Neural network construction device and method for tunnel three-dimensional point cloud intelligent processing

Abstract

The invention discloses a neural network construction device for tunnel three-dimensional point cloud intelligent processing, which comprises a neural network input module, a neural network processing module and a neural network output module, and is characterized in that the neural network processing module preprocesses a feature map of an input point cloud and then performs feature aggregation to obtain an output point cloud feature map; and performing target identification, semantic segmentation and instance segmentation on the tunnel three-dimensional point cloud deep learning network. The tunnel three-dimensional point cloud intelligent processing method has the advantages that the learning and training effects of the neural network model are effectively improved, the accuracy of tunnel three-dimensional point cloud intelligent processing results is improved, and the digitization and intelligence level of tunnel engineering can be improved.

Description

Technical Field

[0001] The present invention relates to the field of tunnel engineering, and particularly to a neural network construction device and method for intelligent processing of tunnel three-dimensional point clouds. Background Art

[0002] The application of laser scanning in tunnel engineering has been gradually promoted to collect three-dimensional point cloud data containing rich spatial information. The three-dimensional point cloud has various application scenarios in tunnel engineering, including asset management, deformation measurement, defect identification, etc. However, the three-dimensional point cloud itself only contains information such as three-dimensional coordinates, laser reflection intensity, and color, and it needs to be processed by object recognition, semantic segmentation, instance segmentation, etc. to extract effective information. With the development of three-dimensional point cloud deep learning technology, neural networks represented by PointNet++ and RandLA-Net have received wide attention and shown great application prospects. However, the existing neural networks are designed for general scenarios and do not consider the characteristics of tunnel three-dimensional point clouds, resulting in poor recognition and segmentation effects. Therefore, it is urgent to carry out neural network construction for tunnel three-dimensional point clouds to promote the application of advanced technologies in tunnel engineering and improve the digital and intelligent level of tunnel engineering. This is where this application needs to be improved. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a neural network construction device and method for intelligent processing of tunnel three-dimensional point clouds, and to improve the automation level of tunnel three-dimensional point cloud processing through targeted neural network construction.

[0004] To solve the above technical problems, the present invention provides a neural network construction device for intelligent processing of tunnel three-dimensional point clouds. The neural network processing module aggregates the features of the input three-dimensional point cloud coordinates and its feature map to obtain an output point cloud feature map, which includes: Neural network input module: The input three-dimensional point cloud coordinates and its feature map are tensors with dimensions of and respectively, where: and are the number of points and the number of features of the input point cloud respectively; Neural network processing module: First preprocess the feature map of the input point cloud to increase the depth of the neural network, and then based on the neighborhood search designed for tunnel three-dimensional point clouds, the three-dimensional coordinates and feature map of all points in the neighborhood jointly constitute the overall information of the tunnel cross-section where the query point is located. On this basis, feature splicing and calculation are performed to obtain an overall neighborhood feature map, and feature pooling is performed on the overall neighborhood feature map to obtain a pooling tensor; Neural network output module: Splice the pooling tensor and the input point cloud feature map, and convert them into the coordinates and feature map of the output point cloud, which are tensors with dimensions of and Zhang of quantity , where: and are the number of points and the number of features of the output point cloud, respectively.

[0005] The neighborhood search is based on the projection positions of each point in the point cloud on the tunnel axis, and the nearest points are used as the search results, where: is the neighborhood size.

[0006] The tunnel axis is determined by cylindrical fitting of the tunnel three-dimensional point cloud.

[0007] The feature splicing and calculation refer to splicing the three-dimensional coordinates and feature maps of all points in the neighborhood, and obtaining the overall neighborhood feature map through convolution, batch normalization, and activation function operations.

[0008] The overall neighborhood feature map is a tensor with a dimension of , where: is the number of features in the feature aggregation process.

[0009] The three-dimensional coordinates of all points in the neighborhood are first encoded for relative positions, that is, the three-dimensional coordinates of the neighborhood points, the three-dimensional coordinates of the query points, the difference between the two, and the L2 norm of the difference between the two are spliced to obtain an overall neighborhood spatial information tensor with a dimension of , and then spliced with the feature map to increase the richness of spatial information. Where: 10 = 3 for the three-dimensional coordinates of the neighborhood points + 3 for the three-dimensional coordinates of the query points + 3 for the difference between the two + 1 for the L2 norm of the difference between the two.

[0010] The feature splicing is performed on the last dimension of the two tensors, and the other dimensions of the two tensors are the same except for the last dimension.

[0011] The feature pooling refers to performing a pooling operation on the second dimension of the overall neighborhood feature map to obtain a pooling tensor with a dimension of , the pooling tensor is spliced with the input point cloud feature map, and then converted to an output point cloud feature map with a dimension of through convolution, batch normalization, and activation function operations.

[0012] The feature pooling uses max pooling or average pooling.

[0013] Preferably, the pooling tensor and the input point cloud feature map are spliced before the convolution operation after pooling to achieve a skip connection, and then converted to an output point cloud feature map through convolution operation.

[0014] All weights and parameters in the neural network are determined by training based on labeled data.

[0015] The calculation includes convolution, batch normalization, and activation functions. The depth of the neural network is increased by adding convolution, batch normalization, and activation functions.

[0016] The present invention also provides a method for constructing a neural network for intelligent processing of tunnel three-dimensional point clouds. The neural network aggregates the features of the input three-dimensional point cloud coordinates and its feature map to obtain an output point cloud feature map. The neural network serves as a deep learning network for target recognition, semantic segmentation, and instance segmentation of tunnel three-dimensional point clouds. The specific steps include: Preprocess the feature map of the input point cloud first to increase the depth of the neural network; Based on the neighborhood search designed for tunnel three-dimensional point clouds, the three-dimensional coordinates and feature maps of all points in the neighborhood together constitute the overall information of the tunnel cross-section where the query point is located. On this basis, feature splicing and calculation are performed to obtain an overall neighborhood feature map, and feature pooling is performed on the overall neighborhood feature map to obtain a pooled tensor; Splice the pooled tensor and the input point cloud feature map to convert it into an output point cloud feature map.

[0017] The output point cloud is a subset of the input point cloud.

[0018] The beneficial effects of the present invention are as follows: 1) Considering the longitudinal distribution characteristics of the tunnel and its three-dimensional point clouds, a method for constructing a neural network for intelligent processing of tunnel three-dimensional point clouds is specifically proposed; 2) Effectively improving the learning and training effects of the neural network model, improving the accuracy of the intelligent processing results of tunnel three-dimensional point clouds, and being beneficial to enhancing the digital and intelligent levels of tunnel engineering. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The specification drawings constituting a part of this application are used to provide a further understanding of the present invention. The schematic 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 Schematic diagram of the neural network architecture of a specific embodiment of the present invention; Figure 2 Schematic diagram of the axial projection of a specific embodiment of the present invention; Figure 3 Schematic diagram of the neighborhood search of a specific embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The embodiments of the present invention are described in detail below with reference to the drawings.

[0021] As Figure 1As shown in the figure, the present invention provides a method for constructing a neural network for intelligent processing of tunnel three-dimensional point clouds. The input of the neural network is the three-dimensional coordinates and feature map of the input point cloud, and the output of the neural network is the three-dimensional coordinates and feature map of the output point cloud, where the output point cloud is a subset of the input point cloud. Specifically, it includes the following: First, perform convolution, batch normalization, and activation function operations on the feature map of the input point cloud to increase the depth of the neural network. The neural network is based on the neighborhood search designed for tunnel three-dimensional point clouds. The designed neighborhood uses the projection positions of each point in the point cloud on the tunnel axis as the search basis. The neighborhood is the points whose axial projections are closest to the axial projection of the query point. The tunnel axis is determined by fitting a cylinder to the tunnel three-dimensional point cloud, and the axial projection is as Figure 2 shown; As Figure 3 shown, the designed neighborhood is the points whose axial projections are closest to the axial projection of the query point; The three-dimensional coordinates and feature maps of all points in the neighborhood together constitute the overall information of the tunnel cross-section where the query point is located. Based on this, feature calculation and splicing are performed. Before splicing the three-dimensional coordinates and feature maps of all points in the neighborhood, perform relative position encoding on the three-dimensional coordinates of all points in the neighborhood and the three-dimensional coordinates of the query point, that is, splice the three-dimensional coordinates of the neighborhood points, the three-dimensional coordinates of the query point, the difference between the two, and the L2 norm of the difference between the two to obtain the overall spatial information tensor of the neighborhood. Splice the three-dimensional coordinates and feature maps of all points in the neighborhood, and obtain the overall feature map of the neighborhood through convolution, batch normalization, and activation function operations. Perform max-pooling on the overall feature map of the neighborhood to obtain a pooling tensor, splice the input point cloud feature map into the pooling tensor to achieve a skip connection, and then perform convolution, batch normalization, and activation function operations on the pooling tensor to convert it into an output point cloud feature map of the required dimension.

[0022] All weights and parameters in the neural network are determined by training based on labeled data.

[0023] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A neural network construction device for intelligent processing of tunnel three-dimensional point clouds, characterized in that: The neural network processing module performs feature aggregation on the input three-dimensional point cloud coordinates and its feature map to obtain an output point cloud feature map, which includes: Neural network input module: The input three-dimensional point cloud coordinates and its feature map are tensors with dimensions of and respectively, where: and are the number of points and the number of features of the input point cloud respectively; The neural network processing module: first preprocesses the feature map of the input point cloud to increase the depth of the neural network, and then, based on the neighborhood search designed for the three-dimensional tunnel point cloud, the three-dimensional coordinates and feature map of all points in the neighborhood together constitute the overall information of the tunnel cross-section where the query point is located. On this basis, feature splicing and calculation are performed to obtain the overall neighborhood feature map, and feature pooling is performed on the overall neighborhood feature map to obtain a pooled tensor; Neural network output module: Concatenate the pooled tensor and the input point cloud feature map, and convert them into the coordinates and feature map of the output point cloud, which are tensors with dimensions of and respectively, where: and are the number of points and the number of features of the output point cloud respectively.

2. The neural network construction device for intelligent processing of tunnel three-dimensional point clouds according to claim 1, characterized in that: The neighborhood search is based on the projection positions of each point in the point cloud on the tunnel axis, and the nearest points are used as the search results, where: is the neighborhood size.

3. The neural network construction device for intelligent processing of tunnel three-dimensional point clouds according to claim 2, wherein: The tunnel axis is determined by performing cylindrical fitting on the three-dimensional tunnel point cloud.

4. The neural network construction device for intelligent processing of tunnel three-dimensional point clouds according to claim 1, characterized in that: The feature splicing and calculation refer to splicing the three-dimensional coordinates and feature map of all points in the neighborhood and obtaining the overall neighborhood feature map through convolution, batch normalization, and activation function operations.

5. The neural network construction device for intelligent processing of tunnel three-dimensional point clouds according to claim 4, characterized in that: The neighborhood global feature map is a tensor with a dimension of , where: is the number of features in the feature aggregation process, is the neighborhood size.

6. The neural network construction device for intelligent processing of tunnel three-dimensional point clouds according to claim 4, characterized in that: The three-dimensional coordinates of all points in the neighborhood are first encoded for relative position and then spliced with the feature map.

7. The neural network construction device for intelligent processing of tunnel three-dimensional point clouds according to claim 6, wherein: The splicing is performed on the last dimension of the two tensors.

8. The neural network construction device for intelligent processing of tunnel three-dimensional point clouds according to claim 1, characterized in that: The feature pooling refers to performing a pooling operation on the second dimension of the overall feature map of the neighborhood to obtain a pooling tensor with a dimension of . The pooling tensor is concatenated with the input point cloud feature map, and then through convolution, batch normalization, and activation function operations, it is transformed into an output point cloud feature map with a dimension of .

9. A neural network construction method for intelligent processing of tunnel three-dimensional point clouds, characterized in that: The neural network performs feature aggregation on the input three-dimensional point cloud coordinates and its feature map to obtain an output point cloud feature map. The neural network, as a deep learning network for target recognition, semantic segmentation, and instance segmentation of three-dimensional tunnel point clouds, includes: Preprocessing the feature map of the input point cloud first to increase the depth of the neural network; Based on the neighborhood search designed for the three-dimensional tunnel point cloud, the three-dimensional coordinates and feature map of all points in the neighborhood together constitute the overall information of the tunnel cross-section where the query point is located. On this basis, feature splicing and calculation are performed to obtain the overall neighborhood feature map, and feature pooling is performed on the overall neighborhood feature map to obtain a pooled tensor; Splicing the pooled tensor and the input point cloud feature map and converting them into an output point cloud feature map.

10. The neural network construction method for intelligent processing of tunnel three-dimensional point clouds according to claim 9, wherein: All weights and parameters in the neural network are determined through training based on labeled data.

Images