includes: when the ground point data is obtained by segmenting each block of data, down-sampling the ground point data so that the number of ground point data is comparable to the number of catenary data; wherein the down-sampling of the ground point data specifically includes: octree space division and coding of the ground point data; calculating the center of gravity of each voxel in the octree, and compressing the ground point data according to the center of gravity of each voxel.

[0015] The method for extracting and reconstructing the railway catenary provided by the embodiment of the present disclosure has at least the following beneficial effects:

[0016] By using the double-branch deep learning network model, the local geometric feature of the point cloud data can be extracted, thereby enriching the local geometric information, the neighborhood feature between the points in the point cloud data can be extracted, the context relationship between the center point and the neighbor point is utilized, the description of the local feature of the network is enhanced, and high-precision extraction of the catenary point cloud data can be realized.

[0017] Further, in the process of aggregating local and global features, by performing multi-scale feature connection on the output of each layer of the multi-layer perception, high-level features containing various semantic information and low-level features containing more geometric details can be obtained, further improving the accuracy of feature extraction, and thus improving the extraction accuracy of the catenary point cloud data.

[0018] Further, in the process of training the dual-branch deep learning network model, data augmentation is performed on the training data and down-sampling is performed on the ground point data, ensuring the consistency of the training data and avoiding the under-segmentation or over-segmentation problem of the deep learning grid, thus ensuring the accuracy of the dual-branch deep learning network model, and further improving the extraction accuracy of the catenary point cloud data.

[0019] In addition, the high-dimensional features of each point are fitted by a random sample consensus model to identify the point cloud data corresponding to the catenary, improving the efficiency of data processing. BRIEF DESCRIPTION OF DRAWINGS

[0020] The above and other objects, features and advantages of the present disclosure will become more apparent from the following description of embodiments of the present disclosure, taken in conjunction with the accompanying drawings, in which:

[0021] Figure 1 A flowchart of a method for extracting and reconstructing a railway catenary provided by an embodiment of the present disclosure is schematically shown.

[0022] Figure 2 A structure diagram of a dual-branch deep learning network model provided by an embodiment of the present disclosure is schematically shown.

[0023] Figure 3 A structure diagram of a geometric feature extraction module provided by an embodiment of the present disclosure is schematically shown.

[0024] Figure 4 A structure diagram of a neighborhood information aggregation module provided by an embodiment of the present disclosure is schematically shown.

[0025] Figure 5 A structure diagram of a PointNet module provided by an embodiment of the present disclosure is schematically shown.

[0026] Figure 6 A structure diagram of an MLP provided by an embodiment of the present disclosure is schematically shown.

[0027] Figure 7 A catenary segmentation result diagram provided by an embodiment of the present disclosure is schematically shown.

[0028] Figure 8 A railway catenary structure diagram provided by an embodiment of the present disclosure is schematically shown.

[0029] Figure 9A raw point cloud span extraction result diagram provided by an embodiment of the present disclosure is schematically shown.

[0030] Figure 10 A railway catenary reconstruction result diagram provided by an embodiment of the present disclosure is schematically shown.

[0031] Figure 11 A data enhancement flowchart provided by an embodiment of the present disclosure is schematically shown.

[0032] Figure 12 An octree structure diagram provided by an embodiment of the present disclosure is schematically shown.

[0033] Figure 13 A preprocessed training data diagram provided by an embodiment of the present disclosure is schematically shown. DETAILED DESCRIPTION

[0034] For the purpose of clarity, technical solutions and advantages of the present disclosure, the present disclosure is further described in detail below in combination with specific embodiments and with reference to the drawings. Obviously, the described embodiments are part of the embodiments of the present disclosure, rather than all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present disclosure.

[0035] The terms used herein are merely used to describe specific embodiments, and are not intended to limit the present disclosure. The terms "include", "contain" and the like used herein indicate the existence of the described features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0036] In the present disclosure, unless otherwise explicitly specified and limited, the terms "mount", "connect", "connect", "fix" and other terms should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or it can be integrated; it can be mechanically connected, or it can be electrically connected or can communicate with each other; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two elements or the interaction relationship between two elements. For those skilled in the art, the specific meaning of the above terms in the present disclosure can be understood according to the specific circumstances.

[0037] In the description of the present disclosure, it should be understood that the orientation or positional relationship indicated by the terms "longitudinal", "length", "circumferential", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present disclosure and simplifying the description, and does not indicate or imply that the indicated subsystem or element must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present disclosure.

[0038] Throughout the drawings, like reference numerals will be used to refer to like or similar elements throughout the description of the disclosure. Conventional structures or configurations can be omitted when they can cause obscurity with regard to the understanding of the present disclosure. Also, the shapes, sizes, locations, and relative positions of the components in the drawings do not reflect actual ones and are merely used to facilitate the description. In addition, in the claims, any reference signs placed between parentheses are not to be construed as limiting the claim.

[0039] Similarly, in order to simplify the present disclosure and to help understand one or more of the various disclosed aspects, in the above description of the exemplary embodiments of the present disclosure, various features of the present disclosure are sometimes grouped together in a single embodiment, figure or description thereof. The description referring to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the particular feature, structure, material or characteristic described in connection with this embodiment or example is included in at least one embodiment or example of the present disclosure. The illustrative representations of the above terms in the specification do not necessarily refer to the same embodiment or example. Moreover, the particular features, structures, materials or characteristics described can be combined in any suitable manner in one or more embodiments or examples.

[0040] In addition, the terms "first", "second", etc. are used only for the purpose of description and should not be understood as indicating or implying relative importance or implying the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present disclosure, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise explicitly specified.

[0041] The present disclosure aims to solve the problems of low processing efficiency and poor extraction accuracy in airborne LiDAR railway catenary point cloud extraction. Based on in-depth analysis of the characteristics of airborne LiDAR railway line point cloud data, through preprocessing, catenary extraction based on deep learning, catenary identification and reconstruction based on random sample consensus algorithm and other methods, high-efficiency processing and high-precision extraction of airborne railway catenary point cloud are realized, which makes up for the lack of airborne railway catenary point cloud extraction and provides protection for railway transportation safety. The specific embodiments will be described in detail below.

[0042] Figure 1 The extraction and reconstruction method of the railway catenary provided by the embodiment of the present disclosure is schematically shown.

[0043] As shown in the figure, the extraction and reconstruction method of the railway catenary may, for example, include operations S101-S104. Figure 1

[0044] ​In operation S101, geometric structure information corresponding to the railway catenary, the tower and the ground point respectively is acquired to obtain initial point cloud data.

[0045] In operation S102, the initial point cloud data is input into a double-branch deep learning network model to extract local features and global features of each point in the initial point cloud data and neighborhood features between the points.

[0046] In operation S103, the local features, the global features and the neighborhood features corresponding to each point are spliced respectively to obtain high-dimensional features of the point.

[0047] In operation S104, point cloud data corresponding to the catenary is identified based on the high-dimensional features, and the catenary is reconstructed based on the random sample consensus model according to the point cloud data corresponding to the catenary.

[0048] Figure 2 A structure diagram of the double-branch deep learning network model provided by the embodiment of the present disclosure is schematically shown.

[0049] As shown in the figure, Figure 2 The double-branch deep learning network model includes a geometric feature extraction (GFE) module, a PointNet module and a neighborhood information aggregation (NIA) module, wherein the GFE module and the PointNet module are connected to constitute one branch, and the NIA module constitutes another branch. The GFE module aims to extract local geometric features of the point cloud, thereby enriching the local geometric information. However, local feature extraction still has challenges in extracting information in incomplete data. Therefore, it is proposed to use the NIA module to aggregate neighborhood information. The NIA module extracts the context relationship between the center point and the neighbor point, thereby enhancing the descriptive of the local features of the network. The PointNet module provides point cloud alignment and rotation invariant operation. The extracted local and global features are fed to a multi-layer perceptron (MLP) and a maximum pooling to obtain global features of the point. Finally, the features obtained by the three modules are concatenated to form high-dimensional features. Then, the scores of each class are output by the MLP.

[0050] In the embodiment of the present disclosure, the point cloud data is input into the double-branch deep learning network model to extract local features and global features of each point in the point cloud data and neighborhood features between the points, specifically including: extracting local geometric features of each point in the initial point cloud data by using the geometric feature extraction module; inputting and transforming the local geometric features of each point by using the PointNet module to obtain local features of the point.

[0051] Figure 3A structural diagram of a geometric feature extraction module type is schematically shown.

[0052] As shown in Figure 3 , the original point cloud data containing only xyz coordinate value information is input into the GFE module, n adjacent points within a radius R are queried using a radius search algorithm, and geometric features are calculated respectively. Then, the 3D coordinates xyz are input into the PointNet model to obtain the local features of each point. Each point is represented by a 6-dimensional (6D) vector, i.e., P(x, y, z, A, L, P).

[0053] Further, the geometric feature extraction module extracting the local geometric features of each point in the initial point cloud data using the radius search algorithm can specifically include: for each point in the point cloud data, determining all points in the neighborhood of the center point with the center point as the center. A three-dimensional structure tensor of the points in the neighborhood is constructed according to the coordinate information of the center point and all points in the neighborhood. Three eigenvalues of the three-dimensional structure tensor are calculated, and the local geometric features of the center point are determined according to the three eigenvalues.

[0054] Exemplarily, since the original point cloud data only contains xyz coordinate values, this may not be sufficient to distinguish catenary and tower, and within a certain neighborhood, catenary, tower and ground points are approximately linear, spherical and planar structures respectively. Therefore, three geometric feature descriptors: anisotropy, linearity and planarity can be selected to enrich the local information of the points. Assuming a 3D point cloud set P, each point in P has xyz coordinate information. Given a center point p∈P and a radius R, a set of neighborhood points {pi||p-pi|<R} is obtained. Then, the three-dimensional structure tensor M of the points in the neighborhood is defined as follows:

[0055]

[0056] where n is the number of neighborhood points. Q=(p1-p, p2-p,..., p n -p) T , p can represent the coordinates of each point in the point cloud data, M is a real symmetric matrix, which can be decomposed as M=RI T , R is a rotation matrix, and I is a positive definite matrix. The elements of I are the eigenvalues of the matrix M. The three eigenvalues are all positive, represented by λ1, λ2 and λ3 respectively, and sorted as λ1>λ2>λ3. Then, the 3D feature is defined as follows:

[0057]

[0058] After the three eigenvalues are calculated, determining the local geometric feature of the center point according to the three eigenvalues can specifically include: sorting the three eigenvalues from large to small. When the maximum eigenvalue is much larger than the intermediate eigenvalue, it is determined that the local geometric feature of the center point is linear. When the maximum eigenvalue is approximately equal to the intermediate eigenvalue and the intermediate eigenvalue is much larger than the minimum eigenvalue, it is determined that the local geometric feature of the center point is planar. When the maximum eigenvalue, the intermediate eigenvalue and the minimum eigenvalue are approximately equal, it is determined that the local geometric feature of the center point is a point.

[0059] Exemplarily, in the case of a linear structure, such as a catenary point, λ1>>λ2 is observed. For a planar structure, such as a ground point, λ1≈λ2>>λ3. For a scattered point, such as a tower point, λ1≈λ2≈λ3. Through these local geometric features, the prior knowledge of the model can be enhanced, and the ability of the model to identify different shaped objects can be improved.

[0060] The GFE module enriches the input information and provides local shape features, but it is difficult to completely retain the geometric information obtained after the MLP. Therefore, a short branch is designed, which aggregates the geometric shallow features through the MLP to describe the local details. In a certain neighborhood, some points do not have obvious geometric structures, especially at the junction. Therefore, the NIA module is proposed, which can extract the context information of the center point and the adjacent points, and enhance the local geometric description ability of the model.

[0061] Figure 4 A structural diagram of the neighborhood information aggregation module provided by the embodiment of the disclosure is schematically shown.

[0062] As shown in Figure 4 , the point cloud data is input into the double-branch deep learning network model, and the neighborhood features between the points in the point cloud data are extracted, which can specifically include: for each point in the point cloud data, taking the point as a center point, using the neighborhood information aggregation module to aggregate the geometric shallow features of the center point and the geometric shallow features of other points in the neighborhood, to obtain the neighborhood features between the center point and other points in the neighborhood.

[0063] Exemplarily, for a center point p g , first, the center point feature f is extracted by the MLP. Then, K neighboring points {p i |i<K} of the center point are found by the K-neighbor algorithm. The neighborhood features Fpi are obtained through the index of the points.

[0064] Then, the center point feature and the neighborhood feature are connected to obtain the semantic feature Fn between the points, so that the network can extract the local neighborhood information and the context between the center point and the neighborhood points. Wherein, Fn can be expressed as:

[0065]

[0066] where C denotes concatenation of data in dimensions, F n represents an N x (16 * K) dimensional feature.

[0067] Figure 5 The structure diagram of the PointNet module provided by the embodiments of the present disclosure is schematically shown.

[0068] As Figure 5 shown, the point cloud data is input into the double-branch deep learning network model, and the global feature of each point in the point cloud data is extracted, which can specifically include: using the PointNet module to perform multi-layer perception and maximum pooling on the local feature of each point to obtain the global feature of the point.

[0069] Exemplarily, each point in the input point cloud data is composed of 6D features calculated by the GEF module, that is, P(x, y, z, A, L, P). The input transformation and feature transformation of the input point cloud based on the original PointNet model are performed to obtain the local feature F l of each point. Then the point features are aggregated through the MLP and the maximum pooling operation to obtain the global feature Fg of the point:

[0070] F g = max(MLP((x, y, z, A, L, P)

[0071] where F g represents the global feature with a shape of N x 1024, and max(·) represents the maximum pooling operation. After maxpool, the repeat(·) function changes the global feature from 1 x 1024 to N x 1024 to achieve dimension alignment, so as to perform the following dimension concatenation operation.

[0072] Further, in order to improve the extraction accuracy, the structure of the MLP is also designed in the embodiments of the present disclosure.

[0073] Figure 6 The structure diagram of the MLP provided by the embodiments of the present disclosure is schematically shown.

[0074] As Figure 6 shown, the MLP is composed of an input layer, an output layer and one or more hidden layers. The input layer inputs the features, the neurons of each fully connected layer can fit the original features, the output layer outputs the features, and the output of each layer is connected with multi-scale features. In this way, the MLP can obtain high-level features containing various semantic information, and the hidden layer can obtain low-level features containing more geometric details.

[0075] After obtaining the local feature, global feature and neighborhood feature, the local and global features of the points can be connected with the neighborhood features of the NIA module to obtain the high-dimensional features of each point. Specifically as follows:

[0076] F c = C(F l , F g , F n )

[0077] where F c represents a feature with a shape of N x 2048. The method uses the softmax loss function of the PointNet model:

[0078] loss = -∑y i ln a i

[0079] where y i represents the true class, and a i represents the output value of softmax. Finally, after the MLP, the network outputs the score of each class.

[0080] Figure 7 A railway catenary segmentation result diagram provided by an embodiment of the disclosure is schematically shown.

[0081] As shown in Figure 7 , the point cloud data is divided into three types: catenary, tower and ground points.

[0082] After obtaining the high-dimensional features, the extracted catenary points need to be further processed to obtain individual conductors. Compared with high-voltage power lines, railway overhead lines are more complex in structure.

[0083] Figure 8 A railway catenary overhead line structure diagram provided by an embodiment of the disclosure is schematically shown.

[0084] As shown in a of Figure 8 , one side of the middle of the two support rods of the railway line respectively contains three catenary lines and one contact conductor. However, due to the low data density obtained by the airborne LiDAR, there are many catenary line breakage cases. In addition, the distance between catenary lines is very short. Therefore, it is difficult to identify a single conductor using the classic density clustering method. Based on the above problems, an automatic catenary line recognition and multi-line fitting algorithm is designed. The algorithm can stably identify each catenary line and reconstruct the catenary line by fitting a quadratic equation. Specifically, it includes three operations: original catenary line span extraction, catenary line recognition and catenary line reconstruction.

[0085] Original catenary line span extraction:

[0086] The catenary obtained by the foregoing segmentation is sampled data. To ensure the accuracy of catenary reconstruction, the sampled data is inversely mapped to obtain the catenary segmentation result of the original data before reconstruction.

[0087] Figure 9 An original point cloud span extraction result diagram provided by the embodiment of the disclosure is schematically shown.

[0088] As Figure 9 shown, the transmission line in the railway line corridor is composed of multiple segments. The tower connects these segments. The complete tower point has been extracted in the foregoing, and the boundary box is calculated by using the extracted tower point, and the catenary point in the middle of the boundary box is the part that needs to be reconstructed. Span extraction reduces the number of data points and makes the catenary reconstruction easier.

[0089] Catenary identification:

[0090] Continuing to refer to Figure 8 , as shown in a of Figure 8 , the contact conductor is a three-dimensional straight line, which is in the same vertical plane as the two catenaries above. That is, in the top view, the eight conductors in the middle of the two support rods are arranged in parallel into four straight lines, as shown in b of Figure 11 .

[0091] Aiming at the structural characteristics of the railway line, an identification method is proposed, which is as follows:

[0092] Let the span point set be {P}, the proposed method is to find the eigenvector corresponding to the maximum eigenvalue of the midpoint O of the boundary box The direction of the vector is the direction of the conductor. Threshold segmentation is performed on the distance from the point to the straight line passing through the midpoint, and the conductor points are divided into four clusters, which are {P1}, {P2}, {P3} and {P4}, as shown in b of Figure 10 . In particular, {P2}, {P3} contains three lines, which are two catenaries and one contact conductor. Since the contact conductor and the catenary use different reconstruction models, here, {P2} and {P3} are divided according to the elevation, and the contact conductor is extracted separately.

[0093] Catenary reconstruction:

[0094] After extracting and identifying the catenary, the last step is to reconstruct the catenary point for subsequent use (such as 3D measurement). The method proposed is a multi-line fitting based on the random sample consensus model. {P} is a point cluster segmented in the previous step, which is as follows:

[0095] First, {P} is translated and rotated to move it to the XOZ plane. Then, three seed points p1, p2, p3 are randomly selected from {P}. A quadratic model is used to perform the catenary fitting on the XOZ plane as follows.

[0096] z = Ax 2 +Bx+C

[0097] Where A, B, C represent the parameters of the fitted curve.

[0098] When the distance d of the remaining points to the curve is less than a given parameter σ, the point is classified as a point on the curve. After iterating t times, the model containing the most catenary points is taken as the optimal model. The corresponding optimal parameters best_A, best_B, best_C are output. Then, the second catenary fitting is performed in a loop until it stops. In this way, the multi-line fitting is completed. For the contact conductor, a straight line model is selected for conductor fitting on the XOZ plane.

[0099] z = ax + b

[0100] Where a, b represent the parameters of the fitted straight line.

[0101] Figure 10 A railway catenary reconstruction result diagram provided by the embodiment of the disclosure is schematically shown.

[0102] As Figure 10 shown, where a is a certain span catenary overhead view, b is a catenary overhead line identification result, c is an overhead view of b, and d is a reconstruction result.

[0103] On the basis of the above-mentioned embodiment, the railway catenary extraction and reconstruction method provided by the disclosure further includes training the double-branch deep learning network model. Training the double-branch deep learning network model can include: obtaining a training data set. The training data set is subjected to data enhancement, and the ground point data in the training data set is down-sampled to obtain an optimized training data set. The double-branch deep learning network model is trained using the optimized training data set.

[0104] Due to the mechanism of acquiring point cloud data by airborne LiDAR and the structural characteristics of railway catenary, the number of catenary points and tower points is less than that of ground points. The inconsistency of data may cause under-segmentation or over-segmentation problems of the deep learning grid. In order to solve this problem, a two-step strategy is applied: first, data enhancement is performed to ensure sufficient training data; second, the ground points are down-sampled to reduce the excessive number of ground points and improve the robustness of the model.

[0105] In this embodiment of the disclosure, data augmentation of the training dataset may specifically include three steps: dividing the training dataset into blocks to obtain multiple data blocks; dividing each data block into catenary data, tower data, and ground point data, and adding different labels to each type of data; and performing rotation augmentation on each data block to obtain an optimized training dataset.

[0106] Figure 11 A flowchart illustrating the data enhancement provided in an embodiment of this disclosure is shown schematically.

[0107] like Figure 11 As shown, 'a' is a side view of the railway line, which is distributed along a corridor. A tower is erected at regular intervals. This method divides the railway line data into blocks along the corridor according to a certain length, ensuring that each block contains two towers, with one tower overlapping between adjacent blocks. This strategy also ensures the consistency of the number of training points and the utilization rate of tower points. The rotation augmentation process is as follows: In addition to the original coordinates, each data block is rotated 90° around the X, Y, and Z axes, respectively, with a point within the point cloud as the origin. For example... Figure 11 As shown in Figure b, the original data is rotated around the X, Y, and Z axes respectively to obtain new point cloud coordinates, thereby obtaining new training samples and alleviating the problem of insufficient training samples for railway lines.

[0108] In this embodiment of the disclosure, downsampling the ground point data in the training dataset specifically includes: when segmenting each data block to obtain ground point data, downsampling the ground point data so that the number of ground point data is comparable to the number of catenary data. Specifically, downsampling the ground point data includes: partitioning and encoding the ground point data into an octree space; calculating the centroid of each voxel in the octree; and compressing the ground point data based on the centroid of each voxel.

[0109] Figure 12 An octree structure diagram provided in an embodiment of this disclosure is illustrated schematically.

[0110] For example, such as Figure 12 As shown, where a is the octree hierarchical structure, b is the octree spatial structure, and c is the centroid of the child nodes. As mentioned above, in airborne point cloud data of elevated railway scenarios, ground points typically account for more than 90% of the total number of points. Using uneven data to train the model leads to undesirable results. Therefore, while segmenting the training data, ground points are downsampled to make the number of ground points and elevated railway points roughly equal. This step includes: first, performing octree spatial partitioning and encoding on the point cloud, such as... Figure 12 As shown in a and b, a maximum recursion depth is set, the maximum size of the ground point cloud is found, and the first cube is built using this size. This is subdivided into eight equal parts until the maximum recursion depth is reached. Then, the centroid of each voxel in the octree is calculated to compress the point cloud, as shown in Figure a.Figure 12 The center c is shown in the middle.

[0111] For example, let P i = (x i , y i , z i ) T be the three-dimensional coordinates of a point P i in the voxel, then the center of gravity P i of all points P g in the voxel is:

[0112]

[0113] In the formula, n is the number of points in the voxel.

[0114] Figure 1 3 shows a pre-processed training data graph provided by the embodiment of the present disclosure.

[0115] In comparison Figure 13 and Figure 12 , the pre-processed data is more uniform.

[0116] So far, the data preprocessing process is completed.

[0117] To sum up, the railway catenary extraction and reconstruction method provided by the embodiment of the present disclosure can efficiently and accurately realize the extraction and reconstruction of the catenary.

[0118] The above specific embodiments further illustrate the purpose, technical solutions and advantages of the present disclosure. It should be understood that the above description is only a specific embodiment of the present disclosure and is not intended to limit the present disclosure. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present disclosure should be included in the protection scope of the present disclosure.

Claims

1. A method of extraction and reconstruction of a railway catenary, characterized in that, The method comprises the following steps: Obtain the geometric structure information corresponding to the railway catenary, the tower and the ground point respectively to obtain initial point cloud data; Input the initial point cloud data into a double-branch deep learning network model to extract the local feature, the global feature and the neighborhood feature of each point in the initial point cloud data; wherein the double-branch deep learning network model comprises a geometric feature extraction module, a PointNet module and a neighborhood information aggregation module, wherein the geometric feature extraction module and the PointNet module are connected to form one branch, and the neighborhood information aggregation module forms another branch; the extraction of the local feature, the global feature and the neighborhood feature of each point in the initial point cloud data comprises: extracting the local geometric feature of each point in the initial point cloud data by using the geometric feature extraction module; inputting and transforming the local geometric feature of each point by using the PointNet module to obtain the local feature of the point; Splice the local feature, the global feature and the neighborhood feature of each point respectively to obtain the high-dimensional feature of the point; Identify the point cloud data corresponding to the catenary based on the high-dimensional feature, and reconstruct the catenary based on the random sample consensus model according to the point cloud data corresponding to the catenary; The method for extracting and reconstructing the railway catenary further comprises training the double-branch deep learning network model; wherein the training of the double-branch deep learning network model comprises: obtaining a training data set; performing data enhancement on the training data set, and downsampling the ground point data in the training data set to obtain an optimized training data set; training the double-branch deep learning network model by using the optimized training data set.

2. The method of claim 1, wherein, Input the point cloud data into the double-branch deep learning network model to extract the global feature of each point in the point cloud data, specifically comprising: Input and transform the local feature of each point by using the PointNet module to obtain the global feature of the point.

3. The method of claim 1, wherein, Input the point cloud data into the double-branch deep learning network model to extract the neighborhood feature between points in the point cloud data, specifically comprising: For each point in the point cloud data, taking the point as a center point, the neighborhood information aggregation module is used to aggregate the geometric shallow features of the center point and the geometric shallow features of other points in the neighborhood to obtain the neighborhood feature between the center point and other points in the neighborhood.

4. The method of claim 1, wherein, The geometric feature extraction module is used to extract the local geometric feature of each point in the initial point cloud data, specifically comprising: For each point in the point cloud data, taking the point as a center point, all points in the neighborhood of the center point are determined; A three-dimensional structure tensor of the points in the neighborhood is constructed according to the coordinate information of the center point and all points in the neighborhood; Three eigenvalues of the three-dimensional structure tensor are calculated; The local geometric feature of the center point is determined according to the three eigenvalues.

5. The method of claim 2, wherein, The multi-layer perception of the PointNet module on the local feature of each point comprises: Multi-scale feature connection is performed on the output of each layer in the multi-layer perception.

6. The method of claim 1, wherein, The training data set is subjected to data enhancement, specifically comprising: The training data set is divided into multiple blocks to obtain multiple blocks of data; Each piece of data is divided into catenary data, tower data and ground point data, and different labels are added for each type of data; Rotational amplification is performed on each piece of data to obtain the optimized training data set.

7. The method of claim 6, wherein, The ground point data in the training data set is down-sampled, specifically including: When the ground point data is obtained by dividing each piece of data, the ground point data is down-sampled so that the number of ground point data is comparable to the number of catenary data; Wherein, the down-sampling of the ground point data specifically includes: The ground point data is octree spatially divided and encoded; The center of gravity of each voxel in the octree is calculated, and the ground point data is compressed according to the center of gravity of each voxel.

Images