Ballast track ballast bed section detection method based on multiple sensors

Abstract

The invention discloses a ballast track bed section detection method based on multiple sensors, and belongs to the technical field of track safety detection. The invention aims to solve the problems of low efficiency and poor precision due to the adoption of a manual measurement mode in the conventional track ballast bed section detection. Comprising the steps that multiple laser scanning sensors are arranged at the bottom of a detection vehicle in the length direction of the ballast bed section, the view fields of the multiple laser scanning sensors cover the whole ballast bed section, and ballast bed section local images are correspondingly collected; local original point clouds are obtained through ballast bed section local image restoration, and then preprocessing is carried out to obtain preprocessed local point clouds; performing coordinate transformation on the plurality of preprocessed local point clouds to obtain a plurality of coordinate-transformed local point clouds under the same coordinate system; fusing the plurality of local point clouds after coordinate transformation to obtain complete ballast bed point cloud data; and according to the complete ballast bed point cloud data, calculating to obtain ballast bed section data. The method is used for detecting the section of the ballast bed of the ballast track.

Description

Technical Field

[0001] The invention relates to a multi-sensor based ballasted track bed section detection method, belonging to the technical field of track safety detection. Background Art

[0002] With the development of railways, the speed of trains continues to increase. People pay more and more attention to the safety issues in vehicle operation, and the requirements for automated maintenance and inspection of trains are also getting higher and higher. In addition to the vehicles themselves, tracks are also an important part of the railway system, and the efficiency and quality of maintenance need to be improved.

[0003] The ballast track bed is the foundation of the track frame, providing lateral and longitudinal resistance for the rail sleepers. The condition of the ballast bed directly affects the stability of the track and the safety of railway vehicles. If the ballast bed of the ballasted track is insufficient, it may cause hazards such as rail expansion, while too much ballast will cause passing vehicles to collide with the ballast and cause damage to the passing vehicles. Insufficient or excessive ballast will cause huge safety problems. The current ballast bed section detection or monitoring has the following problems:

[0004] 1) The existing roadbed section inspection or monitoring is carried out by manual measurement, which consumes a lot of manpower and material resources, has large errors, and is extremely inefficient.

[0005] 2) For manual measurement, the measurement results are highly correlated with the detection point locations and the detection locations are relatively few, which places high technical requirements on the detection personnel and makes it difficult to obtain the detection results accurately and quickly. Summary of the invention

[0006] Aiming at the problem that the existing track bed section detection adopts the manual measurement method with low efficiency and poor accuracy, the present invention provides a ballasted track bed section detection method based on multiple sensors.

[0007] A multi-sensor based ballasted track bed section detection method of the present invention comprises:

[0008] A plurality of laser scanning sensors are arranged at the bottom of the inspection vehicle along the length direction of the roadbed section, so that the field of view of the plurality of laser scanning sensors covers the entire roadbed section, and a local image of the roadbed section is correspondingly collected; a local original point cloud is obtained from the local image of the roadbed section, and then preprocessed to obtain a preprocessed local point cloud;

[0009] The coordinate transformation of multiple pre-processed local point clouds is used to obtain multiple local point clouds after coordinate transformation in the same coordinate system; the multiple local point clouds after coordinate transformation are fused to obtain complete track bed point cloud data;

[0010] The roadbed section data is calculated based on the complete roadbed point cloud data.

[0011] According to the multi-sensor based ballasted track bed section detection method of the present invention, there are four laser scanning sensors, namely sensor one, sensor two, sensor three and sensor four, wherein sensor two and sensor three are located in the middle, and the axial direction of the field of view is perpendicular to the roadbed section; sensor one and sensor four are located on both sides, and the axial direction of the field of view has an inclination angle with the roadbed section; the preprocessed local point clouds corresponding to sensor two and sensor three are in the same spatial coordinate system, and are in a different spatial coordinate system from the preprocessed local point clouds corresponding to sensor one and sensor four.

[0012] According to the multi-sensor based ballasted track bed section detection method of the present invention, conditional point cloud denoising is performed on the local original point cloud based on prior knowledge to obtain the pre-processed local point cloud.

[0013] According to the multi-sensor based ballasted track bed section detection method of the present invention, the method for conditional point cloud denoising based on prior knowledge includes:

[0014] A judgment model F is designed based on prior knowledge to filter out outliers in the local original point cloud to obtain a preliminary denoised local point cloud. A random sample consensus algorithm is then used for further denoising. M pixels are randomly selected from the preliminary denoised local point cloud as inner group points. The inner group model is calculated using the M pixels, and the number of inner group points is calculated based on the inner group model. Repeated iterations are performed to obtain the optimal inner group model and the preprocessed local point cloud.

[0015] According to the multi-sensor based ballasted track section detection method of the present invention, the method of obtaining four local point clouds after coordinate transformation in the same coordinate system by coordinate transformation of four pre-processed local point clouds includes:

[0016] The width direction of the ballast bed is defined as the x direction, the rail direction is defined as the y direction, and the height direction of the ballast bed is defined as the z direction, and a reference space coordinate system is established, wherein the pre-processed local point clouds of sensor two and sensor three are in the reference space coordinate system; the method for coordinate transformation of the pre-processed local point clouds of sensor one and sensor four is the same:

[0017] Assume that the coordinates of the local point cloud of sensor 1 after preprocessing in the space coordinate system of sensor 1 are P(x 0 ,y 0 ,z 0 ), the y coordinate system of the sensor-body space 0 The direction coincides with the y direction of the reference space coordinate system; the x direction of the sensor-body space coordinate system 0 Direction, z 0 The direction has an inclination angle θ with the x direction and z direction of the reference space coordinate system respectively; 0 ,y 0 ,z 0) around y 0 The axis rotates by an angle θ, and P(x 0 ,y 0 ,z 0 ) in the reference space coordinate system is P'(x',y',z'):

[0018]

[0019] Thus, the point cloud data transformation formula from the sensor-body space coordinate system to the reference space coordinate system is obtained:

[0020]

[0021] Thus, the local point cloud is obtained after four coordinate transformations.

[0022] According to the multi-sensor based ballasted track section detection method of the present invention, fusing the local point clouds after four coordinate transformations includes extracting features from the local point clouds after the coordinate transformations:

[0023] The local point cloud after coordinate transformation is input into 3D convolution layer 1, and then passes through 3D convolution layer 2, 3D convolution layer 3, 3D convolution layer 4 and 3D convolution layer 5. The output of 3D convolution layer 5 and the output of 3D convolution layer 4 are used as the input of 3D deconvolution layer 1, the output of 3D deconvolution layer 1 and the output of 3D convolution layer 3 are used as the input of 3D deconvolution layer 2, the output of 3D deconvolution layer 2 and the output of 3D convolution layer 2 are used as the input of 3D deconvolution layer 3, the output of 3D deconvolution layer 3 and the output of 3D convolution layer 1 are used as the input of 3D deconvolution layer 4, and the local point cloud features are output by 3D deconvolution layer 4;

[0024] The convolution parameters of 3D convolution layer 4 and 3D deconvolution layer 4 are the same, the convolution parameters of 3D convolution layer 3 and 3D deconvolution layer 3 are the same, the convolution parameters of 3D convolution layer 2 and 3D deconvolution layer 2 are the same, and the convolution parameters of 3D convolution layer 1 and 3D deconvolution layer 1 are the same;

[0025] Each 3D convolutional layer is configured with a batch normalization layer and an activation layer.

[0026] According to the multi-sensor based ballasted track section detection method of the present invention, fusing the local point clouds after the four coordinate transformations also includes key point matching of the extracted local point cloud features:

[0027] Assume that the two local point cloud features for key point matching are F C1 and F C2 , introduce the cross attention mechanism to F C1 and F C2 Perform deep information interaction to obtain the associated information interaction feature CrossAttention (FC1 ,F C2 ):

[0028]

[0029] Where Softmax represents the normalized exponential function, Q represents the cross-attention query vector, K represents the cross-attention key vector, V represents the cross-attention value vector, and d k is the dimension of the cross attention query vector;

[0030] Q=F C1 W Q ,

[0031] K=V=F C2 W K ,

[0032] W Q is the learnable projection matrix corresponding to the cross-attention query vector Q, W K is the learnable projection matrix corresponding to the cross-attention key vector K.

[0033] According to the multi-sensor based ballasted track section detection method of the present invention, for each associated information interaction feature CrossAttention (F C1 ,F C2 ) Use rigid transformer to generate multiple point cloud fusion hypotheses, assuming C 1 is the local point cloud feature F C1 The corresponding point cloud to be fused in C 2 is the local point cloud feature F C2 The corresponding point cloud to be fused in the 1 and the point cloud C to be fused 2 Form a set of point cloud fusion hypotheses; calculate C 1 and C 2 The chamfer distance d after fusion CD (C 1 ,C 2 )for:

[0034]

[0035] Where a and b are the point clouds C to be fused. 1 , C 2 The points in the point cloud C to be fused are m and n. 1 , C 2 The number of midpoints, ‖ba‖ represents the distance between point a and point b in space;

[0036] The fusion result with the smallest chamfer distance after fusion among the multiple point cloud fusion hypotheses is taken as the final fusion result;

[0037] The complete track bed point cloud data is obtained from all the final fusion results.

[0038] According to the multi-sensor based ballasted track bed section detection method of the present invention, the bed section data includes calculation of shoulder width and pile height, specifically:

[0039] The complete roadbed point cloud data is cut along the y direction to obtain multiple groups of roadbed profile data;

[0040]

[0041] Where w shoulder Indicates shoulder width, h shoulder Indicates stack height,

[0042] According to the prior knowledge, the coordinate x of point A on the inner side of the ballast shoulder in the reference space coordinate system is used. A Calculate the coordinate z of point A in the reference space coordinate system A As ballast shoulder height;

[0043] Traverse the point clouds in multiple groups of roadbed profile data to obtain the coordinates of the highest point (x B , z B ), and the coordinates of point C on the outside of the ballast shoulder at the same height as point A (x C ,z C ).

[0044] According to the multi-sensor based ballasted track bed section detection method of the present invention, the bed section data also includes slope height, slope length, slope ratio and ballast surplus and deficit, and the calculation method is:

[0045]

[0046] Where h slope is the slope height, l slope is the slope length, r slope is the slope ratio; (x D ,z D ) is the coordinate of the lowest point of the roadbed obtained by traversing the complete roadbed point cloud data;

[0047] Methods for calculating ballast profit and loss include:

[0048] First, calculate the average height E(Z) of the point cloud between rails and the discrete value D(Z) of the point cloud between rails:

[0049]

[0050] Where I is the number of points in the point cloud between rails, z i is the point height of the point cloud between rails;

[0051] The points of the point cloud between rails corresponding to the discrete value D(Z) in the set interval are shielded;

[0052] The linear equation f is established by using the undisturbed ballast point cloud data after removing the shielding points from the complete ballast point cloud data. 1 (x), f 2 (x), …, f N (x), where N is the number of points in the undisturbed ballast point cloud data, and the ballast surplus or deficit S is calculated as:

[0053]

[0054] Where z j is the height of the jth point in the interference track bed point cloud data.

[0055] Beneficial effects of the present invention: The method of the present invention uses four high-precision laser scanning sensors at the bottom of the high-speed inspection vehicle to obtain the data of the ballast track bed; converts the original data of the ballast bed into point cloud data, and performs denoising on the point cloud data; performs coordinate transformation on the point cloud so that the point cloud data is in the same coordinate system, and uses point cloud matching to merge the four point cloud data into a complete ballast bed point cloud data; finally, the point cloud data is calculated to obtain the ballast bed section data such as shoulder height, slope ratio, ballast surplus and deficit. According to the calculation results, an alarm can be issued when an abnormality is determined.

[0056] The method of the present invention introduces automatic identification technology into the roadbed section detection, which can realize automatic fault identification, effectively save labor costs, and improve work quality and efficiency; multiple sensors are used to obtain complete roadbed splicing data, which improves the detection accuracy. Compared with a single laser scanning sensor to complete the task of scanning the entire roadbed, the accuracy of the detection results obtained by increasing the measurement distance is greatly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 1 is a schematic diagram of the ballast track section data collection of the multi-sensor based ballast track section detection method of the present invention; in the figure, 1 represents the laser scanning sensor, 2 represents the ballast track, and 3 represents the rail;

[0058] Figure 2 It is the original image of the roadbed section formed by stitching together the local images of the roadbed section;

[0059] Figure 3 It is the local original point cloud connection map corresponding to the images collected by sensor one and sensor four;

[0060] Figure 4 It is the local original point cloud connection map corresponding to the images collected by sensor two and sensor three;

[0061] Figure 5 yes Figure 3The corresponding preprocessed local point cloud mosaic;

[0062] Figure 6 yes Figure 4 The corresponding preprocessed local point cloud mosaic;

[0063] Figure 7 It is a network structure diagram for feature extraction of local point cloud after coordinate transformation;

[0064] Figure 8 It is a roadbed cross-section curve diagram drawn based on the complete roadbed point cloud data;

[0065] Fig. 9 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0066] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0068] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0069] Specific implementation method 1. Combination Figures 1 to 9 As shown, the present invention provides a multi-sensor based ballasted track bed section detection method, comprising:

[0070] A plurality of laser scanning sensors are arranged at the bottom of the inspection vehicle along the length direction of the roadbed section, so that the field of view of the plurality of laser scanning sensors covers the entire roadbed section, and a local image of the roadbed section is correspondingly collected; a local original point cloud is obtained from the local image of the roadbed section, and then preprocessed to obtain a preprocessed local point cloud;

[0071] The coordinate transformation of multiple pre-processed local point clouds is used to obtain multiple local point clouds after coordinate transformation in the same coordinate system; the multiple local point clouds after coordinate transformation are fused to obtain complete track bed point cloud data;

[0072] The roadbed section data is calculated based on the complete roadbed point cloud data.

[0073] The local image of the roadbed section is a 16-bit tiff image. The position in the image represents the position of the point cloud, and the pixel value in the image represents the point cloud height value at that position. The point cloud data can be restored using the local image of the roadbed section. Figure 1 shown.

[0074] Further, combined with Figure 1 As shown, there are four laser scanning sensors, namely sensor one, sensor two, sensor three and sensor four, wherein sensor two and sensor three are located in the middle, and the axial direction of the field of view is perpendicular to the roadbed section; sensor one and sensor four are located on both sides, and the axial direction of the field of view has an inclination angle with the roadbed section; the preprocessed local point clouds corresponding to sensor two and sensor three are in the same spatial coordinate system, and are in a different spatial coordinate system from the preprocessed local point clouds corresponding to sensor one and sensor four.

[0075] In this implementation, conditional point cloud denoising is performed on the local original point cloud based on prior knowledge to obtain the pre-processed local point cloud. The original point cloud obtained by the sensor scanning often contains noise points, also called outliers and miscellaneous points. These noise points may be caused by various reasons, such as scanning dust, reflection on the surface of the target to be scanned, etc. This implementation performs conditional point cloud denoising based on prior knowledge, which can better filter the noise points in the point cloud.

[0076] The distribution of the track bed point cloud data should be regularly within a certain range, while the noise data, especially the noise caused by foreign objects, is generally located outside the track bed data. Therefore, this embodiment uses prior knowledge to perform conditional screening before point cloud denoising, which can better remove large continuous noise groups caused by external factors.

[0077] Methods for conditional point cloud denoising based on prior knowledge include:

[0078] Design a judgment model F based on prior knowledge, traverse all points in the original point cloud, and filter out the local original point cloud C p The outliers in the local point cloud are used to obtain the initial denoising local point cloud; the RANSAC algorithm is used for further denoising, and M pixels are randomly selected from the local point cloud after the initial denoising as the inner group points, and the inner group model is calculated using the M pixels. The remaining points are then brought into the inner group model, and the number of inner group points is further calculated based on the inner group model; repeated iterations are performed to obtain the optimal inner group model, and the preprocessed local point cloud is obtained. Figures 3 to 6 shown.

[0079] The judgment model F can be designed to make the output of the point cloud data whose input exceeds the track bed data 0, and the corresponding points are deleted at this time.

[0080] Further denoising methods can also be selected according to needs, such as radius filtering, statistical filtering, Euclidean clustering, etc.

[0081] Furthermore, the method of obtaining four local point clouds after coordinate transformation in the same coordinate system by coordinate transformation of the four pre-processed local point clouds includes:

[0082] Sensor 1 and Sensor 4 are not placed horizontally in order to ensure that the entire roadbed can be scanned.

[0083] The width direction of the ballast bed is defined as the x direction, the rail direction is defined as the y direction, and the height direction of the ballast bed is defined as the z direction, and a reference space coordinate system is established, wherein the pre-processed local point clouds of sensor two and sensor three are in the reference space coordinate system; the method for coordinate transformation of the pre-processed local point clouds of sensor one and sensor four is the same:

[0084] Assume that the coordinates of the local point cloud of sensor 1 after preprocessing in the space coordinate system of sensor 1 are P(x 0 ,y 0 ,z 0 ), the y coordinate system of the sensor-body space 0 The direction coincides with the y direction of the reference space coordinate system; the x direction of the sensor-body space coordinate system 0 Direction, z 0 The direction has an inclination angle θ with the x direction and z direction of the reference space coordinate system respectively; 0 ,y 0 ,z 0 ) around y 0 The axis rotates by an angle θ, and P(x 0 ,y 0 ,z 0 ) in the reference space coordinate system is P'(x',y',z'):

[0085]

[0086] Thus, the point cloud data transformation formula from the sensor-body space coordinate system to the reference space coordinate system is obtained:

[0087]

[0088] Thus, the local point cloud is obtained after four coordinate transformations.

[0089] After the coordinate transformation, the four local point clouds are all in the reference space coordinate system, which meets the conditions for point cloud fusion.

[0090] Point cloud data fusion: Among the point cloud registration and fusion methods, common methods include the iterative closest point (ICP) method and the normal distribution transform (NDT) method. However, these traditional algorithms are often slow and have low accuracy. In this embodiment, a point cloud fusion method based on deep learning is used, and a general multi-channel embedding framework is adopted to register and fuse point clouds from different data sources and partially overlap. The general multi-channel embedding method includes three components: feature extraction component, key point matching component, and hypothesis selection component.

[0091] Combination Figure 7 As shown, the feature extraction component is used to fuse the four local point clouds after coordinate transformation, including feature extraction of the local point clouds after coordinate transformation:

[0092] The local point cloud after coordinate transformation is input into 3D convolution layer 1, and then passes through 3D convolution layer 2, 3D convolution layer 3, 3D convolution layer 4 and 3D convolution layer 5. The output of 3D convolution layer 5 and the output of 3D convolution layer 4 are used as the input of 3D deconvolution layer 1, the output of 3D deconvolution layer 1 and the output of 3D convolution layer 3 are used as the input of 3D deconvolution layer 2, the output of 3D deconvolution layer 2 and the output of 3D convolution layer 2 are used as the input of 3D deconvolution layer 3, the output of 3D deconvolution layer 3 and the output of 3D convolution layer 1 are used as the input of 3D deconvolution layer 4, and the local point cloud features are output by 3D deconvolution layer 4;

[0093] The convolution parameters of 3D convolution layer 4 and 3D deconvolution layer 4 are the same, the convolution parameters of 3D convolution layer 3 and 3D deconvolution layer 3 are the same, the convolution parameters of 3D convolution layer 2 and 3D deconvolution layer 2 are the same, and the convolution parameters of 3D convolution layer 1 and 3D deconvolution layer 1 are the same;

[0094] Each 3D convolutional layer is configured with a batch normalization layer and an activation layer.

[0095] This implementation method extracts point cloud features based on deep learning, laying the foundation for more accurate data fusion.

[0096] Figure 7 Medium k 1 to k 4 is the 3D convolution kernel size of the corresponding convolution layer, S 1 To S 4 is the step size of the corresponding convolutional layer, d out1 to d out4is the output of the corresponding convolutional layer. ⊕ indicates a cascade operation. The feature extraction component follows the structure of the basic neural network Unet, while using sparse 3D convolution and adding jump connections. At the same time, each convolution module in the network is connected to the batch normalization (Batch Normalization, BN) layer and the activation (ReLU) layer. The feature extraction component assigns a feature vector to each point in the input point cloud. These features encapsulate global information and local information.

[0097] Furthermore, the fusion of the local point clouds after the four coordinate transformations also includes using a key point matching component to perform key point matching on the extracted local point cloud features:

[0098] Assume that the two local point cloud features for key point matching are F C1 and F C2 , at this time F C1 and F C2 are independent, and a cross-attention mechanism is introduced in the key point matching component to C1 and F C2 Perform deep information interaction to associate the features of each point with another set of point clouds, and obtain the associated information interaction feature CrossAttention (F C1 ,F C2 ):

[0099]

[0100] Where Softmax represents the normalized exponential function, Q represents the cross-attention query vector, K represents the cross-attention key vector, V represents the cross-attention value vector, and d k is the dimension of the cross attention query vector;

[0101] Q=F C1 W Q ,

[0102] K=V=F c2 W K ,

[0103] W Q is the learnable projection matrix corresponding to the cross-attention query vector Q, W K is the learnable projection matrix corresponding to the cross-attention key vector K.

[0104] By comparing the key points from the two point clouds in pairs, it is determined whether they are located in the overlapping area, and then the overlapping area and key point matching information of the two point clouds are obtained.

[0105] The hypothesis selection component is used to select each associated information interaction feature CrossAttention (F C1,F C2 ) Use rigid transformer to generate multiple point cloud fusion hypotheses, assuming C 1 is the local point cloud feature F C1 The corresponding point cloud to be fused in C 2 is the local point cloud feature F C2 The corresponding point cloud to be fused in the 1 and the point cloud C to be fused 2 Form a set of point cloud fusion hypotheses; calculate C 1 and C 2 The chamfer distance d after fusion CD (C 1 ,C 2 )for:

[0106]

[0107] Where a and b are the point clouds C to be fused. 1 , C 2 The points in the point cloud C to be fused are m and n. 1 , C 2 The number of midpoints, ‖ba‖ represents the distance between point a and point b in space;

[0108] The hypothesis selection component selects a set of fusion results with the smallest chamfer distance after fusion among multiple sets of point cloud fusion hypotheses as the final fusion result;

[0109] The complete track bed point cloud data is obtained from all the final fusion results.

[0110] Furthermore, the ballast section data calculation and section generation: the ballast section data includes the calculation of shoulder width and pile height, specifically:

[0111] Combination Figure 1 As shown, after the fused point cloud data is obtained, the complete roadbed point cloud data is cut along the y direction to obtain multiple groups of roadbed profile data; data calculation is performed and cross-sectional images are generated;

[0112]

[0113] Where w ahoulder Indicates shoulder width, h shoulder Indicates stack height,

[0114] According to the prior knowledge, the coordinate x of point A on the inner side of the ballast shoulder (edge ​​of the sleeper) in the reference space coordinate system is used. A Calculate the coordinate z of point A in the reference space coordinate system A As ballast shoulder height;

[0115] Traverse the point clouds in multiple groups of roadbed profile data to obtain the coordinates of point B (xB , z B ), and the coordinates of point C on the outside of the ballast shoulder at the same height as point A (x C ,z C ).

[0116] The ballast section data also includes slope height, slope length, slope ratio and ballast surplus and deficit, and the calculation method is:

[0117]

[0118] Where h slope is the slope height, l slope is the slope length, r slope is the slope ratio; (x D ,z D ) is the coordinate of point D, the lowest point of the roadbed, obtained by traversing the complete roadbed point cloud data;

[0119] Methods for calculating ballast profit and loss include:

[0120] Ballast surplus or deficit is the difference between the actual ballast volume and the standard ballast volume per unit length. In the calculation of ballast surplus or deficit, the biggest interference source is the sleepers in the trackbed. However, the point cloud distribution characteristics of sleepers and ballast are significantly different. Therefore, the discrete value of the point cloud in the trackbed is used to shield the interference of sleepers.

[0121] First, calculate the average height E(Z) of the point cloud between rails and the discrete value D(Z) of the point cloud between rails:

[0122]

[0123] Where I is the number of points in the point cloud between rails, z i is the point height of the point cloud between rails;

[0124] Shielding the points of the rail point cloud corresponding to the discrete value D(Z) in the set interval; the set interval is an interval outside the distribution range of the ballast point cloud set according to the empirical value;

[0125] The linear equation f is established by using the undisturbed ballast point cloud data after removing the shielding points from the complete ballast point cloud data. 1 (x), f 2 (x), …, f N (x), where N is the number of points in the undisturbed ballast point cloud data, and the ballast surplus or deficit S is calculated as:

[0126]

[0127] Where z j is the height of the jth point in the interference track bed point cloud data.

[0128] Finally, the cross-section curve can be drawn by traversing the complete track bed point cloud data, such as Figure 8 shown.

[0129] Data analysis and alarm: After obtaining the corresponding data of the roadbed section, the calculated data is compared with the standard data. If the difference with the standard data is greater than a certain value, an alarm is issued. The overall process is as follows: Fig. 9 shown.

[0130] Although the present invention is described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the present invention. It should therefore be understood that many modifications may be made to the exemplary embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the various dependent claims and features described herein may be combined in a manner different from that described in the original claims. It should also be understood that the features described in conjunction with a single embodiment may be used in other described embodiments.

Claims

1. A multi-sensor based ballasted track bed section detection method, characterized in that: include: A plurality of laser scanning sensors are arranged at the bottom of the inspection vehicle along the length direction of the roadbed section, so that the field of view of the plurality of laser scanning sensors covers the entire roadbed section, and a local image of the roadbed section is correspondingly collected; a local original point cloud is obtained from the local image of the roadbed section, and then preprocessed to obtain a preprocessed local point cloud; The coordinate transformation of multiple pre-processed local point clouds is used to obtain multiple local point clouds after coordinate transformation in the same coordinate system; the multiple local point clouds after coordinate transformation are fused to obtain complete track bed point cloud data; The roadbed section data is calculated based on the complete roadbed point cloud data.

2. The multi-sensor based ballasted track bed section detection method according to claim 1, characterized in that: There are four laser scanning sensors, namely sensor one, sensor two, sensor three and sensor four, among which sensor two and sensor three are located in the middle, and the axial direction of the field of view is perpendicular to the roadbed section; sensor one and sensor four are located on both sides, and the axial direction of the field of view has an inclination angle with the roadbed section; the preprocessed local point clouds corresponding to sensor two and sensor three are in the same spatial coordinate system, and are in a different spatial coordinate system from the preprocessed local point clouds corresponding to sensor one and sensor four.

3. The multi-sensor based ballasted track bed section detection method according to claim 2 is characterized in that: Conditional point cloud denoising is performed on the local original point cloud based on prior knowledge to obtain the preprocessed local point cloud.

4. The multi-sensor based ballasted track bed section detection method according to claim 3 is characterized in that: Methods for conditional point cloud denoising based on prior knowledge include: A judgment model F is designed based on prior knowledge to filter out outliers in the local original point cloud to obtain a preliminary denoised local point cloud. A random sample consensus algorithm is then used for further denoising. M pixels are randomly selected from the preliminary denoised local point cloud as inner group points. The inner group model is calculated using the M pixels, and the number of inner group points is calculated based on the inner group model. Repeated iterations are performed to obtain the optimal inner group model and the preprocessed local point cloud.

5. The multi-sensor based ballasted track bed section detection method according to claim 4, characterized in that: The method of obtaining four local point clouds after coordinate transformation in the same coordinate system by coordinate transformation of four preprocessed local point clouds includes: The width direction of the ballast bed is defined as the x direction, the rail direction is defined as the y direction, and the height direction of the ballast bed is defined as the z direction, and a reference space coordinate system is established, wherein the pre-processed local point clouds of sensor two and sensor three are in the reference space coordinate system; the method for coordinate transformation of the pre-processed local point clouds of sensor one and sensor four is the same: Assume that the coordinates of the preprocessed local point cloud of sensor 1 in the sensor 1 body space coordinate system are P(x0, y0, z0), and the y0 direction of the sensor 1 body space coordinate system coincides with the y direction of the reference space coordinate system; the x0 direction and the z0 direction of the sensor 1 body space coordinate system have an inclination angle θ with the x direction and the z direction of the reference space coordinate system respectively; rotate P(x0, y0, z0) around the y0 axis by an angle θ, and obtain the coordinates of P(x0, y0, z0) in the reference space coordinate system as P'(x', y', z'): x′=z′sinθ+x0cosθ y′=y0, z′=z0cosθ-x0sinθ Thus, the point cloud data transformation formula from the sensor-body space coordinate system to the reference space coordinate system is obtained: Thus, the local point cloud is obtained after four coordinate transformations.

6. The multi-sensor based ballasted track bed section detection method according to claim 5, characterized in that: The fusion of the four local point clouds after coordinate transformation includes feature extraction of the local point clouds after coordinate transformation: The local point cloud after coordinate transformation is input into 3D convolution layer 1, and then passes through 3D convolution layer 2, 3D convolution layer 3, 3D convolution layer 4 and 3D convolution layer 5. The output of 3D convolution layer 5 and the output of 3D convolution layer 4 are used as the input of 3D deconvolution layer 1, the output of 3D deconvolution layer 1 and the output of 3D convolution layer 3 are used as the input of 3D deconvolution layer 2, the output of 3D deconvolution layer 2 and the output of 3D convolution layer 2 are used as the input of 3D deconvolution layer 3, the output of 3D deconvolution layer 3 and the output of 3D convolution layer 1 are used as the input of 3D deconvolution layer 4, and the local point cloud features are output by 3D deconvolution layer 4; The convolution parameters of 3D convolution layer 4 and 3D deconvolution layer 4 are the same, the convolution parameters of 3D convolution layer 3 and 3D deconvolution layer 3 are the same, the convolution parameters of 3D convolution layer 2 and 3D deconvolution layer 2 are the same, and the convolution parameters of 3D convolution layer 1 and 3D deconvolution layer 1 are the same; Each 3D convolutional layer is configured with a batch normalization layer and an activation layer.

7. The multi-sensor based ballasted track bed section detection method according to claim 6, characterized in that: The fusion of the local point clouds after the four coordinate transformations also includes key point matching of the extracted local point cloud features: Assume that the two local point cloud features for key point matching are F C1 and F C2 , introduce the cross attention mechanism to F C1 and F C2 Perform deep information interaction to obtain the associated information interaction feature CrossAttention (F C1 ,F C2 ): Where Softmax represents the normalized exponential function, Q represents the cross-attention query vector, K represents the cross-attention key vector, V represents the cross-attention value vector, and d k is the dimension of the cross attention query vector; Q=F C1 W Q , K=V=F C2 W K , W Q is the learnable projection matrix corresponding to the cross-attention query vector Q, W K is the learnable projection matrix corresponding to the cross-attention key vector K.

8. The multi-sensor based ballasted track bed section detection method according to claim 7, characterized in that: For each associated information interaction feature CrossAttention (F c1 ,F C2 ) Use rigid transformer to generate multiple sets of point cloud fusion hypotheses, assuming C1 is the local point cloud feature F C1 The corresponding point cloud to be fused, C2 is the local point cloud feature F C2 The corresponding point cloud to be fused in the point cloud to be fused, the point cloud to be fused C1 and the point cloud to be fused C2 form a set of point cloud fusion hypotheses; calculate the chamfer distance d after fusion of C1 and C2 CD (C1,C2) is: Where a and b are points in the point clouds C1 and C2 to be fused, respectively; m and n are the number of points in the point clouds C1 and C2 to be fused, respectively; ‖ba‖ represents the distance between point a and point b in space; The fusion result with the smallest chamfer distance after fusion among the multiple point cloud fusion hypotheses is taken as the final fusion result; The complete trackbed point cloud data is obtained from all the final fusion results.

9. The multi-sensor based ballasted track bed section detection method according to claim 8, characterized in that: The ballast section data includes the calculation of shoulder width and pile height, specifically: The complete roadbed point cloud data is cut along the y direction to obtain multiple groups of roadbed profile data; w shoulder =x C -x A h shoulder =z B -z A , Where w shoulder Indicates shoulder width, h shoulder Indicates stack height, According to the prior knowledge, the coordinate x of point A on the inner side of the ballast shoulder in the reference space coordinate system is used. A Calculate the coordinate z of point A in the reference space coordinate system A As ballast shoulder height; Traverse the point clouds in multiple groups of roadbed profile data to obtain the coordinates of the highest point (x B , z B ), and the coordinates of point C on the outside of the ballast shoulder at the same height as point A (x C ,z C ).

10. The multi-sensor based ballasted track bed section detection method according to claim 9, characterized in that: The ballast section data also includes slope height, slope length, slope ratio and ballast surplus and deficit, and the calculation method is: h slope =z C -z D r slope ＝h slope / l slope Where h slope is the slope height, l slope is the slope length, r slope is the slope ratio; (x D ,z D ) is the coordinate of the lowest point of the roadbed obtained by traversing the complete roadbed point cloud data; Methods for calculating ballast profit and loss include: First, calculate the average height E(Z) of the point cloud between rails and the discrete value D(Z) of the point cloud between rails: Where I is the number of points in the point cloud between rails, z i is the point height of the point cloud between rails; The points of the point cloud between rails corresponding to the discrete value D(Z) in the set interval are shielded; The linear equations f1(x), f2(x), ..., f are established by using the undisturbed ballast point cloud data after removing the shielding points from the complete ballast point cloud data. N (x), where N is the number of points in the undisturbed ballast point cloud data, and the ballast surplus or deficit S is calculated as: Where z j is the height of the jth point in the interference track bed point cloud data.

Images