Shield tunnel single-ring point cloud segmentation method based on point cloud feature deep learning

Abstract

The invention discloses a shield tunnel single-ring point cloud segmentation method based on point cloud feature deep learning, and the method comprises the steps: carrying out the training of a point cloud feature deep learning model through the shield tunnel single-ring point cloud data containing marking information, and achieving the intelligent segmentation of the shield tunnel single-ring point cloud through a point cloud feature deep learning model obtained through training. The input point cloud is the shield tunnel single-ring original point cloud, the output point cloud is the shield tunnel single-ring point cloud containing semantic information, and the method is applied to a noise segmentation task and a segment segmentation task of the shield tunnel single-ring point cloud. The method has the advantages that automatic extraction of single-ring space information and deformation of the shield tunnel is achieved, and the accuracy of the segmentation task is improved.

Description

Technical Field

[0001] The present invention relates to the field of shield tunnel engineering, and in particular to a shield tunnel single-ring point cloud segmentation method based on point cloud feature deep learning. Background Art

[0002] In the wave of development and utilization of underground space, shield tunnels have been built in large numbers due to their technical and economic advantages. Shield tunnels are usually built in densely populated cities as shared infrastructure such as subways and pipe galleries. Their structural safety and performance status are of great significance to urban governance and social function guarantee. The shield tunnel structure is an assembled structure. The shield segments are connected into rings through bolts and other connectors. At the same time, the rings are also connected by bolts to form a tunnel. Each ring of shield lining is composed of a certain number of shield segments, forming a unit in the tunnel that bears the lateral load of the tunnel. Its spatial position and deformation can reflect the structural safety and performance status.

[0003] Since shield tunnels do not have the structural characteristics of key components, each segment and joint may deform, and it is difficult to measure the deformation in the entire area. Using technologies such as 3D laser scanning, LiDAR and photogrammetry, we can quickly obtain massive shield tunnel point clouds. By analyzing the information contained in the point clouds, we can measure the spatial position and deformation of each ring of shield tunnels, providing a basis for the safety and performance status evaluation and analysis of shield tunnels.

[0004] However, the collected point cloud data does not contain semantic information, making it difficult to directly obtain deformation information of each segment and joint and conduct various analyses. Using intelligent methods to segment shield tunnel point clouds and provide point-by-point semantic information for single-ring point clouds is the development direction and key technology for shield tunnel structure measurement based on point clouds.

[0005] At present, the main methods for single-ring point cloud segmentation of shield tunnels are as follows: CN108225184A proposes a method for denoising point cloud data of three-dimensional laser scanning of a shield tunnel, comprising the steps of: reading scanned point cloud data; establishing a Cartesian rectangular coordinate system and partitioning the scanned points into quadrants; extracting scanned point cloud data; performing preliminary noise elimination based on x-values ​​or y-values ​​and distance change trends of different quadrants according to the structural characteristics of the shield tunnel and the scanned characteristics of the three-dimensional laser scanner; performing secondary elimination of the scanned point data by circular fitting after the preliminary elimination of the noise points; performing final denoising of the three-dimensional laser scanning point cloud data by an ellipse fitting method based on the least squares method of geometric distance; repeating the above steps to process the point cloud data of all scanned rings; and outputting the effective point cloud data of the scanned points.

[0006] CN114109410A proposes a method and device for measuring the misalignment and opening of shield tunnel segments, wherein the steps involved in point cloud segmentation include: obtaining three-dimensional expanded point cloud data of a target ring of the shield tunnel based on the central axis of the shield tunnel; and obtaining boundary point cloud data of the segment assembly position within the target ring based on the three-dimensional expanded point cloud data of the target ring of the shield tunnel.

[0007] CN114370828A proposes a method for detecting the diameter convergence and radial misalignment of a shield tunnel based on laser scanning, wherein the steps involved in point cloud segmentation include: mapping three-dimensional point cloud data into a two-dimensional grayscale image, and enhancing the two-dimensional grayscale image; performing feature detection on the enhanced grayscale image to separate the pipe segment gaps with line segment features.

[0008] The shortcomings of the above methods are that they are generally composed of multiple feature extraction components, and require manual design to segment the single-ring shield tunnel segments, which usually involve processes such as point cloud rotation, expansion and fitting. The errors generated by each component will be transmitted in the method flow, reducing the accuracy of the method; at the same time, it is difficult to take into account different segmentation tasks, and the degree of automation is low. In addition, the non-end-to-end algorithm involves difficulties in precision optimization.

[0009] This is an area where this application needs to be improved. Summary of the invention

[0010] The technical problem to be solved by the present invention is to provide a single-ring point cloud segmentation method for a shield tunnel based on deep learning of point cloud features, so as to realize the automatic extraction of spatial information and deformation of a single-ring of a shield tunnel and increase the accuracy of the segmentation task.

[0011] In order to solve the above technical problems, the present invention provides a shield tunnel single-loop point cloud segmentation method based on point cloud feature deep learning, which uses the shield tunnel single-loop point cloud data containing annotation information to perform point cloud feature deep learning model training, and uses the trained point cloud feature deep learning model to realize the intelligent segmentation of the shield tunnel single-loop point cloud. The input point cloud is the shield tunnel single-loop original point cloud, and the output point cloud is the shield tunnel single-loop point cloud containing semantic information. The method is applied to the noise segmentation task and the segment segmentation task of the shield tunnel single-loop point cloud, and includes the following steps: Step S1, determining the shield tunnel single-loop point cloud deep learning task, the task types include: Task type 1 is single-loop point cloud noise segmentation, which is to divide all points in the point cloud into two categories, namely, the point cloud formed by the pipe segment and the point cloud formed by all objects except the pipe segment; Task type 2 is single-ring point cloud segment segmentation, which is to divide all points in the point cloud into n categories, that is, the point cloud formed by each segment, where n is the number of segments contained in each ring of the shield; Task type three is the overall segmentation of a single-ring point cloud, which is a combination of task type one and task type two. It divides all points in the point cloud into n+1 categories, that is, the point cloud formed by each segment and the point cloud formed by all objects except the segment.

[0012] Step S2, according to the shield tunnel single-loop point cloud deep learning task, annotate the shield tunnel single-loop point cloud data for model training and create a single-loop point cloud data set; Among them: the object represented by the point in the point cloud is the single-ring shield tunnel itself or the internal object; Corresponding to task type 1 and task type 3, the point cloud includes the point cloud formed by the pipe segment and the point cloud formed by all objects except the pipe segment; Corresponding to task type 2, the point cloud includes the point cloud formed by the pipe segment; Each single-ring point cloud contains M points. Each point in the point cloud has the same dimension, at least 3 coordinate dimensions, namely spatial coordinate information x, y, z; each point has C feature dimensions, including: normal vector information nx, ny, nz, color information r, g, b, reflection intensity information i, and all dimension values ​​are normalized to real numbers in the interval [0,1]; Step S21, annotating each point in the single-ring point cloud of the shield tunnel by a k-dimensional vector, wherein each segmentation category uniquely corresponds to a vector; Preferably, for task type 1, the point cloud is annotated with a 1-dimensional vector, the point cloud formed by the pipe segment is annotated as 1, and the point cloud formed by all objects other than the pipe segment is annotated as 0; Preferably, for task type 2 and task type 3, the point cloud is annotated with a 2D vector, the capping block is taken as the first segment, and the numbering is performed in a counterclockwise or clockwise increasing manner. The point cloud formed by the first segment is annotated as (cos(a), sin(a)), and the point cloud formed by the jth segment (j=2,3,…,n) is annotated as (cos(a+b j ),sin(a+b j )), where: a is an arbitrary angle value, b j The angle required for the center of the first segment to rotate along the segment circle center in the numbering direction to the center of the jth segment; Step S22, performing overall rotation or partial rotation on the basis of the manually annotated data to achieve data augmentation; During overall rotation, the spatial coordinate information x, y, z and optional normal vector information nx, ny, nz of all points in the point cloud are rotated to obtain a new point cloud and add it to the data set; During local rotation, the spatial coordinate information x, y, z and optional normal vector information nx, ny, nz of one or more categories of corresponding points in the point cloud are rotated to obtain a new point cloud and add it to the dataset.

[0013] Step S3, inputting the single-ring point cloud data of the shield tunnel containing the annotation information into the deep learning network for model training to construct a segmentation task model; Step S31, the deep learning network is an end-to-end model, and the prediction result error obtained from the input end to the output end will be used to adjust each layer of the model until the model converges or achieves the expected effect. The input of the network is a tensor of shape (N, 3 + C), and the output of the network is a tensor of shape (N, k), where: N is the number of points in the input network, C is the number of input channels excluding coordinates, and k is the number of prediction result channels; Step S32, the number of points M contained in each ring of shield tunnel point cloud data is greater than or equal to N. When M is greater than N, the point cloud needs to be downsampled before inputting network training. The number of downsampling times is the integer part of M / N. After each downsampling is completed, the points contained in the downsampling are removed from the original point cloud.

[0014] Step S4, using the constructed point cloud feature deep learning model to calculate the single-ring point cloud of the shield tunnel to be segmented; Step S41, the number of points M contained in each ring of shield tunnel point cloud data is greater than or equal to N. When M is greater than N, the point cloud needs to be downsampled before inputting into the model calculation. The number of downsampling is the integer part of M / N. After each downsampling is completed, the points contained in the downsampling are removed from the original point cloud; Step S42: After the model outputs the result, all downsampled data corresponding to each ring point cloud needs to be merged.

[0015] The superior effects of the present invention are: 1) The end-to-end cloud feature deep learning network is used to segment the single-ring shield segment point cloud, which can quickly and accurately perform various segmentation tasks, provide sufficient semantic information for the single-ring point cloud of the shield tunnel, extract the point clouds corresponding to each segment and other objects, and realize the automatic extraction of the single-ring spatial information and deformation of the shield tunnel; 2) K-dimensional vector encoding is used to annotate point clouds. Different from the commonly used one-hot encoding method, the method used in the present invention can reflect the inherent mutual position relationship between point clouds of different segments of shield tunnels and point clouds of other objects in the label, and improve from a mathematical perspective, effectively avoiding the overfitting phenomenon of deep learning networks, increasing the accuracy of segmentation tasks, and effectively improving the prediction performance of deep learning models; 3) Using overall rotation and local rotation to augment data can reduce the workload of manual point cloud annotation, achieve excellent single-ring point cloud segmentation effect of shield tunnel when measured data is insufficient, and improve the application level of deep learning methods in the field of shield tunnel engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The drawings constituting a part of the present application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 It is a work flow chart of a specific embodiment of the present invention; Figure 2 is a top view of a single-ring point cloud of a tunnel according to a specific embodiment of the present invention; Figure 3 is a front view of a single-ring point cloud of a tunnel according to a specific embodiment of the present invention; Figure 4 is a three-dimensional view of a single-loop point cloud of a tunnel according to a specific embodiment of the present invention; Figure 5 is the annotation result of the single-ring point cloud data set corresponding to the noise segmentation task in an embodiment of the present invention; Figure 6 is the labeling result of the single-ring point cloud data set corresponding to the segment segmentation task in the embodiment of the present invention; Figure 7 It is a schematic diagram of the process of noise segmentation of a single-ring point cloud of a shield tunnel according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0018] like Figure 1 As shown, the present invention provides a single-loop point cloud segmentation method for a shield tunnel based on deep learning of point cloud features, determines a deep learning task of a single-loop point cloud for a shield tunnel; according to the deep learning task of the single-loop point cloud for a shield tunnel, annotates the single-loop point cloud data of the shield tunnel for model training, and produces a single-loop point cloud data set; inputs the single-loop point cloud data of the shield tunnel containing the annotation information into a deep learning network for model training, and constructs a segmentation task model; and uses the constructed point cloud feature deep learning model to calculate the single-loop point cloud of the shield tunnel that needs to be segmented. Specific embodiments

[0019] Taking a shield tunnel consisting of 6 segments as an example, its single ring point cloud is as follows Figure 2 , 3 , 4, which includes the point clouds formed by 6 segments and the point clouds formed by other objects except the segments. The dataset annotation results corresponding to the shield tunnel single ring point cloud noise segmentation task are shown in Figure 5 As shown in , the point cloud is annotated with a 1-dimensional vector. The point cloud formed by the segment is annotated as 1, and the point cloud formed by all objects other than the segment is annotated as 0. The dataset annotation results corresponding to the shield tunnel single-ring point cloud noise segmentation task are shown in Figure 6As shown, the point cloud is annotated with a two-dimensional vector. The first to sixth segments are annotated as (1, 0), (0.71, 0.71), (-0.38, 0.92), (-1, 0), (-0.38, -0.92), and (0.71, -0.71), respectively.

[0020] The specific implementation method is introduced by taking the noise segmentation of single-ring point cloud of shield tunnel as an example. The detailed process is as follows: Figure 7 As shown: Step S1: Each point in the point cloud has 6 dimensions, including spatial coordinate information x, y, z and normal vector information nx, ny, nz. All dimensional values ​​are normalized to real numbers in the interval [0, 1].

[0021] Step S2: Using the measured 67-ring shield tunnel point cloud to create a data set, the data is rotated as a whole or partially based on the manually annotated data to achieve data augmentation; During overall rotation, the spatial coordinate information x, y, z and optional normal vector information nx, ny, nz of all points in the point cloud are rotated to obtain a new point cloud and add it to the dataset; during local rotation, the spatial coordinate information x, y, z and normal vector information nx, ny, nz of one or more categories of corresponding points in the point cloud are rotated to obtain a new point cloud and add it to the dataset; after data augmentation and downsampling, the total dataset has 2193 samples.

[0022] Step S3: Input the single-ring point cloud data of the shield tunnel containing the annotation information into PointNet++ for model training to build a segmentation task model; The deep learning network is an end-to-end model. The prediction result error obtained from the input to the output will be used to adjust each layer of the model until the model converges or achieves the expected effect. The training is terminated. The input of the network is a tensor with a shape of (2048,6), and the output of the network is a tensor with a shape of (2048,2). Verified on the test set, the constructed point cloud feature deep learning model has a segmentation accuracy of 0.89, which is used for noise segmentation of single-ring point clouds in measured shield tunnels.

[0023] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A shield tunnel single-ring point cloud segmentation method based on deep learning of point cloud features, characterized by: The point cloud feature deep learning model is trained through the single-loop point cloud data of the shield tunnel containing annotated information, and the trained point cloud feature deep learning model is used to realize the intelligent segmentation of the single-loop point cloud of the shield tunnel. The input point cloud is the original point cloud of the single-loop point cloud of the shield tunnel, and the output point cloud is the single-loop point cloud of the shield tunnel containing semantic information. It is applied to the noise segmentation task and segment segmentation task of the single-loop point cloud of the shield tunnel.

2. According to claim 1, a shield tunnel single-ring point cloud segmentation method based on point cloud feature deep learning is characterized in that: The following steps are involved: Step S1, determining a single-loop point cloud deep learning task for a shield tunnel; Step S2, according to the shield tunnel single-loop point cloud deep learning task, annotate the shield tunnel single-loop point cloud data for model training and create a single-loop point cloud data set; Step S3, inputting the single-ring point cloud data of the shield tunnel containing the annotation information into the deep learning network for model training to construct a segmentation task model; Step S4, using the constructed point cloud feature deep learning model to calculate the single-ring point cloud of the shield tunnel that needs to be segmented.

3. According to claim 2, a shield tunnel single-ring point cloud segmentation method based on point cloud feature deep learning is characterized in that: The step S2 comprises the following steps: Step S21, annotating each point in the single-ring point cloud of the shield tunnel by a k-dimensional vector, wherein each segmentation category uniquely corresponds to a vector; Step S22, performing overall rotation or partial rotation on the basis of the manually annotated data to achieve data augmentation; During overall rotation, the spatial coordinate information x, y, z and optional normal vector information nx, ny, nz of all points in the point cloud are rotated to obtain a new point cloud and added to the dataset; during local rotation, the spatial coordinate information x, y, z and optional normal vector information nx, ny, nz of one or more categories of corresponding points in the point cloud are rotated to obtain a new point cloud and added to the dataset.

4. The shield tunnel single-ring point cloud segmentation method based on point cloud feature deep learning according to claim 2 is characterized in that: The step S3 comprises the following steps: Step S31, the deep learning network is an end-to-end model, and the prediction result error obtained from the input end to the output end will be used to adjust each layer of the model until the model converges or achieves the expected effect, and then the training ends; Step S32, the number of points M contained in each ring of shield tunnel point cloud data is greater than or equal to N. When M is greater than N, the point cloud needs to be downsampled before inputting network training. The number of downsampling times is the integer part of M / N. After each downsampling is completed, the points contained in the downsampling are removed from the original point cloud.

5. The shield tunnel single-ring point cloud segmentation method based on point cloud feature deep learning according to claim 2 is characterized in that: The step S4 comprises the following steps: Step S41, the number of points M contained in each ring of shield tunnel point cloud data is greater than or equal to N. When M is greater than N, the point cloud is downsampled before input into the model calculation, and the number of downsampling is the integer part of M / N. After each downsampling is completed, the points contained in the downsampling are removed from the original point cloud; Step S42: After the model outputs the result, all downsampled data corresponding to each ring point cloud are merged.

6. A computer system comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the shield tunnel single-ring point cloud segmentation method described in any one of claims 1-5.

Images