Fully automatic search and extraction method for contact network facility point cloud

Abstract

The present invention discloses a method for fully automatic search and extraction of contact network facility point cloud, including: S1, collecting laser point cloud data along the target railway, and automatically searching and extracting the contact network point cloud in the laser point cloud data based on a double-selection stereo frame: determining the search range by a rough selection stereo frame, and extracting the track line points within the range according to the frame; tracking, cutting and extracting along the track direction by a selected stereo frame; S2, semantic segmentation of the cut and extracted contact network point cloud based on deep learning, including the following sub-steps: S2‑1, manually marking the point cloud data; S2‑2, constructing an MFF_A model; S3, reconstructing a contact network three-dimensional model based on geometric features; S4, detecting geometric parameters based on the contact network model. The method is fast, has high extraction accuracy, and is suitable for terrain undulations and curves; the accuracy and computational efficiency of the contact network semantic segmentation method are better than similar algorithms.

Description

Technical Field

[0001] The invention relates to the field of railway facility monitoring, and in particular to a fully automatic search and extraction method for a contact network facility point cloud. Background Art

[0002] Railway transportation has become a major mode of modern transportation due to its advantages of high speed, high safety, good comfort and high cost performance. At present, railway transportation has gradually become one of the most important means of transportation between cities. Electrified railways are the main form of existing railways. Overhead Catenary System (OCS) refers to the electromechanical system that provides electrical energy to the electric traction unit through the collector. Figure 1 As shown, it is usually composed of several parts such as contact suspension, supporting device, positioner, pillar and foundation. Among them, the contact suspension includes contact wire, suspension string, load-bearing cable and connecting parts. The contact suspension is erected on the column through the supporting device. Its function is to transmit the electric energy obtained from the traction substation to the electric locomotive. The supporting device is used to support the contact suspension and transmit its load to the column or other buildings. The supporting device usually includes a flat wrist arm, an inclined wrist arm, a horizontal pull rod and an insulator. The function of the positioner is to fix the position of the contact wire so that the contact wire is within the operating track of the pantograph slide, ensure that the contact wire and the pantograph are not separated, and transmit the horizontal load of the contact wire to the pillar.

[0003] As the core part of electrified railways, OCS plays an important role in ensuring railway transportation safety, improving transportation efficiency, and reducing transportation energy consumption. However, since the contact network is in an open-air environment and is affected by severe natural weather such as wind, rain, snow, and dust, it is inevitable that the structure will become loose, age, and change in geometric position. At the same time, during train operation, the contact network will be deformed or even broken due to irregular operation, abnormal relationship between the bow and the catenary, long-term structural vibration, excessive current, electrical corrosion, and other reasons. When the contact network fails, the operation of the entire railway line will be affected due to the lack of a backup line. In order to avoid failures of the contact network during operation and improve the safety and reliability of the traction power supply system, it is necessary to conduct regular inspections of the contact network and timely grasp the geometric status of the contact network of the entire line.

[0004] The Vehicle Mobile Measurement System (VMMS) collects high-density laser point cloud data along the line by mounting the measurement system on a track flatbed vehicle. Compared with traditional image-based and LiDAR-based detection methods, VMMS data acquires richer dimensional information and has the advantages of fast measurement speed, wide scanning range (suitable for double-track overhead contact network data collection) and high accuracy (linear error better than 5mm).

[0005] The key to VMMS detection is the accurate semantic segmentation of point cloud components. At present, three-dimensional point cloud segmentation methods can be divided into two categories: statistical analysis-based methods and deep learning-based methods. The principle of statistical analysis-based methods is to achieve fitting and analysis of specific targets or planes based on the feature differences such as distance, shape, echo intensity, etc. between different parts in point cloud data, and then achieve semantic segmentation of different components. There are mainly machine learning algorithms such as random sample consensus (RANSAC) clustering or region growing. However, such algorithms rely more on prior knowledge, require a lot of manual intervention in the implementation process, and need to set more parameters, making it difficult to achieve automatic extraction and wide application promotion. The semantic segmentation method based on deep learning has better application prospects. It no longer relies on manual extraction of target features, but automatically learns to extract multi-dimensional features through convolution, and uses activation functions to perform semantic segmentation on high-dimensional features. Although many scholars have conducted in-depth research on the segmentation of contact network facilities based on deep learning, the existing methods still have the following defects:

[0006] (1) It relies on a lot of manual preliminary sample extraction and processing, and lacks methods for automatically searching and extracting overhead line facilities;

[0007] (2) In deep neural networks, from the perspective of feature extraction, many networks map features to high-dimensional space to obtain more and richer high-dimensional features, but they cannot effectively distinguish the feature weight matrix according to the importance of features, resulting in a large amount of important feature information being discarded because it cannot be effectively transmitted. After the attention mechanism was introduced in a few network models, this situation has improved, and the segmentation accuracy has been slightly improved, but the amount of calculation has increased dramatically, and there is a lack of effective attention mechanism to improve the weight coefficients of important features;

[0008] (3) The refinement structure plays an increasingly important role in the image deep semantic segmentation network. Currently, the refinement structure is rarely applied to the preliminary segmentation results of point clouds, and the preliminary segmentation results cannot meet the segmentation accuracy required by the application.

[0009] (4) There is a lack of extraction and fusion of multi-scale features, and the ability to fuse shallow features with deep features is insufficient, resulting in serious loss of effective feature expression. Summary of the invention

[0010] In view of the deficiencies in the prior art in the search and extraction capabilities of contact network facility point clouds, the present invention provides a fully automatic search and extraction method for contact network facility point clouds based on double-selection stereo boxes and semantic segmentation.

[0011] To this end, the present invention adopts the following technical solutions:

[0012] A fully automatic search and extraction method for a contact network facility point cloud comprises the following steps:

[0013] S1, collect laser point cloud data along the target railway, and automatically search and extract the contact network point cloud in the laser point cloud data based on the double-selected stereo frame: first, determine the search range through the coarse selection stereo frame, and use this frame to extract the track line points within the range; then, track, crop and extract along the track direction through the selected stereo frame;

[0014] S2, semantic segmentation of the cropped and extracted contact network point cloud based on deep learning, including the following sub-steps:

[0015] S2-1, manually labeling the point cloud data;

[0016] S2-2, construct MFF_A model;

[0017] S3, reconstruction of the contact network 3D model based on geometric features;

[0018] S4, geometric parameter detection based on contact network model.

[0019] Among them, in step S1, the method for automatically searching and extracting the contact network point cloud in the laser point cloud data based on the double-selection stereo box includes the following sub-steps:

[0020] S1-1, rough selection of stereoscopic frame posture and positioning: the rough selection of stereoscopic frame satisfies the following formula:

[0021] BBox={min_point,max_point}

[0022] In the formula, min_point and max_point represent the minimum and maximum points of the coordinate values ​​within the range of the roughly selected stereo frame, respectively.

[0023] The center point of the rough selection stereo frame is located near the center points of two pairs of adjacent columns; the area of ​​the rough selection stereo frame includes the track line points and the point cloud area of ​​all facilities of the contact network; the six planes of the rough selection stereo frame are parallel to the XOY, XOZ and YOZ planes respectively; on the XOY projection plane, the center point of the rough selection stereo frame falls in the middle of the two pairs of adjacent columns and on the track line point; during the cutting and extraction process, the rough selection stereo frame moves along the track direction, and during the movement process, the rough selection stereo frame is moved to the next position according to the spacing between the columns;

[0024] S1-2, selected stereo frame posture and positioning: the selected stereo frame satisfies the following formula:

[0025] CBox=RMatrix*minBox+T

[0026] Where minBox is the minimum three-dimensional box for cutting a pair of overhead line column support facilities; RMatrix is ​​the rotation matrix; T is the translation vector, where:

[0027]

[0028] The selected stereo frame satisfies the requirement that the minimum cropped stereo frame is completely contained in the rough selected stereo frame, and the long side of the selected stereo frame is parallel to the track, and the width is perpendicular to the track; the unprocessed minimum stereo frame minBox is rotated and translated to obtain the final selected stereo frame CBox;

[0029] S1-3, determining the offset of the double-selection stereo box, includes the following steps:

[0030] S1-3-1, Determination of offset distance: First, extract the position of the center point of each column from the scene {Centerpoint1,...,Centerpoint i ,...,Centerpoint n}, by projecting these center points onto the XOY plane, the distances from the centers of the adjacent columns are calculated to obtain the offset distance information of the double-selection stereo box:

[0031] PoleDis[i]=Dis(Centerpoint[i+1]-centerpoint[i]),i=1,...,N-1

[0032] Where PoleDis[i] is the Euclidean distance between the i-th column and the i+1-th column in the XOY plane; Dis(·) is the function for calculating the Euclidean distance between two points in the XOY plane; Centerpoint[i+1] is the coordinate information of the center point of the i+1-th column;

[0033] S1-3-2, determination of the offset direction: obtain all POS track points in the rough selection stereo frame, and use two adjacent POS track points as the direction of the railway track. When the rough selection stereo frame moves and searches along the track, the offset direction formula of two adjacent rough selection stereo frames is:

[0034] V dif =V i+1 -V i

[0035] Where V i+1 is the POS trajectory direction vector in the i-th rough selected stereo frame, V i+1 is the POS trajectory direction vector in the i+1th rough selected stereo frame, V dif Represents the direction vector between two adjacent coarse selection stereo frames i and i+1, and obtains the offset vector T of the double selection stereo frame;

[0036] S1-4, trajectory data-assisted automatic adjustment of selected stereoscopic frame poses, includes the following steps:

[0037] S1-4-1, center adjustment of the selected stereoscopic frame: after determining the offset vector T of the double-selected stereoscopic frame, the selected stereoscopic frame is translated to the next area to be cropped. The specific steps are as follows:

[0038] (1) adjusting the double-selection stereo frame in translation according to the offset vector;

[0039] (2) updating the current dual-selection stereo frame so that the dual-selection stereo frame just calculated becomes the current dual-selection stereo frame;

[0040] (3) recalculate the center of the currently selected stereo frame to make it the new rotation center;

[0041] S1-4-2, automatic adjustment of the selected stereoscopic frame posture: using formula V dif =V i+1 -V i The direction difference calculation of two adjacent selected stereo frames is realized, and the difference is used as the basic parameter for adjusting the selected stereo frame to obtain the complete transformation matrix TMatrix at the i+1th position in the selected stereo frame, and the posture of the selected stereo frame is adjusted by using TMatrix;

[0042] S1-5, cutting and extraction of contact network facilities.

[0043] The specific method of step S1-5 is: perform point-by-point distance accumulation calculation on the POS trajectory points, and when the accumulated distance is greater than or equal to the offset distance, crop the point cloud in the area of ​​the box by selecting the stereo box CBox to complete the cropping and extraction operations.

[0044] The specific method of step S2-2 is: first, based on the attention mechanism feature extraction method of ECA and CBAM, the important features in the point cloud are strengthened from the channel and spatial domains; then, a residual refinement structure based on the OCS initial classification result is introduced, which realizes the refinement of the contact network facility classification results through multi-scale receptive field feature extraction and fusion.

[0045] The specific method of step S2-2 is: eliminate the T-shaped input feature transformation in the original structure of the MFF_A model, use the MLP module to extract point cloud features, and use the ECA channel attention mechanism to realize channel enhancement of the extracted features; then use the CBAM channel and spatial attention mechanism to perform the shallow features processed by ECA to enhance the shallow extracted features, and fuse the multiple layers of shallow features with the deep features; then use the ECA channel attention mechanism to process the fused features, and use the activation function to divide the point cloud into n categories of contact network facilities, input the initial classification results into the Refine structure, and finally output the refined classification results.

[0046] The MFF_A model includes the following modules:

[0047] 1) ECA module: The original ECA module is improved to global average pooling (GAP) and global maximum pooling (GMP) and used together:

[0048]

[0049] Where |t| odd Find the nearest odd number to t; γ and b are 2 and 1 respectively;

[0050] After the shared weight MLP feature extraction, a multi-dimensional feature map is obtained, which is used as the input of the ECA module, and the input feature dimension is kept consistent with the output feature dimension to prevent the feature dimension from being reduced; the grouped convolution strategy is used to capture cross-channel interactions; for a given fully connected layer, the grouped convolution divides it into multiple groups and performs linear transformations independently in each group, and the interaction between each channel and its adjacent channels, so the weight calculation formula is:

[0051]

[0052] In the formula, Represents a set of adjacent channels;

[0053] In this way, each channel attention module affects the K*C parameters, making all channels share the same learning parameters;

[0054] 2) CBAM module: The direct jump connection adopts the jump connection based on CBAM to strengthen the important features from the two dimensions of channel and space. To this end, the input features pass through the channel and spatial attention mechanisms in turn, so that each branch can learn "What" and "Where" on the channel and spatial axes respectively, strengthening the effective transmission of important features in the network;

[0055] 3) Pyramid pooling module: obtain feature tensors through convolution operations, then obtain feature tensors of different scales through global maximum pooling operations, then perform convolution operations on feature tensors of different scales to reduce their dimensions, and finally perform deconvolution operations to facilitate the subsequent superposition and fusion operations of these feature tensors;

[0056] 4) Channel feature enhancement structure: On the one hand, the feature dimension is converted from the original 64 dimensions to 16 dimensions through the convolution layer, and the ECA module is introduced to strengthen the learning of channel features and improve the expression ability of this feature dimension in important features; then the feature enhancement results are superimposed with the feature extraction results of each scale in PPM, and the features of these four scales are superimposed and fused through concat connection to achieve the enhancement and fusion of multi-scale features; on the other hand, the average pooled feature tensor of the feature tensor (24, 4096, 1, 64) is obtained through GAP and superimposed with the multi-scale feature fusion tensor obtained in the previous step to achieve feature enhancement.

[0057] The CBAM module includes a channel attention submodule and a spatial attention submodule, wherein:

[0058] In the channel attention submodule, the spatial information of the feature map is first aggregated using average and maximum pooling operations on the input feature F to generate two different spatial context descriptors: and Represent the average pooling feature and the maximum pooling feature respectively; then, both descriptors are fed into a shared MLP structure to generate a channel attention feature map Set the hidden activation size to Where r is the reduction ratio; after applying the shared network to each descriptor, the output feature vector is merged by element accumulation, and the channel attention is calculated as follows:

[0059]

[0060] Where σ represents the activation function, W0 and W1 refer to the weights of the MLP, and the condition W0 and W1 share weights between the input features and the subsequent ReLU activation function;

[0061] The spatial attention submodule uses GAP and GMP to perform pooling operations on the channel enhancement feature F' and connects the two to generate an effective feature descriptor. On the connected feature description, a convolutional layer is used to generate a spatial attention feature map. And encode it to achieve feature emphasis or suppression; aggregate the channel information of the feature map through two pooling operations to generate two types of feature maps: and Each average and maximum pooling feature adopts the channel attention mechanism, and then they are connected and convolved through a standard convolution layer to generate a 2D spatial attention feature map. The spatial attention formula is shown as follows:

[0062]

[0063] In the formula, σ represents the activation function, f 7×7 Represents a 7×7 convolution kernel for convolution operation.

[0064] The reconstruction of the contact network 3D model in S3 includes:

[0065] 1) Model reconstruction of contact line and suspension string: First, the trajectory line segmented fitting algorithm is used to realize the straight line fitting and model reconstruction of the suspension string; secondly, the segmented fitting suspension string model is used to further segment the contact line point cloud, and the contact line between adjacent suspension strings is also fitted and model reconstructed using the segmented fitting algorithm;

[0066] 2) Model reconstruction of the locator component: The locator adopts a rectangular tube, and a standard locator three-dimensional model is established according to the geometric parameters of the rectangular tube. The length direction vector of the established model is set to (1,0,0); the nearest point iterative registration algorithm is used to align the locator point cloud obtained by semantic segmentation with the standard locator model to achieve automatic reconstruction of the locator point cloud.

[0067] The geometric parameter detection based on the contact network model in step S4 includes:

[0068] 1) Detection of contact wire height and pull-out value: The detection result of the track centerline is the measurement baseline of the height and pull-out value. The height of the contact network is obtained by subtracting the elevation of the reconstructed contact wire suspension point and the elevation of the track centerline at the corresponding position; the pull-out value of the contact network in the straight section is measured as the distance from the center line of the line after the vertical projection of the contact wire, and is compared with the reconstructed contact wire model parameters and the center line of the line; the pull-out value of the contact network in the curved section is calculated according to the following formula:

[0069] a=m+c,where:

[0070] In the formula: a: contact network pull-out value; m: horizontal distance between the contact line at the positioning point and the center of the line; c: horizontal distance between the center of the pantograph at the positioning point and the center of the line; h: outer rail superelevation; H: contact line height; L: track gauge.

[0071] 2) Locator tilt angle detection: The first endpoint P of the locator model is obtained by iteratively registering the locator model with the locator point cloud. s and the end point P e The spatial coordinate value of .

[0072] In the positioner tilt angle detection, first the end point P e Project it onto the XOY plane to get the projection point P of the end point e '; Secondly, calculate the vector With vector The angle between the two is the slope angle of the positioner. The vector angle θ is calculated as follows:

[0073] In step S1, when collecting laser point cloud data along the railway, the OptechLynx HS 600 vehicle-mounted mobile measurement system is used, in which two laser scanning heads are installed at an angle, and the vehicle-mounted radar system is installed to the rear of the flatbed vehicle through a rigid heightening frame. The laser measurement frequency of the scanner is 600kHz, and the scanning frequency of a single laser head is 400 lines / second; a high-precision laser gyroscope is used to provide accurate position information; a GNSS base station is set up at intervals of about 15 kilometers along the railway, and an odometer is installed on the wheels of the train to assist the positioning and attitude measurement system in performing zero-speed correction.

[0074] Compared with the prior art, the present invention has the following beneficial effects:

[0075] 1. The present invention sequentially removes background interference points through double-selection stereo frame posture determination and positioning, double-selection stereo frame offset vector determination, and automatic adjustment of the selected stereo frame posture assisted by along-track POS data, thereby realizing automatic search and extraction of contact network facility point cloud information from the original three-dimensional point cloud scene;

[0076] 2. The present invention is based on the feature extraction method of the attention mechanism of ECA (Efficient Channel Attention) and CBAM (Convolutional Block Attention Module), which strengthens the feature extraction capability, promotes the fusion of multi-level shallow features and deep features, and adopts the residual refinement structure of the initial segmentation result. Based on this, a semantic segmentation model based on MFF_A is proposed to realize the classification of contact network facilities; the three-dimensional models of the contact line and the suspension string are reconstructed by piecewise straight line fitting, and the three-dimensional model of the locator is reconstructed by cube fitting, realizing the geometric parameter detection based on the contact network model;

[0077] 3. The fully automatic search and extraction method of the present invention is fast and has high extraction accuracy, and is applicable not only to straight lines but also to undulating terrain and curves;

[0078] 4. The present invention performs semantic segmentation of the contact network based on deep learning, making the accuracy and computational efficiency better than similar algorithms;

[0079] 5. The present invention introduces a three-dimensional reconstruction method for the contact network by fitting straight lines and cubes, and the measurement accuracy of the contact network geometric parameters can meet the measurement requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 This is the basic structural diagram of the overhead contact network;

[0081] Figure 2 A flowchart of a method for automatically searching and extracting overhead line facility samples in a three-dimensional scene of the present invention;

[0082] Figure 3 A study plot diagram used in an embodiment of the present invention;

[0083] Figures 4a-4d They are respectively partial laser point clouds collected in the station range, roadbed range, bridge range, and tunnel range in the embodiments of the present invention;

[0084] Figure 5 It is a schematic diagram of automatically searching and extracting the contact network point cloud based on the double-selection stereo frame in the present invention;

[0085] Figure 6 is a point cloud and background point cloud detail image of an extraction result in an embodiment of the present invention;

[0086] Figure 7 is a point cloud and background point cloud detail image of another extraction result in an embodiment of the present invention;

[0087] Figure 8a-8c The detailed diagram of the extraction results of the two arms in the overhead line are the side view, top view and front view respectively;

[0088] Figure 9a-9b The detailed images of the extraction results of the two arms in the overhead contact network are the side view, top view and front view respectively;

[0089] Fig.10 Detailed front view of the extraction results of the two arms in the overhead contact network

[0090] Fig.11 The following is a schematic diagram of the eight component categories of overhead line facilities;

[0091] Fig.12 The MFF_A network model diagram constructed for the present invention;

[0092] Fig.13 It is the schematic diagram of ECA;

[0093] Fig.14 This is the schematic diagram of CBAM;

[0094] Figure 15-18 The experimental results of the three existing methods and the method of the present invention are classified into detailed diagrams;

[0095] Fig.19 It is a comparison chart of MACs and parameter quantities;

[0096] Fig. 20 Reconstruction diagram of the contact line segmented three-dimensional model;

[0097] Fig.21 , 22 They are schematic diagrams of the positions of the locators at the hard span and the locators at the column respectively;

[0098] Fig.23 This is a schematic diagram of the method for measuring the conduction height and pull-out value;

[0099] Fig.24 It is a schematic diagram of the locator and slope;

[0100] Fig.25 This is a rendering of the effect of automatically searching and extracting the contact network based on the double-selection stereo frame in the present invention. DETAILED DESCRIPTION

[0101] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0102] Embodiment 1

[0103] This embodiment selects a section of the Guiyang-Guangzhou railway to conduct an experiment to verify the effectiveness of the method of the present invention. Figure 3 As shown in the figure, this line is an inter-regional high-speed railway in China connecting Guiyang and Guangzhou, with a total length of 201.32km. All tracks in Guangxi are ballastless tracks, with an actual operating speed of 250km / h. The line from Hezhou Station to Huaiji Station, about 100.66km, was selected for round-trip scanning, and measurements were taken along the left and right lines of the railway.

[0104] In order to quickly and effectively collect data along the railway, the present invention improves the existing VMMS so that it can more quickly and effectively obtain point cloud data of the contact network. For LiDAR point cloud data, the Optech Lynx HS 600 vehicle-mounted mobile measurement system is used, and its minimum measurement distance is 1.5 meters. The two laser scanning heads of the Lynx HS 600 are installed at an angle, which can effectively reduce the scanning blind area along the railway. A customized rigid heightening frame is used to install the vehicle-mounted radar system to the rear of the flatbed truck. The heightening frame can increase the distance from the center of the scanner to the bottom rail. The top of the bracket has a quick fixing device for easy installation. The heightening bracket and the vehicle are fixed with binding straps. The acquisition parameters of the scanner are: the laser measurement frequency is 600kHz, and the scanning frequency of a single laser head is 400 lines / second.

[0105] In order to obtain accurate position information, a high-precision laser gyroscope is used. When the satellite signal is good, the gyroscope has a roll and pitch accuracy of 0.005 degrees and a heading accuracy of 0.015 degrees. When the satellite signal is lost, it can maintain its nominal accuracy within 60 seconds. During the scan, a GNSS base station is set up at intervals of about 15 kilometers along the line, with a total of 7 stations set up. The sampling frequency of the base station is uniformly set to 1HZ, and GNSS signals are received uninterruptedly throughout the whole process. When the VMMS enters the railway tunnel, the positioning attitude measurement system (POS) cannot receive satellite signals. When the driving time in the tunnel exceeds 60 seconds, it is necessary to stop and stand still at the entrance and exit of the tunnel respectively to improve the position and attitude accuracy of the POS. Since the position drifts when the POS stops and stands still, the odometer encoder (DMI) is installed on the train wheel to assist the POS in zero-speed correction. In order to achieve precise synchronization between the rotation of the DMI and the train wheel, the DMI needs to be fitted on the metal protective cover, and then the metal protective cover is installed on the outside of the train drive shaft. To do this, first remove the protective cover on the outside of the vehicle's drive shaft, and then fix the metal protective cover with integrated DMI on the outside of the vehicle's drive shaft.

[0106] Finally, in the actual data collection phase, since the railway was in normal operation, we chose to scan from 23:00 to 03:00 on two days (no other trains were running on the railway line), for a total of 8 hours. During the scanning process, the average vehicle speed was about 60km / h. Since there were many long tunnels or tunnel groups in this section, it was necessary to stop the vehicle before entering the tunnel to allow the positioning and attitude measurement system to be stationary. Each stop lasted 5 to 10 minutes. During the waiting period, the scanner stopped data collection, but the positioning and attitude measurement system continued to work normally. The final point cloud data was about 180GB. Figures 4a-4d Shown are partial laser point clouds of stations, roadbeds, bridges, and tunnels collected in this embodiment.

[0107] The fully automatic search and extraction method of contact network facilities based on a stereoscopic frame of the present invention includes four steps: automatic search and extraction of contact network facilities, semantic segmentation of contact network facility point cloud based on deep learning, reconstruction of contact network facility three-dimensional model and detection of contact network geometric parameters. Each step is described in detail below:

[0108] S1, automatic search and extraction method of contact network based on double-selection stereo frame:

[0109] In the point cloud data obtained, the roadbed section accounts for 21.03%, the tunnel section accounts for 38.30%, the bridge section accounts for 40.67%, and the contact network facility point cloud accounts for no more than 1% of the entire point cloud scene. The point cloud scene mainly includes 87% of ground points, 8% of track points and 4% of facade baffles, vegetation, etc. If the deep learning network is directly used to perform semantic segmentation on the original data point cloud, the extremely unbalanced category samples (Classimblance) may lead to poor generalization ability of the network and easy overfitting, which will eventually lead to misclassification and omission of contact network components. Therefore, it is necessary to pre-extract the contact network facilities from the original data point cloud scene to eliminate the interference of other categories of point cloud objects on the identification of contact network facilities.

[0110] In order to reduce the amount of calculation in translation and rotation of point clouds and assist in cutting POS track points within a certain range, the present invention adopts a scene-based double-selection stereo frame contact network automatic search and extraction method based on POS information: first, the search range is determined by roughly selecting a stereo frame, and the track line points within the range are extracted according to this frame; then, the tracking, cutting and extraction are performed along the track direction by selecting a stereo frame, and the effect is as follows: Fig.25 See Figure 5 , the method comprises the following steps:

[0111] S1-1, rough selection of stereo frame pose and positioning:

[0112] The rough selection stereo box is a boundary stereo box within a certain range constructed in the three-dimensional scene. The three-dimensional point cloud that meets the conditions is selected from the scene with this box as the constraint range, which satisfies the following formula:

[0113] BBox={min_point,max_point}

[0114] Where min_point and max_point represent the minimum and maximum points of the coordinate values ​​within the range of the rough selected stereo frame, respectively. Generally, the center point of the frame is located near the center points of two pairs of adjacent columns.

[0115] The first rough selection stereo frame in the scene is set manually. When setting the maximum and minimum points, it is necessary to consider that the area of ​​the rough selection stereo frame formed by it can include the point cloud area of ​​the track line points and all facilities of the contact network. In terms of posture determination, the six planes formed are parallel to the XOY, XOZ and YOZ planes respectively. In terms of positioning, the focus is on placing the center point of the frame in the middle of two pairs of adjacent columns and the track line points on the XOY projection plane, thereby reducing the selection error caused by translation on the track. In the process of cutting and extracting, the rough selection stereo frame moves along the track direction. No processing is required in the posture of the frame. In terms of positioning, the rough selection stereo frame needs to be moved to the next position based on the spacing between the columns.

[0116] S1-2, Selected stereo frame pose and positioning:

[0117] The selected stereo frame refers to the minimum cropped stereo frame along the track direction, and the point cloud of the contact network facilities along the track is obtained by using this frame, which satisfies the following formula:

[0118] CBox=RMatrix*minBox+T

[0119] In the formula, minBox is the minimum three-dimensional frame for cutting a pair of overhead line column support facilities. RMatrix is ​​the rotation matrix, and T is the translation vector. The main function of this operation is to rotate and translate minBox, transform its long side to be parallel to the track, and the center point is located near the center points of the two support facilities, so as to achieve the goal of cutting the minimum three-dimensional frame along the track. The specific implementation is shown in the following formula:

[0120]

[0121] The first selected stereo frame in the scene is also set manually. The selected stereo frame needs to have appropriate length, width and height, and must satisfy the minimum cropped stereo frame to be completely contained in the rough selection frame, and the long side of the selected stereo frame is parallel to the track and the width is perpendicular to the track. The unprocessed minimum stereo frame minBox is rotated and translated to obtain the final selected stereo frame CBox.

[0122] S1-3, determination of the offset of the double-selection stereo frame:

[0123] Due to factors such as terrain undulations and obstacles, the track inevitably has undulating terrain and curves. In the along-track search calculation of the coarse selection stereo frame and the fine selection stereo frame, it is necessary to further consider how to set the offset vector of the center of the double selection stereo frame along the track line to realize the along-track movement and clipping of the double selection stereo frame. Among them, the calculation of the offset vector T of the adjacent stereo frame includes the determination of the offset direction and the offset distance. The specific process is as follows:

[0124] S1-3-1, Determination of offset distance:

[0125] The offset distance of the double-selection stereo frame along the track in the scene needs to be adjusted based on the spacing between adjacent columns. First, the position of the center point of each column {Centerpoint1,...,Centerpoint i ,...,Centerpoint n}, by projecting these center points onto the XOY plane, the distances from the centers of the adjacent columns are calculated to obtain the offset distance information of the double-selection stereo box:

[0126] PoleDis[i]=Dis(Centerpoint[i+1]-centerpoint[i]),i=1,...,N-1

[0127] Where PoleDis[i] refers to the Euclidean distance between the i-th pole and the i+1-th pole in the XOY plane; Dis(·) is the function for calculating the Euclidean distance between two points in the XOY plane. Centerpoint[i+1] refers to the coordinate information of the center point of the i+1-th pole.

[0128] S1-3-2, Determination of offset direction:

[0129] The offset direction is determined by calculating the vector difference between the POS tracks of two adjacent rough selection boxes. All POS track points in the rough selection box are obtained, and the two adjacent POS track points are used as the direction of the railway track. When the rough selection box moves along the track for searching, Figure 5 As shown in the figure, the offset direction formula of two adjacent coarse-selected stereo frames is:

[0130] V dif =V i+1 -V i

[0131] Where V i+1 is the POS trajectory direction vector in the i-th rough selected stereo frame, V i+1 is the POS trajectory direction vector in the i+1th rough selected stereo frame, V dif Represents the direction vector between two adjacent coarse-selected cubic boxes i and i+1.

[0132] The double selection stereo box offset vector T is obtained through the above method.

[0133] S1-4, trajectory data-assisted automatic adjustment of selected stereoscopic frame poses, includes the following steps:

[0134] S1-4-1 Selected stereo frame center adjustment:

[0135] After determining the offset vector of the double-selected stereo frame, the selected stereo frame needs to be translated to the next area to be cropped. The specific steps are as follows:

[0136] (1) adjusting the double-selection stereo frame in translation according to the offset vector;

[0137] (2) updating the current dual-selection stereo frame so that the dual-selection stereo frame just calculated becomes the current dual-selection stereo frame;

[0138] (3) Recalculate the center of the currently selected stereo frame to make it the new rotation center.

[0139] S1-4-2 Selected stereo frame posture automatic adjustment:

[0140] The posture of the selected stereo frame should change with the ups and downs of the railway track and the changes in the curves, so that it can be more helpful to cut and extract the contact network facilities that meet the accuracy requirements. dif =V i+1 -V i The direction difference calculation of two adjacent selected stereo frames is realized. This difference is used as the basic parameter for adjusting the selected stereo frame, and the complete transformation matrix TMatrix (rotation center, translation, rotation) at the i+1th position in the selected stereo frame is obtained, and the posture of the selected stereo frame is adjusted using TMatrix.

[0141] S1-5 Cutting and extraction of contact network facilities:

[0142] After obtaining the selected stereo frame of the OCS facility, the 3D point cloud in the scene is cropped. The distance of the POS trajectory points is accumulated point by point. When the accumulated distance is greater than or equal to the offset distance (i.e., the distance between the adjacent columns), the point cloud in the area of ​​the selected stereo frame CBox is cropped to complete the cropping and extraction operations.

[0143] After completing the automatic search and extraction of the contact network facilities in the stereo frame, the point cloud data of the target domain is cropped by searching the stereo frame, which can not only effectively eliminate the interference of the non-contact network point cloud, but also automatically eliminate the interference of the column point cloud that accounts for a relatively large proportion in the contact network, and obtain the segmented contact network facility point cloud data without columns.

[0144] Finally, each selected stereo frame is cropped and the extracted point cloud data is stored separately to facilitate the training and prediction of the deep learning-based network model in S2.

[0145] See also Figure 5 , the contact network point cloud is segmented from the original laser point cloud using the above steps S1-1 to S1-5, as follows:

[0146] The original laser point cloud data is used to conduct an automatic search and extraction test of contact network facilities based on a stereo frame. The first rough selection stereo frame and the fine selection stereo frame are manually set, and the GNSS trajectory points in the experimental data area are used as auxiliary in the algorithm. This embodiment uses the upper left corner of the full point cloud as the starting point to determine the position of the rough selection frame; the fine selection frame is within the rough selection frame, and its length, width and height parameters are 49, 11, 2. The cropping and extraction accuracy is 96.4%, and the point cloud processing rate is 2 million points per second. Figure 6-Figure 10 It can be seen that ground points, track points, facade baffles and columns have been excluded from the search and extraction selection box, and the cropped data can meet the OCS component segmentation requirements.

[0147] S2, semantic segmentation of contact point cloud based on deep learning:

[0148] In this step, the overhead contact network facilities are divided into eight categories, namely, straight arm, inclined arm, elastic sling, conductor, positioning tube, positioner, suspension string, and load-bearing cable. Fig.11 The specific steps include the following:

[0149] S2-1, manually mark the point cloud data:

[0150] Considering that the density and quantity of point clouds between different contact network facilities are relatively different, the semantic segmentation of contact network based on deep learning needs to focus on the imbalance of classification samples. To this end, spatial and channel attention mechanisms are introduced in the construction of deep learning networks, and the semantic segmentation model based on multi-scale feature fusion and attention mechanism (MFF_A) is used to classify contact network facilities. First, based on the feature extraction method of the attention mechanism of ECA and CBAM, the important features in the point cloud are strengthened from the channel and spatial domains; then, the residual refinement structure based on the initial OCS classification result is introduced, which realizes the refinement of the contact network facility classification results through multi-scale receptive field feature extraction and fusion.

[0151] S2-2, construct MFF_A model:

[0152] The MFF_A in the present invention uses PointNet as the skeleton structure, eliminates the T-shaped input feature transformation in the original structure, adopts the MLP module (two or more convolutional layers with weight sharing) to extract point cloud features, and uses the ECA channel attention mechanism to achieve channel enhancement of the extracted features. Then, the shallow features processed by ECA are subjected to the channel and spatial attention mechanism of CBAM to enhance the shallow extracted features, and fuse the multiple layers of shallow features with the deep features. Next, the fused features are processed by the ECA channel attention mechanism, and the point cloud is divided into n categories of contact network facilities through the activation function. Next, the initial classification results are input into the Refine structure, and finally the refined classification results are output, such as Fig.12 shown.

[0153] The MFF_A model of the present invention mainly includes the following modules:

[0154] 1)ECA module

[0155] Existing studies have shown that embedding the attention module into CNN can bring significant performance improvements (such as SENet, CBAM, ECANet, EPSANet). However, in the traditional PointNet network structure, shared weight MLP is used to implement point cloud feature extraction, and different numbers of convolution kernels are combined to achieve dimensionality increase or decrease of extracted features, and the convolution kernel size is [1,1]. The introduction of ECA improves the ability to extract important features during global feature extraction. For the original ECA module, it is improved to Global Average Pooling (GAP) and Global Max Pooling (GMP) and used together:

[0156]

[0157] Where |t| odd Find the nearest odd number to t. In the present invention, γ and b are set to 2 and 1, respectively. Obviously, the ψ function enables larger-scale channels to have long-range interactions, and vice versa.

[0158] See also Fig.13 , it can be seen that after the shared weight MLP feature extraction, a multi-dimensional feature map is obtained, which is used as the input of the ECA module, and the input feature dimension is kept consistent with the output feature dimension, which can effectively prevent the reduction of feature dimension. In addition, the grouped convolution strategy is applied to capture cross-channel interactions. For a given fully connected layer, the grouped convolution divides it into multiple groups and performs linear transformations independently in each group. The interaction between each channel and its adjacent channels. Therefore, the weight calculation formula is:

[0159]

[0160] In the formula, Represents a set of adjacent channels. This formula captures the local interaction across channels. This local range constraint effectively avoids the interaction across all channels, thereby improving the efficiency of the network model. In this way, each channel attention module affects the K*C parameter. In order to further reduce the complexity of the model and improve efficiency, all channels share the same learning parameters.

[0161] 2) CBAM module

[0162] CBAM is mainly used in the encoder-decoder network architecture to solve the problem that although the network skip connection can improve the fusion of shallow features and deep features, it cannot effectively reduce the degradation of the neural network. Since the MLP operation in the present invention extracts features by mixing cross-channel and spatial information, the direct skip connection can be replaced by a CBAM-based skip connection to strengthen important features from both the channel and spatial dimensions. Fig.14 To achieve this goal, the input features are sequentially passed through the channel and spatial attention mechanisms so that each branch can learn "What" and "Where" on the channel and spatial axes respectively, thereby strengthening the effective transmission of important features in the network. It mainly includes the channel attention submodule and the spatial attention submodule, as follows:

[0163] (1) Channel attention submodule:

[0164] The implementation of the channel attention structure in CBAM is as follows:

[0165] First, the spatial information of the feature map is aggregated using average and maximum pooling operations on the input feature F to generate two different spatial context descriptors: and Represent the average pooling feature and the maximum pooling feature respectively. Next, both descriptors are fed into a shared MLP structure to generate a channel attention feature map To reduce the number of parameters, the hidden activation size is set to Where r is the reduction ratio. After applying the shared network to each descriptor, the output feature vector is merged by element accumulation. The channel attention calculation formula is as follows:

[0166]

[0167] Where σ represents the activation function, W0 and W1 refer to the weights of the MLP, and the condition W0 and W1 are shared weights by the input features and the subsequent ReLU activation function.

[0168] (2) Spatial Attention Submodule:

[0169] The spatial attention focuses on the "Where" part. Specifically, GAP and GMP are used to perform pooling operations on the channel enhancement feature F', and the two are connected to generate an effective feature descriptor. On the connected feature description, a convolutional layer is used to generate a spatial attention feature map. And encode it to achieve feature emphasis or suppression. Through two pooling operations, the channel information of the feature map is aggregated to generate two types of feature maps: and Each average and maximum pooling feature adopts the channel attention mechanism, and then they are connected and convolved through a standard convolution layer to generate a 2D spatial attention feature map. The spatial attention formula is shown as follows:

[0170]

[0171] In the formula, σ represents the activation function, f 7×7 Represents a 7×7 convolution kernel for convolution operation.

[0172] 3) Pyramid pooling module (PPM)

[0173] In complex scene parsing tasks, contextual relationships are very important. If they are separated from contextual relationships, many prediction results may have incorrect relationship matches; small-sized targets in complex scenes are likely to be ignored, while large-sized targets may exceed the receptive field of network feature extraction, resulting in discontinuous segmentation problems. The introduction of the pyramid pooling module (PPM) can solve this problem. The pyramid pooling module mainly obtains feature tensors through convolution operations, and then obtains feature tensors of different scales through global maximum pooling operations. Convolution operations are then performed on feature tensors of different scales to reduce dimensionality. Finally, a deconvolution operation is performed to facilitate the subsequent superposition and fusion operations of these feature tensors.

[0174] The PPM structure can be used to effectively obtain contact network point cloud feature information of different scales, which is helpful for the identification of small target objects. At the same time, the PPM structure can obtain context-based contact network facility feature information and effectively extract contact network facility target information with spatial relationships, such as the catenary should be above the conductor, the suspension string and the catenary, the conductor maintain an adjacent relationship, the straight arm and the inclined arm maintain an adjacent relationship, the positioning tube and the locator also maintain an adjacent relationship, etc.

[0175] 4) Channel feature enhancement structure, including the following two aspects:

[0176] On the one hand, the feature dimension is converted from the original 64 dimensions to 16 dimensions through the convolution layer, and the ECA module is introduced to strengthen the learning of channel features and improve the expression ability of this feature dimension in important features. Then the feature enhancement results are superimposed with the feature extraction results of each scale in PPM, and the features of these four scales are superimposed and fused through concat connection to achieve the enhancement and fusion of multi-scale features.

[0177] On the other hand, the average pooled feature tensor (24, 4096, 1, 64) obtained through GAP is superimposed with the multi-scale feature fusion tensor obtained in the previous step to achieve the goal of feature enhancement.

[0178] This channel feature enhancement module only adds a small number of parameters to achieve significant performance gains. Through the analysis of the channel attention module in ECA, experience shows that avoiding dimensionality reduction is very important for learning channel attention, and appropriate cross-channel interaction can significantly reduce the complexity of the model while maintaining performance.

[0179] The experimental environment of this embodiment is: the deep learning framework is keras, and the python integrated development environment is Pycharm. Hardware: the CPU is Intel(R) Core(TM) i7-9700K 3.60GHz, the memory is 32GB, and the graphics card is NVIDIAGeForce RTX 2080Ti 11G.

[0180] Through step S2, the contact network point cloud is divided into 8 categories: straight arm, inclined arm, elastic sling, conductor, positioning tube, locator, suspension string and load-bearing cable. The specific classification accuracy is as follows:

[0181] The present invention selects representative typical point cloud segmentation algorithms PointNet, PointNet++, comparative literature (Lin, S.; Xu, C.; Chen, L.; Li, S.; Tu, X. LiDAR point cloud recognition of overheadcatenary system with deep learning. Sensors 2020, 20, 2212., hereinafter referred to as “comparative literature”) and the algorithm of the present invention, wherein the number of neighboring points in the KNN algorithm is set to 16.

[0182] The comparison results of the algorithm of the present invention with the three comparison algorithms in terms of precision (P), intersection over union (IoU) and average precision (MA) are shown in the following table:

[0183] Comparison with similar algorithms (%)

[0184]

[0185] (1) Compared with three other similar algorithms, our algorithm performs best in terms of average precision and intersection-over-union ratio, with the highest precision and intersection-over-union ratio of 96.7% and 93.70% respectively.

[0186] (2) The order of average precision and intersection-over-union ratio from high to low among the four algorithms is the algorithm in this paper, the comparative literature, PointNet++ (MSG), and PointNet. Among them, PointNet performs the worst because PointNet mainly uses MLP to extract global features of contact network facilities and lacks local feature extraction; secondly, Multi-scale Grouping PointNet++ (MSG) improves this problem by performing local sampling and grouping of point clouds, and introducing step-by-step downsampling and Encoder-decoder structure, and using jump connections to fuse shallow features with deep features to obtain good semantic segmentation results; in the comparative literature, the context features of each single-frame point cloud are mainly extracted through the improved PointNet, and then the single-frame data is analyzed by neighboring point distances and combined with feature extraction units to achieve semantic segmentation of the contact network. The algorithm in this paper introduces spatial and channel attention mechanisms, combines multi-level shallow features with deep features, and uses the residual refinement structure of the initial segmentation results to more effectively extract multi-scale feature information. Therefore, the algorithm in this paper is leading in accuracy, intersection-over-union ratio, and average precision;

[0187] (3) For most types of overhead line facility point clouds, the proposed algorithm is superior to other algorithms in terms of accuracy and intersection-over-combination ratio, especially for positioning tubes, straight arms, inclined arms, conductors, and catenary cables, which are leading in terms of accuracy; and straight arms, inclined arms, and conductors are leading in terms of intersection-over-combination ratio;

[0188] (4) Overall, it can be seen that the accuracy and intersection ratio of the straight wrist arm and oblique wrist arm are at least 1.8% and 1.2% higher than those of other algorithms. Therefore, the algorithm has better practicality.

[0189] The classification details of the experimental results of the four methods are further compared, such as Figure 15-18 As shown in the figure, we can see the local classification diagram for the same line segment, with different colors representing different types of overhead contact network facilities. The following conclusions can be drawn:

[0190] (1) The classification accuracy of straight wrist and oblique wrist is lower than that of the other three algorithms, and more misclassification occurs. Fig.15 The straight arm and the oblique arm in the middle rectangular box are mistakenly classified as wires. This is because the PointNet network uses the global features of the entire point cloud sample and does not extract local feature information. In the case where the types of contact network facilities are similar, it may not be possible to effectively identify the corresponding contact network facility type, resulting in misclassification. However, PointNet++ (MSG) and the comparative literature can effectively obtain local features and fuse them with global features to improve feature expression capabilities. The algorithm of the present invention uses a multi-scale feature extraction and fusion module to effectively focus on feature information of different details, see Figure 16-18Compared with the four algorithms, the proposed algorithm is the best in the segmentation of straight wrist and oblique wrist.

[0191] (2) For the classification accuracy of elastic slings, the comparison literature and the algorithm of the present invention have better segmentation accuracy than PointNet and PointNet++ (MSG). The worst performance is PointNet++ (MSG), where a large number of elastic sling point clouds are misclassified as wire types. The algorithm of the present invention is better than the comparison literature in the segmentation of elastic sling details. For example, a small number of point clouds in the elastic sling area shown in the rectangular box in Figure 17 are misclassified as wires, while Fig.18 There has been significant improvement.

[0192] (3) Regarding the classification accuracy of positioning tubes and locators, both PointNet++ (MSG) and the algorithm of the present invention performed well. There were no large numbers of cases where positioning tubes were misclassified as locators or wires. The recognition accuracy of positioning tubes was higher than that of locators. However, the algorithm of the present invention misclassified a small part of the point cloud where the locator touched the wire as a locator.

[0193] (4) The classification accuracy of the suspension string is superior to other algorithms in the literature, with fewer misclassifications. This is because the feature extraction unit (FEU) can effectively extract the distance information of neighboring point clouds and adopt the MLP structure to extract higher-dimensional features.

[0194] In order to evaluate the parameter complexity of the algorithm of the present invention, MACs calculation and comparison are used, and the results are as follows. Fig.19 It can be seen that the present invention has the lowest MACs and total parameters in the segmentation algorithm. This is because the method of the present invention uses fewer fully connected operations, uses an MLP structure with a smaller convolution kernel, and introduces a lightweight attention mechanism ECA, which basically does not increase the number of parameters. In addition, the multi-scale feature extraction and fusion module helps to improve the accuracy while the number of parameters is effectively controlled. By further segmenting the initial segmentation results with a refined structure, a higher-precision result is obtained. This reduces the complexity of the model.

[0195] S3. Reconstruction of contact network 3D model based on geometric features

[0196] After completing the point cloud segmentation of the eight components in step S2, the contact network 3D model needs to be reconstructed. Here, only the model reconstruction method of the representative contact wire, dropper and locator is described. These three reconstructed models can be used for geometric detection in S4.

[0197] Considering the model parameters of the various components of the contact network, the three-dimensional model of the contact line and the suspension string can be reconstructed by piecewise straight line fitting; the three-dimensional model of the oblique arm and the flat arm can be reconstructed by cylindrical fitting; the three-dimensional model of the locator can be reconstructed by cubic fitting, as follows:

[0198] 1) Model reconstruction of contact line and dropper

[0199] The contact wire is a key facility for power supply of electrified railway locomotives. In order to ensure that the pantograph on the top of the train and the contact wire are evenly stressed, vertical suspension strings are placed at certain intervals between adjacent contact network supports to raise the height of the contact wire so that the height of the contact wire relative to the rail remains fixed. Affected by gravity, the contact network suspension strings are in the shape of vertical straight lines, which meet the characteristic requirements of straight line fitting. First, the aforementioned trajectory line segmented fitting algorithm of the present invention is used to achieve straight line fitting and model reconstruction of the suspension strings. Secondly, the segmented fitting suspension string model is used to further segment the contact line point cloud. The contact lines between adjacent suspension strings are also fitted and model reconstructed using the segmented fitting algorithm. The segmented reconstructed contact line model is as follows: Fig. 20 shown.

[0200] 2) Model reconstruction of locator components

[0201] Positioners are mostly made of rectangular tubes, such as Fig.21 , Fig. 22 As shown, the present invention establishes a standard locator three-dimensional model based on the geometric parameters of the locator rectangular tube. Because the locator slope value is usually set to 6° to 16°, and the angle with the horizontal plane is small, the length direction vector of the constructed model is set to (1,0,0). The closest point iterative registration (ICP) algorithm is used to align the locator point cloud obtained by semantic segmentation with the standard locator model to achieve automatic reconstruction of the locator point cloud. During the iterative registration process, noise points far away from the locator model will also be eliminated.

[0202] S4, Geometric parameter detection based on contact network model

[0203] See also Fig.23 , the contact network geometric parameters are important data for evaluating the state of the contact network, mainly including the height of the conductor and the pull-out value. In order to make the carbon slide plate of the pantograph of the electric locomotive wear evenly, the overhead contact line needs to be reasonably placed in a "zip-zap" shape. The offset of the positioning point to the center line of the pantograph is called the pull-out value. Due to the limitation of the length of the pantograph carbon slide plate, the pull-out value of the contact line is required to be limited to a certain range. When the pull-out value is too large, the contact line is easy to exceed the limit in bad weather such as strong winds, resulting in bow-net accidents such as bow scraping or bow drilling; when the pull-out value is too small, the area where the contact line acts on the pantograph carbon slide plate is too concentrated, which consumes the service life of the carbon slide plate. The height of the conductor refers to the vertical distance from the bottom of the contact wire to the rail surface connection line. The height of the contact line is an important indicator for evaluating the working state of the contact network. If the height of the conductor is too large, the pantograph may be offline, arcing may occur, and the contact line and pantograph may be worn; if the height of the conductor is too small, bow drilling accidents are prone to occur, affecting the safety of passengers and cargo.

[0204] The positioner is an important component of the contact network structure. It interacts directly with the contact line and the pantograph to complete the current collection of the train. The slope of the positioner is crucial to the contact performance of the pantograph and the safety of operation. During the rapid operation of the train, the vibration and excitation caused by the pantograph-catenary coupling usually loosen the bolt and nut structure of the positioner, resulting in abnormal slope values ​​of the positioner. Abnormal slope values ​​will cause the pantograph to wear faster, hit the positioner, and affect the current collection quality of the pantograph and the catenary. Therefore, regular detection of the positioner slope is of great significance to ensure the safe operation of the train.

[0205] "Interim Technical Requirements for Contact Network Suspension State Detection and Monitoring Devices (4C)" (TJ / GD006-2014) stipulates the range and accuracy of static geometric parameters measured by contact network suspension state detection and monitoring devices, as shown in Table 1.

[0206] Table 1 Measurement range and accuracy of contact network geometric parameters

[0207]

[0208] 1) The detection method of contact line height and pull-out value is as follows:

[0209] The detection result of the track centerline is the measurement baseline for the height and pullout value. The height of the contact network is obtained by taking the difference between the elevation of the reconstructed contact line suspension point and the elevation of the track centerline at the corresponding position. The pullout value of the contact network in the straight section is measured as the distance from the centerline of the line after the vertical projection of the contact line, and is compared with the reconstructed contact line model parameters and the centerline of the line. The pullout value of the contact network in the curved section is calculated according to the following formula:

[0210] a=m+c,where:

[0211] In the formula: a: contact network pull-out value; m: horizontal distance between the contact line at the positioning point and the center of the line; c: horizontal distance between the center of the pantograph at the positioning point and the center of the line; h: outer rail superelevation; H: contact line height; L: track gauge.

[0212] 2) Positioner tilt angle detection

[0213] Positioner and slope diagram as shown Fig.24 As shown. By iteratively registering the locator model with the locator point cloud (ICP), the first endpoint P of the locator model is obtained. s and the end point P e First, the end point P e Project it onto the XOY plane to get the projection point P of the end point e '; Secondly, calculate the vector With vector The angle between the two is the slope angle of the positioner. The vector angle θ is calculated as follows:

[0214] In order to evaluate the reliability and accuracy of the invented method for detecting the geometric parameters of the contact network, the DJJ-8 laser contact network detector was used to manually measure the collective parameters of the contact network in a certain section in the experimental area. The comparison results are shown in Tables 2 to 4.

[0215] Table 2 Comparison of locator slope detection results

[0216]

[0217] Table 3 Comparison of contact line pull-out value detection results

[0218]

[0219]

[0220] Table 4 Comparison of positioning point and hanging string height detection results

[0221]

[0222] According to the comparative analysis of the detection results of the above two different measurement methods, it can be concluded that, regarding the locator slope, the difference between the two detection values ​​is within 0.5; regarding the contact line pull-out value, the maximum difference between the two detection values ​​is 7mm, and the average value is 3mm; regarding the contact line height, the maximum difference between the two detection values ​​is 8mm, and the average value is 3mm.

[0223] In summary, the detection accuracy of the detection method of the present invention for contact line geometric parameters such as contact line height, pull-out value, and locator slope is comparable to that of conventional special measuring instruments.

Claims

1. A fully automatic search and extraction method for a contact network facility point cloud, comprising the following steps: S1, collect laser point cloud data along the target railway, and automatically search and extract the contact network point cloud in the laser point cloud data based on the double-selection stereo frame: first, determine the search range by roughly selecting the stereo frame, and use this frame to extract the track line points within the range; then, track, crop and extract along the track direction by selecting the stereo frame; wherein: The rough selection stereo frame satisfies the following formula: BBox={min_point,max_point} Where min_point and max_point represent the minimum and maximum coordinate values ​​within the coarse selection stereo frame respectively; The selected stereoscopic frame satisfies the following formula: CBox=RMatrix*minBox+T Where minBox is the minimum three-dimensional box for cutting a pair of overhead line column support facilities; RMatrix is ​​the rotation matrix; T is the translation vector, where: S2, semantic segmentation of the cropped and extracted contact network point cloud based on deep learning, including the following sub-steps: S2-1, manually labeling the point cloud data; S2-2, construct MFF_A model; S3, reconstruction of the contact network 3D model based on geometric features; S4, geometric parameter detection based on contact network model.

2. The method for fully automatic search and extraction of contact network facility point cloud according to claim 1, characterized in that: In step S1, the method for automatically searching and extracting the contact network point cloud in the laser point cloud data based on the double-selection stereo box includes the following sub-steps: S1-1, rough selection of stereo frame pose and positioning: The center point of the rough selection stereo frame is located near the center points of two pairs of adjacent columns; the area of ​​the rough selection stereo frame includes the track line points and the point cloud area of ​​all facilities of the contact network; the six planes of the rough selection stereo frame are parallel to the XOY, XOZ and YOZ planes respectively; on the XOY projection plane, the center point of the rough selection stereo frame falls in the middle of the two pairs of adjacent columns and on the track line point; during the cutting and extraction process, the rough selection stereo frame moves along the track direction, and during the movement process, the rough selection stereo frame is moved to the next position according to the spacing between the columns; S1-2, Selected stereo frame pose and positioning: The selected stereo frame satisfies the requirement that the minimum cropped stereo frame is completely contained in the rough selected stereo frame, and the long side of the selected stereo frame is parallel to the track, and the width is perpendicular to the track; the unprocessed minimum stereo frame minBox is rotated and translated to obtain the final selected stereo frame CBox; S1-3, determining the offset of the double-selection stereo box, includes the following steps: S1-3-1, Determination of offset distance: First, extract the position of the center point of each column from the scene {Centerpoint1,...,Centerpoint i ,...,Centerpoint n }, by projecting these center points onto the XOY plane, the distances from the centers of the adjacent columns are calculated to obtain the offset distance information of the double-selection stereo box: PoleDis[i]=Dis(Centerpoint[i+1]-centerpoint[i]),i=1,...,N-1 Where PoleDis[i] is the Euclidean distance between the i-th column and the i+1-th column in the XOY plane; Dis(·) is the function for calculating the Euclidean distance between two points in the XOY plane; Centerpoint[i+1] is the coordinate information of the center point of the i+1-th column; S1-3-2, determination of the offset direction: obtain all POS track points in the rough selection stereo frame, and use two adjacent POS track points as the direction of the railway track. When the rough selection stereo frame moves and searches along the track, the offset direction formula of two adjacent rough selection stereo frames is: V dif =V i+1 -V i Where V i+1 is the POS trajectory direction vector in the i-th rough selected stereo frame, V i+1 is the POS trajectory direction vector in the i+1th rough selected stereo frame, V dif Represents the direction vector between two adjacent coarse selection stereo frames i and i+1, and obtains the offset vector T of the double selection stereo frame; S1-4, trajectory data-assisted automatic adjustment of selected stereoscopic frame poses, includes the following steps: S1-4-1, center adjustment of the selected stereoscopic frame: after determining the offset vector T of the double-selected stereoscopic frame, the selected stereoscopic frame is translated to the next area to be cropped. The specific steps are as follows: (1) adjusting the double-selection stereo frame in translation according to the offset vector; (2) updating the current dual-selection stereo frame so that the dual-selection stereo frame just calculated becomes the current dual-selection stereo frame; (3) recalculate the center of the currently selected stereo frame to make it the new rotation center; S1-4-2, automatic adjustment of the selected stereoscopic frame posture: using formula V dif =V i+1 -V i The direction difference calculation of two adjacent selected stereo frames is realized, and the difference is used as the basic parameter for adjusting the selected stereo frame to obtain the complete transformation matrix TMatrix at the i+1th position in the selected stereo frame, and the posture of the selected stereo frame is adjusted by using TMatrix; S1-5, cutting and extraction of contact network facilities.

3. The fully automatic search and extraction method for contact network facility point cloud according to claim 2, characterized in that: The specific method of step S1-5 is: perform point-by-point distance accumulation calculation on the POS trajectory points, and when the accumulated distance is greater than or equal to the offset distance, crop the point cloud in the area of ​​the box by selecting the stereo box CBox to complete the cropping and extraction operations.

4. The fully automatic search and extraction method for contact network facility point cloud according to claim 2 is characterized in that: The specific method of step S2-2 is: first, based on the attention mechanism feature extraction method of ECA and CBAM, the important features in the point cloud are strengthened from the channel and spatial domains; then, a residual refinement structure based on the OCS initial classification result is introduced, which realizes the refinement of the contact network facility classification results through multi-scale receptive field feature extraction and fusion.

5. The fully automatic search and extraction method for contact network facility point cloud according to claim 2, characterized in that: The specific method of step S2-2 is: eliminate the T-shaped input feature transformation in the original structure of the MFF_A model, use the MLP module to extract point cloud features, and use the ECA channel attention mechanism to realize channel enhancement of the extracted features; then use the CBAM channel and spatial attention mechanism to perform the shallow features processed by ECA to enhance the shallow extracted features, and fuse the multiple layers of shallow features with the deep features; then use the ECA channel attention mechanism to process the fused features, and use the activation function to divide the point cloud into n categories of contact network facilities, input the initial classification results into the Refine structure, and finally output the refined classification results.

6. The fully automatic search and extraction method for contact network facility point cloud according to claim 5, characterized in that: The MFF_A model includes the following modules: 1) ECA module: The original ECA module is improved to global average pooling (GAP) and global maximum pooling (GMP) and used together: Where |t| odd Find the nearest odd number to t; γ and b are 2 and 1 respectively; After the shared weight MLP feature extraction, a multi-dimensional feature map is obtained, which is used as the input of the ECA module, and the input feature dimension is kept consistent with the output feature dimension to prevent the feature dimension from being reduced; the grouped convolution strategy is used to capture cross-channel interactions; for a given fully connected layer, the grouped convolution divides it into multiple groups and performs linear transformations independently in each group, and the interaction between each channel and its adjacent channels, so the weight calculation formula is: In the formula, Represents a set of adjacent channels; In this way, each channel attention module affects the K*C parameters, making all channels share the same learning parameters; 2) CBAM module: The direct jump connection adopts the jump connection based on CBAM to strengthen the important features from the two dimensions of channel and space. To this end, the input features pass through the channel and spatial attention mechanisms in turn, so that each branch can learn "What" and "Where" on the channel and spatial axes respectively, strengthening the effective transmission of important features in the network; 3) Pyramid pooling module: obtain feature tensors through convolution operations, then obtain feature tensors of different scales through global maximum pooling operations, then perform convolution operations on feature tensors of different scales to reduce their dimensions, and finally perform deconvolution operations to facilitate the subsequent superposition and fusion operations of these feature tensors; 4) Channel feature enhancement structure: On the one hand, the feature dimension is converted from the original 64 dimensions to 16 dimensions through the convolution layer, and the ECA module is introduced to strengthen the learning of channel features and improve the expression ability of this feature dimension in important features; then the feature enhancement results are superimposed with the feature extraction results of each scale in PPM, and the features of these four scales are superimposed and fused through concat connection to achieve the enhancement and fusion of multi-scale features; on the other hand, the average pooled feature tensor of the feature tensor (24, 4096, 1, 64) is obtained through GAP and superimposed with the multi-scale feature fusion tensor obtained in the previous step to achieve feature enhancement.

7. The fully automatic search and extraction method for contact network facility point cloud according to claim 6, characterized in that: The CBAM module includes a channel attention submodule and a spatial attention submodule. In the channel attention submodule, the spatial information of the feature map is first aggregated using average and maximum pooling operations on the input feature F to generate two different spatial context descriptors: and Represent the average pooling feature and the maximum pooling feature respectively; Then, both descriptors are fed into a shared MLP structure to generate channel attention feature maps Set the hidden activation size to Where r is the reduction ratio; after applying the shared network to each descriptor, the output feature vector is merged by element accumulation, and the channel attention is calculated as follows: Where σ represents the activation function, W0 and W1 refer to the weights of the MLP, and the condition W0 and W1 share weights between the input features and the subsequent ReLU activation function; The spatial attention submodule uses GAP and GMP to perform pooling operations on the channel enhancement feature F' and connects the two to generate an effective feature descriptor. On the connected feature description, a convolutional layer is used to generate a spatial attention feature map. And encode it to achieve feature emphasis or suppression; aggregate the channel information of the feature map through two pooling operations to generate two types of feature maps: and Each average and maximum pooling feature adopts the channel attention mechanism, and then connects and convolves them through a standard convolutional layer to generate a 2D spatial attention feature map. The spatial attention formula is shown as follows: In the formula, σ represents the activation function, f 7×7 Represents a 7×7 convolution kernel for convolution operation.

8. The fully automatic search and extraction method for contact network facility point cloud according to claim 7, characterized in that: The reconstruction of the contact network 3D model in S3 includes: 1) Model reconstruction of contact line and suspension string: First, the trajectory line segmented fitting algorithm is used to realize the straight line fitting and model reconstruction of the suspension string; secondly, the segmented fitting suspension string model is used to further segment the contact line point cloud, and the contact line between adjacent suspension strings is also fitted and model reconstructed using the segmented fitting algorithm; 2) Model reconstruction of the locator component: The locator adopts a rectangular tube, and a standard locator three-dimensional model is established according to the geometric parameters of the rectangular tube. The length direction vector of the established model is set to (1,0,0); the nearest point iterative registration algorithm is used to align the locator point cloud obtained by semantic segmentation with the standard locator model to achieve automatic reconstruction of the locator point cloud.

9. The fully automatic search and extraction method for contact network facility point cloud according to claim 8, characterized in that: The geometric parameter detection based on the contact network model in step S4 includes: 1) Detection of contact wire height and pull-out value: The detection result of the track centerline is the measurement baseline of the height and pull-out value. The height of the contact network is obtained by subtracting the elevation of the reconstructed contact wire suspension point and the elevation of the track centerline at the corresponding position; the pull-out value of the contact network in the straight section is measured as the distance from the center line of the line after the vertical projection of the contact wire, and is compared with the reconstructed contact wire model parameters and the center line of the line; the pull-out value of the contact network in the curved section is calculated according to the following formula: a=m+c,where: Where: a: contact network pull-out value; m: horizontal distance between the contact line at the positioning point and the center of the line; c: horizontal distance between the pantograph center at the positioning point and the center of the line; h: outer rail superelevation; H: contact line height; L: track gauge; 2) Locator tilt angle detection: The first endpoint P of the locator model is obtained by iteratively registering the locator model with the locator point cloud. s and the end point P e The spatial coordinate value of .

10. The fully automatic search and extraction method for contact network facility point cloud according to claim 9, characterized in that: In the positioner tilt angle detection, first the end point P e Project it onto the XOY plane to get the projection point P of the end point e '; Secondly, calculate the vector With vector The angle between the two is the slope angle of the positioner. The vector angle θ is calculated as follows:

Images