Subway shield tunnel point cloud filtering method combining offset features and semantic segmentation

By combining offset features and semantic segmentation, a point cloud filtering method for subway shield tunnels was developed, which solved the problem of noise point removal in complex tunnel environments. This method achieves efficient and accurate extraction of tunnel wall point clouds, ensuring the accuracy and safety of subway tunnel deformation monitoring.

CN118587398BActive Publication Date: 2024-11-15CHINA UNIV OF MINING & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410951437.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-16
Publication Date
2024-11-15
Estimated Expiration
2044-07-16

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively remove noise points in complex subway tunnel environments, affecting the accuracy and reliability of point cloud data and leading to inaccurate monitoring of subway tunnel deformation.

Method used

A point cloud filtering method for subway shield tunnels combining offset features and semantic segmentation is proposed. By constructing a radial offset feature module for the tunnel point cloud cross section, local translation alignment and local normalization are performed to generate multi-channel features, and a semantic segmentation network is used to extract the tunnel wall point cloud.

Benefits of technology

It significantly improves noise point filtering efficiency and filtering accuracy, enhances system adaptability, and ensures the accuracy and safety of subway tunnel deformation monitoring.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118587398B_ABST
    Figure CN118587398B_ABST
Patent Text Reader

Abstract

A subway shield tunnel point cloud filtering method combining offset features and semantic segmentation comprises the following steps: constructing a tunnel point cloud cross-section radial offset feature extraction module, and extracting the cross-section radial offset feature #imgabs0# of the tunnel point cloud relative to the cross-section center of the same ring segment; performing local translation alignment and local normalization processing on the point cloud according to the extracted cross-section radial offset feature #imgabs1# to obtain a local translation feature #imgabs2# and a local normalization feature #imgabs3#; performing weighted fusion on the cross-section radial offset feature #imgabs4#, the local translation feature #imgabs5# and the local normalization feature #imgabs6# to generate a multi-channel feature #imgabs7#; sending the multi-channel feature #imgabs8# into a #imgabs9# semantic segmentation network for training, and extracting the tunnel wall point cloud according to the classification result. The present invention combines the semantic segmentation network with the ellipse fitting filtering method to provide higher filtering accuracy and robustness while maintaining computational efficiency, effectively removing the point cloud of ancillary facilities that are close to the tunnel lining structure, making the safety monitoring and maintenance of subway tunnels more efficient and reliable.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a subway shield tunnel point cloud filtering method combining offset features and semantic segmentation, and belongs to the technical field of subway tunnel point cloud filtering. Background Art

[0002] Due to the complex geological environment in which they are located, subway tunnels are easily affected by natural factors such as geological movements, groundwater level changes, and earthquakes. In addition, human activities such as ground construction and vibrations from heavy vehicles will also have an impact on the tunnel structure. These factors will cause tunnel deformation. When the deformation exceeds a certain limit, it will cause cracks in the tunnel wall and deformation of the track, which will directly threaten the operational safety of the subway. Therefore, regular deformation monitoring and timely maintenance and repair of subway tunnels are the key to ensuring the safe operation of the subway.

[0003] Compared with the traditional total station detection method, laser radar has been widely used in the field of precision engineering measurement due to its advantages of high precision, fast speed and easy operation. Laser radar can obtain high-density point cloud data on the tunnel surface by emitting laser beams and receiving reflected signals. These point cloud data can accurately reflect the three-dimensional shape of the tunnel and provide rich data support for subsequent deformation analysis. However, the internal environment of the tunnel is complex. In addition to the tunnel wall itself, there are also a large number of ancillary facilities such as brackets, cables, power supply and distribution equipment, signal equipment, lighting equipment, etc. These facilities will generate a large number of non-tunnel wall points, namely noise points, during the laser radar scanning process, which are mixed in the point cloud data. The noise points interfere with the subsequent deformation analysis and tunnel modeling, affecting the accuracy and reliability of the detection. Therefore, how to remove noise points to the maximum extent while retaining the point cloud data of the inner wall of the tunnel has become a key issue that needs to be solved urgently.

[0004] Noise filtering methods mainly include statistical filtering, radius filtering and bilateral filtering. These methods usually identify and remove abnormal points by analyzing the neighborhood characteristics of each point in the point cloud. However, there are limitations when processing the data of ancillary facilities in the tunnel, because the facilities usually have the characteristics of continuity and uniform density, so noise filtering cannot effectively identify and remove them. Based on this, the prior art has proposed a variety of point cloud filtering methods based on advanced algorithms and technologies, including random sampling consistent multi-model fitting, ellipse fitting combining least squares method and Lagrange multiplier method, improved bilateral filtering algorithm, point cloud filtering method based on elliptical cylinder model, and cloth simulation filtering. Although these methods have their own advantages, they still have limitations in complex tunnel environments. Therefore, how to more effectively filter out noise points in complex environments and retain the real tunnel inner wall point cloud is still a problem that needs further research and optimization. Summary of the invention

[0005] In view of the problems existing in the above-mentioned prior art, the present invention provides a subway shield tunnel point cloud filtering method combining offset features and semantic segmentation. The method can separate the non-inner wall point cloud and the inner wall point cloud in the tunnel, more effectively filter out noise points, improve the accuracy of subway tunnel deformation monitoring, and ensure safe operation of the subway.

[0006] In order to achieve the above object, the present invention provides a subway shield tunnel point cloud filtering method combining offset features and semantic segmentation, comprising the following steps:

[0007] S1: Construct a tunnel point cloud section radial offset feature extraction module and extract the section radial offset features of the tunnel point cloud relative to the center of the same ring segment section ;

[0008] S2: Based on the extracted cross-section radial offset feature Perform local translation alignment and local normalization on the point cloud to obtain local translation features and local normalized features ;

[0009] S3: Offset the section radially , local translation feature With local normalized features Perform weighted fusion to generate multi-channel features ;

[0010] S4: Multi-channel features Feed Trained in the semantic segmentation network, the tunnel wall point cloud is extracted based on the classification results.

[0011] Furthermore, the construction process of the tunnel point cloud section radial offset feature extraction module in step S1 is as follows:

[0012] S1-1: Calculate the center coordinates of the cross section where each point is located ;

[0013] S1-2: Calculate the radial distance between each point and the center coordinate of its cross section ;

[0014] S1-3: Calculate the radial distance of each point Difference from the section radius , and obtain the radial offset characteristics of each point in the tunnel section .

[0015] Furthermore, the specific process of step S3 is: according to the preset weight, the cross-section radial offset feature , local translation feature With local normalization Features are linearly combined to generate multi-channel features .

[0016] Furthermore, the process of step S4 is: use the training model to perform semantic segmentation on the point cloud and generate a label for each point ; Filter non-tunnel wall point clouds according to classification labels and extract tunnel wall point clouds.

[0017] Furthermore, in step S1-1, the center coordinates of the cross section where each point is located are calculated. The steps are:

[0018] S1-1.1: Project the tunnel point cloud along the elliptical cylinder and unfold it, extract the tunnel bolt hole point cloud according to the distance and normal vector constraints, and segment the tunnel point cloud ring by ring according to the bolt hole point cloud; The process of projecting the tunnel point cloud onto the elliptical cylinder and unfolding it is as follows: Assume the fitting cylinder equation is: , is a point on the cylinder axis, is the axis direction vector, a point in the tunnel point cloud The corresponding cylinder axis point The relationship is: ;

[0019] The direction vector from point P to point A is: ;

[0020] Project the point P Translate along the projection direction to the cylindrical surface: , r is the radius of the cylinder;

[0021] The coordinates after unfolding the cylindrical surface are obtained by the following formula: ;

[0022] In the formula, are the coordinates of a point on the cylindrical surface, is the coordinate of the point after expansion, R is the design radius of the shield tunnel, is the angle between the straight line from the center of the section to the point and the vertical line;

[0023] S1-1.2: Extract the central axis L of each ring and rotate the point cloud as a whole until the central axis is parallel to the Y direction; extract the coordinates of the center point of the central section based on the least squares circle fitting , get each point Coordinates of the center of the section ; The central axis L is expressed as follows: ;

[0024] in, is the equation of the line on the YOZ plane, is the equation of the line on the XOY plane;

[0025] ;

[0026] In the formula, is the coordinate of a point in the segment ring, and n is the number of point clouds in the segment ring;

[0027] The center coordinates of the cross section of each ring center point are extracted by least squares circle fitting , so that the objective function is minimized, the objective function is expressed as: ;

[0028] In the formula, is the coordinate of a point on the cross section of each ring center, is the fitting radius of the cross-section point cloud; therefore, any point of each ring The coordinates of the center point of the cross section are obtained as ;

[0029] In step S1-2, the radial distance between each point and the center coordinate of its cross section is calculated. The formula is: ;

[0030] In step S1-3, the radial distance between each point and the center coordinate of its cross section is calculated. The formula is: .

[0031] Furthermore, the process of performing local translation alignment and local normalization processing on the point cloud in step S2 is as follows: according to the cross-sectional radial offset feature , translate the point cloud along the tunnel's forward direction and elevation direction to generate local translation features ; Perform minimum-maximum local normalization processing on the point cloud in the X, Y, and Z directions of the tunnel to generate local normalized features ; Among them, the local translation feature The calculation formula is: ;

[0032] In the formula, is the coordinate after translation;

[0033] Local normalized features The calculation formula is: ;

[0034] In the formula, is the local normalized coordinate, It is the extreme value of the point cloud coordinates in three directions within each segment ring.

[0035] Furthermore, in step S3, the multi-channel feature The calculation formula is: ;

[0036] In the formula, Section radial offset features , local translation features and local normalized features The weight of .

[0037] S4-1: Multi-channel features Input to Use the trained model to perform semantic segmentation on the point cloud and generate a label for each point ;

[0038] S4-2: Fusion of multi-channel features Local features after hierarchical feature extraction Generate global eigenvector , the global eigenvector Input to the fully connected layer and use multi-layer perceptron to analyze the global features Processed and passed The activation function generates the classification probability of each point , the formula is as follows: ;

[0039] In the formula, Yes The probability distribution of semantic labels;

[0040] S4-3: Train the model by minimizing the cross entropy loss function. The cross entropy loss function is as follows: ;

[0041] In the formula, Yes The real label, is the probability of predicting the label, N refers to the total number of points involved in the loss calculation, and C is the number of categories;

[0042] S4-4: Perform semantic segmentation on the new point cloud based on the trained model and generate the label of each point. The formula is as follows: ;

[0043] In the formula, are the trained network parameters, is the predicted semantic label set, PY is the cross-sectional radial offset feature extraction module, for Network model;

[0044] S4-5: Based on the predicted label , filter out points classified as tunnel walls: ;

[0045] In the formula, Yes The predicted label of .

[0046] The present invention generates more accurate multi-channel features by extracting the cross-sectional radial offset features of the tunnel point cloud, and performing local translation alignment and local normalization processing, which not only improves the filtering efficiency, but also significantly enhances the system's adaptability to complex tunnel environments; by combining the semantic segmentation network with the ellipse fitting filtering method, while maintaining computational efficiency, it provides higher filtering accuracy and robustness, effectively removes the point clouds of ancillary facilities that are close to the tunnel lining structure, and makes the safety monitoring and maintenance of subway tunnels more efficient and reliable. The present invention realizes the separation of non-inner wall point clouds and inner wall point clouds in the tunnel, more effectively filters out noise points, improves the accuracy of deformation monitoring of subway tunnels, and ensures the safe operation of subways. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of the method of the present invention;

[0048] Figure 2 It is a schematic diagram of the process of segmenting the tunnel point cloud ring by ring according to the present invention;

[0049] Figure 3 It is a schematic diagram of the rotation of a full ring point cloud of the present invention;

[0050] Figure 4 is a schematic diagram of tunnel wall point cloud extraction according to the present invention;

[0051] Figure 5 This is a comparison chart of the accuracy results of the present invention and the existing PointNet++;

[0052] Figure 6 This is a comparison chart of the loss convergence results of the present invention and the existing PointNet++;

[0053] Figure 7 2 is a comparison diagram of noise point recognition results, wherein Figure (a) is the noise point recognition result diagram of the existing PointNet++, and Figure (b) is the noise point recognition result diagram of the present invention. DETAILED DESCRIPTION

[0054] The present invention will be further described below in conjunction with the accompanying drawings.

[0055] like Figure 1 As shown, a subway shield tunnel point cloud filtering method combining offset features and semantic segmentation includes the following steps:

[0056] S1: Construct a tunnel point cloud section radial offset feature extraction module and extract the section radial offset features of the tunnel point cloud relative to the center of the same ring segment section ;

[0057] S2: Based on the extracted cross-section radial offset feature Perform local translation alignment and local normalization on the point cloud to obtain local translation features and local normalized features ;

[0058] S3: Offset the section radially , local translation feature With local normalized features Perform weighted fusion to generate multi-channel features ;

[0059] S4: Multi-channel features Feed Trained in the semantic segmentation network, the tunnel wall point cloud is extracted based on the classification results.

[0060] Furthermore, the construction process of the tunnel point cloud section radial offset feature extraction module in step S1 is as follows:

[0061] S1-1: Calculate the center coordinates of the cross section where each point is located ;

[0062] S1-2: Calculate the radial distance between each point and the center coordinate of its cross section ;

[0063] S1-3: Calculate the radial distance of each point Difference from the section radius , and obtain the radial offset characteristics of each point in the tunnel section .

[0064] Furthermore, the specific process of step S3 is: according to the preset weight, the cross-section radial offset feature , local translation feature With local normalization Features are linearly combined to generate multi-channel features .

[0065] Furthermore, the process of step S4 is: use the training model to perform semantic segmentation on the point cloud and generate a label for each point ; Filter non-tunnel wall point clouds according to classification labels and extract tunnel wall point clouds.

[0066] Furthermore, if Figure 2 As shown, in step S1-1, the center coordinates of the cross section where each point is located are calculated. The steps are:

[0067] S1-1.1: Project the tunnel point cloud along the elliptical cylinder and unfold it, extract the tunnel bolt hole point cloud according to the distance and normal vector constraints, and segment the tunnel point cloud ring by ring according to the bolt hole point cloud; The process of projecting the tunnel point cloud onto the elliptical cylinder and unfolding it is as follows: Assume the fitting cylinder equation is: , is a point on the cylinder axis, is the axis direction vector, a point in the tunnel point cloud The corresponding cylinder axis point The relationship is: ; The direction vector from point P to point A is: ;

[0068] Project the point P Translate along the projection direction to the cylindrical surface: , r is the radius of the cylinder;

[0069] The coordinates after unfolding the cylindrical surface are obtained by the following formula: ;

[0070] In the formula, are the coordinates of a point on the cylindrical surface, is the coordinate of the point after expansion, R is the design radius of the shield tunnel, is the angle between the straight line from the center of the section to the point and the vertical line;

[0071] S1-1.2: If Figure 3 As shown, the central axis L of each ring is extracted, and the point cloud is rotated as a whole until the central axis is parallel to the Y direction; the coordinates of the center point of the central section are extracted based on the least squares circle fitting , get the center coordinates of the cross section where each point is located ; The central axis L is expressed as follows: ;

[0072] in, is the equation of the line on the YOZ plane, is the equation of the line on the XOY plane;

[0073] ;

[0074] In the formula, is the coordinate of a point in the segment ring, and n is the number of point clouds in the segment ring; the center coordinates of the cross section of the center point of each ring are extracted by least squares circle fitting , so that the objective function is minimized, the objective function is expressed as: ;

[0075] In the formula, is the coordinate of a point on the cross section of each ring center, is the fitting radius of the cross-section point cloud; therefore, any point of each ring The coordinates of the center point of the cross section are obtained as ;

[0076] In step S1-2, the radial distance between each point and the center coordinate of its cross section is calculated. The formula is: ;

[0077] In step S1-3, the radial distance between each point and the center coordinate of its cross section is calculated. With cross-sectional radius The difference between : .

[0078] Furthermore, the process of performing local translation alignment and local normalization processing on the point cloud in step S2 is as follows: according to the cross-sectional radial offset feature , translate the point cloud along the tunnel's forward direction and elevation direction to generate local translation features ; Perform minimum-maximum local normalization processing on the point cloud in the X, Y, and Z directions of the tunnel to generate local normalized features ; Among them, the local translation feature The calculation formula is: ;

[0079] In the formula, is the coordinate after translation;

[0080] Local normalized features The calculation formula is: ;

[0081] In the formula, is the local normalized coordinate, It is the extreme value of the point cloud coordinates in three directions within each segment ring.

[0082] Furthermore, in step S3, the multi-channel feature The calculation formula is: ;

[0083] In the formula, Section radial offset features , local translation feature and local normalized features The weight of .

[0084] Furthermore, in step S4, the multi-channel features Feed The steps of training the semantic segmentation network and extracting the tunnel wall point cloud according to the classification results are as follows: S4-1: multi-channel features Input to Use the trained model to perform semantic segmentation on the point cloud and generate a label for each point ;like Figure 4 As shown, the non-tunnel wall point cloud is filtered according to the classification label, and the tunnel wall point cloud is extracted;

[0085] S4-2: Fusion of multi-channel features Local features after hierarchical feature extraction Generate global eigenvector , the global eigenvector Input to the fully connected layer, using a multi-layer perceptron ( MLP ) for global features Processed and passed The activation function generates the classification probability of each point , the formula is as follows: ;

[0086] In the formula, Yes The probability distribution of semantic labels;

[0087] S4-3: Train the model by minimizing the cross entropy loss function. The cross entropy loss function is as follows: ;

[0088] In the formula, Yes The real label, is the probability of predicting the label, N refers to the total number of points involved in the loss calculation, and C is the number of categories;

[0089] S4-4: Perform semantic segmentation on the new point cloud based on the trained model and generate the label of each point. The formula is as follows: ;

[0090] In the formula, are the trained network parameters, is the predicted semantic label set, PY is the cross-sectional radial offset feature extraction module, for Network model;

[0091] S4-5: Based on the predicted label , filter out points classified as tunnel walls: ;

[0092] In the formula, Yes The predicted label of .

[0093] like Figure 5 As shown in Figure 2, the performance of the present invention in semantic segmentation tasks is better than that of traditional Significantly improved. During the training process, the present invention can reach a convergence state after only 20 iterations, showing a faster convergence speed; at the same time, the segmentation accuracy of the model is also significantly improved, reflecting a more superior classification ability;

[0094] like Figure 6 As shown in Figure 2, the loss function (Loss) value of the present invention also shows faster convergence, can be stable after about 20 iterations, and the final Loss value is significantly smaller. These performances show that the present invention is superior to the traditional one in terms of accuracy and training efficiency. , significantly improving the overall performance of the semantic segmentation task.

[0095] like Figure 7 As shown, the present invention is compared with the traditional When processing the point cloud of ancillary facilities very close to the tunnel wall, it can more accurately identify and separate noise points, which not only improves the accuracy of denoising, but also ensures the integrity and accuracy of the tunnel wall point cloud, significantly improves the quality and reliability of the point cloud data, and provides a more reliable data basis for subsequent tunnel structure analysis and safety assessment.

Claims

1. A subway shield tunnel point cloud filtering method combining offset features and semantic segmentation, It is characterized in that The following steps are involved: S1: Construct a tunnel point cloud section radial offset feature extraction module and extract the section radial offset feature f of the tunnel point cloud relative to the center of the same ring segment section d ; S2: According to the extracted cross-section radial offset feature f d Perform local translation alignment and local normalization on the point cloud to obtain the local translation feature f p and the local normalized feature f g ; Among them, the process of local translation alignment and local normalization of the point cloud is: according to the cross-sectional radial offset feature f d , translate the point cloud along the tunnel forward direction and elevation direction to generate the local translation feature f p ; Perform minimum-maximum local normalization processing on the point cloud in the X, Y, and Z directions of the tunnel to generate local normalized features f g ; Among them, the local translation feature f p The calculation formula is: In the formula, (x i ,y i ,z i ) is the coordinate of a point inside the segment ring, (x 0 ,y 0 ,z 0 ) is a point on the cylinder axis, is the coordinate after translation; Local normalized feature f g The calculation formula is: In the formula, is the local normalized coordinate, x min ,x max ,y min ,y max ,z min ,z max are the extreme values ​​of the point cloud coordinates in three directions within each segment ring; S3: radially offset the section feature f d , local translation feature f p With the local normalized feature f g Perform weighted fusion to generate multi-channel features f m ; S4: Multi-channel features f m Send it to the PointNet++ semantic segmentation network for training, and extract the tunnel wall point cloud based on the classification results. The specific steps are as follows: S4-1: Multi-channel features f m Input into the PointNet++ network for training; use the trained model to perform semantic segmentation on the point cloud and generate labels for each point S4-2: Fusion of multi-channel features f m And the local features f after hierarchical feature extraction i Generate global eigenvector f global , the global eigenvector f global Input to the fully connected layer, use multi-layer perceptron to analyze the global feature f global Process it and generate the classification probability y of each point through the softmax activation function i , the formula is as follows: and i =softmax(MLP(f global )); In the formula, y i It's point p i The probability distribution of semantic labels; S4-3: Train the model by minimizing the cross entropy loss function. The cross entropy loss function is as follows: In the formula, y i,c It's point p i The real label, is the probability of predicting the label, N refers to the total number of points involved in the loss calculation, and C is the number of categories; S4-4: Perform semantic segmentation on the new point cloud based on the trained model and generate the label of each point. The formula is as follows: Where P is the cylinder axis point, θ * are the trained network parameters, is the predicted semantic label set, PY is the cross-section radial offset feature extraction module, and PointNet++ is the PointNet++ network model; S4-5: Based on the predicted label Filter out points classified as tunnel walls: In the formula, It's point p i The predicted label of .

2. According to claim 1, the subway shield tunnel point cloud filtering method combining offset features and semantic segmentation, It is characterized in that The construction process of the tunnel point cloud section radial offset feature extraction module in step S1 is as follows: S1-1: Calculate the center coordinates p of the cross section where each point is located 0 (x o ,y i ,z o ); S1-2: Calculate the radial distance D of each point from the center coordinate of its cross section i ; S1-3: Calculate the radial distance D of each point i The difference between the cross-sectional radius ΔD i , and obtain the radial offset characteristics f of each point in the tunnel section d .

3. The subway shield tunnel point cloud filtering method combining offset features and semantic segmentation according to claim 1 or 2, It is characterized in that The specific process of step S3 is: according to the preset weight, the cross-section radial offset feature f d , local translation feature f p With local normalization f g Features are linearly combined to generate multi-channel features f m .

4. According to claim 3, the subway shield tunnel point cloud filtering method combining offset features and semantic segmentation, It is characterized in that The process of step S4 is: use the training model to perform semantic segmentation on the point cloud and generate a label for each point The non-tunnel wall point cloud is filtered according to the classification label and the tunnel wall point cloud is extracted.

5. According to claim 2, the subway shield tunnel point cloud filtering method combining offset features and semantic segmentation, It is characterized in that In step S1-1, the center coordinates p of the cross section where each point is located are calculated. 0 (x o ,y i ,z o ) are as follows: S1-1.1: Project the tunnel point cloud along the elliptical cylinder and unfold it, extract the tunnel bolt hole point cloud according to the distance and normal vector constraints, and segment the tunnel point cloud ring by ring according to the bolt hole point cloud; The process of projecting the tunnel point cloud onto the elliptical cylinder and unfolding it is as follows: Assume the fitting cylinder equation is: (x 0 ,y 0 ,z 0 ) is a point on the cylinder axis, (l,m,n) is the axis direction vector, and a point A(x A ,y A ,z A ) and the corresponding cylinder axis point P(x P ,y P ,z P )The relationship is: The direction vector from point P to point A is: Translate the projection point P to the cylindrical surface along the projection direction: r is the radius of the cylinder; The coordinates after unfolding the cylindrical surface are obtained by the following formula: In the formula, (x 1 ,y 1 ,z 1 ) is the coordinate of a point on the cylindrical surface, (x 2 ,y 2 ,z 2 ) is the coordinate of the point after expansion, R is the design radius of the shield tunnel, α c is the angle between the straight line from the center of the section to the point and the vertical line; S1-1.2: Extract the central axis L of each ring and rotate the point cloud as a whole until the central axis is parallel to the Y direction; extract the coordinates of the center point of the central section (x 0 ,y 0 ,z 0 ), and get each point p(x i ,y i ,z i ) is the center coordinate of the cross section 0 (x o ,y i ,z o ); the central axis L is expressed as follows: Among them, L 1 is the equation of the line on the YOZ plane, L 2 is the equation of the line on the XOY plane; In the formula, (x i ,y i ,z i ) is the coordinate of a point in the segment ring, and n is the number of point clouds in the segment ring; The center coordinates (x 0 ,y 0 ,z 0 ), so that the objective function is minimized, and the objective function is expressed as: In the formula, (x i ,z i ) is the coordinate of a point on the cross section of the center point of each ring, r 0 is the fitting radius of the cross-section point cloud; therefore, any point P(x i ,y i ,z i ) is obtained as p 0 (x o ,y i ,z o ); In step S1-2, the radial distance D of each point from the center coordinate of its cross section is calculated. i The formula is: In step S1-3, the radial distance D of each point from the center coordinate of its cross section is calculated. i The formula is:

6. According to claim 1, the subway shield tunnel point cloud filtering method combining offset features and semantic segmentation, It is characterized in that In step S3, the multi-channel feature f m The calculation formula is: f m =w d ·f d +w p ·f p +w g ·f g ; In the formula, w d ,w p ,w g They are the cross-sectional radial offset characteristics f d , local translation feature f p and the local normalized feature f g The weight of .