Tunnel point cloud high-fidelity completion method based on double-branch coding and geometric prior guidance

Abstract

The application discloses a kind of based on double branch coding and geometric prior guide's high fidelity completion method of roadway point cloud, wherein double branch anisotropy-axial prior network includes: a double branch parallel encoder, which works in parallel by a local geometric perception branch and an axial perception branch, the local geometric perception branch is improved by adopting the anisotropic neighborhood construction strategy guided by central axis Point Transformer module, so that it is adapted to the extension direction of lane / tunnel to accurately extract micro features;Axial perception branch utilizes central axis to serialize point cloud, effectively captures macro structural dependence;A structured prior guided decoder, which can introduce the inherent geometric continuity prior of roadway as a dynamic constraint in the process of generating point cloud from coarse to fine, to ensure that the completed surface is smooth and consistent with the structure logic.The application can improve the global structure consistency, geometric regularity and high fidelity of the completed point cloud.

Description

Technical Field

[0001] This invention relates to the field of point cloud processing technology, and in particular to a high-fidelity point cloud completion method for roadways based on bi-branch coding and geometric prior guidance. Background Technology

[0002] The core of 3D reconstruction of coal mine roadways lies in acquiring complete and accurate 3D point cloud data. However, in the real underground environment, point cloud data acquisition faces severe challenges. The ubiquitous dust and water mist severely interfere with lidar signals; simultaneously, the widely distributed support structures, pipelines, conveyor belts, and other equipment on the roadway walls cause large-scale physical obstructions. These factors combined result in the acquired roadway point clouds often exhibiting problems such as partial missing data, sparse areas, and uneven density, affecting subsequent applications such as engineering quantity calculations, geological structure analysis, equipment path planning, and safety monitoring.

[0003] To address the problem of incomplete point clouds, traditional geometric methods typically rely on interpolation, symmetry assumptions, or geometric primitive fitting. These methods can handle objects with simple structures or small missing parts, but their performance drops sharply when faced with large-scale, slender, topologically complex scenes such as coal mine tunnels that lack obvious symmetry, making it difficult to generate satisfactory results.

[0004] In recent years, point cloud completion methods based on deep learning have become a research hotspot. However, existing methods still have significant shortcomings when dealing with special industrial scenarios such as tunnels. In mainstream encoder-decoder architectures, multilayer perceptron (MLP) encoders based on PointNet series process each point independently, ignoring the spatial topology and geometric relationships between points, making it difficult to capture the fine structure of tunnel walls. While the standard Point Transformer module introduces a self-attention mechanism to model local relationships, its K-nearest neighbor (KNN) is an isotropic neighborhood search, which cannot adapt to the anisotropic distribution characteristics of tunnel point clouds that are continuous along the extension direction but finite in the radial direction. It is also prone to introducing interference from irrelevant points such as opposite walls. On the decoder side, methods based on folding (FoldingNet) or parametric surfaces (AtlasNet) are difficult to generate structures with open boundaries and complex topologies like tunnels, and are prone to producing unnatural deformations. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a high-fidelity completion method for roadway point clouds based on bi-branch coding and geometric prior guidance, which aims to solve the problems of poor structural consistency, distortion of local detail recovery, and performance degradation caused by inability to adapt to scene characteristics when processing roadway scenes.

[0006] Technical solution: A high-fidelity completion method for roadway point clouds based on bi-branch coding and geometric prior guidance, comprising the following steps:

[0007] S1, Data Acquisition and Preprocessing: Acquire an incomplete 3D tunnel point cloud containing missing data regions as input, and preprocess the incomplete 3D tunnel point cloud to generate standardized training samples.

[0008] S2. Construct a DAAP-Net network and employ a dual-branch feature encoder to fuse local geometric features and macroscopic structural features of the tunnel. The dual-branch feature encoder includes a local geometry perception branch and an axis perception branch. The local geometry perception branch is used to extract a local geometric feature at the highest abstraction level that can characterize the details of the point cloud. The axial sensing branch is used to extract axially dependent features that characterize the topological continuity and structural patterns of the tunnel. ; local geometric features With axial dependence features The features are spliced ​​and fused along the feature channel dimension to obtain the fused features. ;

[0009] S3 utilizes a guided decoder to fuse features. A coarse-to-fine strategy is adopted to generate a layered point cloud to fill in the missing data areas; the guided decoder uses the inherent geometric continuity of the surrounding rock of the tunnel as a dynamic geometric constraint during the point cloud generation process.

[0010] S4 merges the completed point cloud generated by the guided decoder with the original incomplete input point cloud to form a structurally complete and high-fidelity alley / tunnel point cloud.

[0011] Furthermore, the detailed steps for generating standardized training samples include:

[0012] S11 employs a statistical filtering algorithm to identify and remove outliers by analyzing the statistical characteristics of each point and its neighborhood.

[0013] S12 uses a voxel downsampling algorithm to divide the entire space into cubic voxels of a preset size, and uses the three-dimensional coordinate centroid of all points in each non-empty voxel to represent the voxel, thereby obtaining a point cloud with uniform density and preserving the key geometric contours.

[0014] S13 employs a filtering method guided by the roadway centerline to eliminate interference from non-main structures;

[0015] S14 uses a sliding window strategy to cut the point cloud into multiple overlapping local point cloud blocks, and uses the farthest point sampling algorithm to uniformly sample each point cloud block to a preset number of input points N; then, the uniformly sampled point cloud is normalized; finally, by simulating single viewpoint occlusion, missing data is generated for each point cloud block as a standardized training sample.

[0016] Furthermore, the detailed implementation steps of the filtering method based on the roadway centerline guidance are as follows:

[0017] S131, using principal component analysis to determine the local principal direction of the roadway point cloud, and then combining B-spline curve fitting, extracting a three-dimensional central axis L from the global point cloud that can smoothly represent the overall direction of the roadway.

[0018] S132, proceeds along the three-dimensional central axis L with a fixed step size. At each step position, based on the cross-sectional distribution of the local point cloud around that position, an adaptive-sized inner bounding box is dynamically constructed. This bounding box can tightly enclose the main passage space of the alley.

[0019] S133 identifies all points falling into these inner bounding boxes distributed along the three-dimensional central axis L as structured interference and removes them.

[0020] Furthermore, the local geometry-aware branch outputs multi-scale local feature maps through a multi-layered stacked improved Point Transformer module and FPS downsampling operation. The axis-aware branch outputs axis-dependent features through a stacked multi-layer 1D-CNN vertical structure convolutional network. ;

[0021] Local geometric features With axial dependence features The features are concatenated along the feature channel dimension, and then subjected to deep nonlinear fusion through a fusion MLP consisting of multiple fully connected layers to obtain the final comprehensive fused features. :

[0022] ,

[0023] MLP() represents a multilayer perceptron network.

[0024] Furthermore, local feature maps are obtained. The implementation steps are as follows:

[0025] SA11, constructing an anisotropic neighborhood guided by the central axis;

[0026] For each center point to be processed First, calculate the Find the nearest projection point of the point onto the central axis L, and take the tangent vector of the central axis at that projection point, denoted as . The vector The representative alleyway is at the point The local extension direction in the vicinity is this vector. When searching for the K nearest neighbors, a weighted, direction-aware distance metric is used. :

[0027] ,

[0028] in, ,express and Euclidean distance between them; It is a positive hyperparameter used to control the intensity of anisotropy;

[0029] By using As a search metric for the K-nearest neighbor algorithm, the improved Point Transformer module constructs an anisotropic neighborhood extending along the lane direction. ;

[0030] SA12, computational position-aware self-attention;

[0031] center point Input features and neighboring points Input features Each is achieved through a learnable linear projection matrix. , , Mapped to query vectors respectively Key vector Sum value vector :

[0032] ,

[0033] Simultaneously, calculation and neighboring points relative position coordinate vector And through a separate small Mapped to relative position encoding :

[0034] ,

[0035] in, This represents a mapping function implemented by a small multilayer perceptron (MLP);

[0036] Ultimately, the center point Update features It is the weighted sum of all value vectors in its neighborhood:

[0037] ,

[0038] ,

[0039] in, It represents the dimension of the key vector, "T" indicates matrix transpose, and softmax indicates the normalized exponential function.

[0040] Furthermore, axial dependence features were obtained. The implementation steps are as follows:

[0041] SA21, for the input point cloud X in Each point in Calculation points Find the orthogonal projection point on the three-dimensional central axis 𝐿, and obtain the arc length distance of this projection point along the starting point of the central axis. According to distance The size of the points is sorted, transforming the unordered point cloud set into an ordered point cloud feature sequence tensor. Where 𝐶 is the feature dimension, It is the set of real numbers;

[0042] SA22 uses the sequence tensor S as the input to the axial sensing branch. For the nth point in the sequence, the output feature of the lth convolutional layer is... The expression is as follows:

[0043] ,

[0044] in, It represents the output features of the previous layer. ReLU stands for Modified Linear Unit Activation Function, and the convolution kernel... The size is ;

[0045] By stacking multiple layers of 1D-CNN, the final output is axially dependent features. .

[0046] Compared with the prior art, the significant advantages of this invention are as follows:

[0047] 1. Collaborative understanding of global and local information, strong structural consistency: This invention adopts a dual-branch encoder, in which the axial sensing branch treats the tunnel as a quasi-one-dimensional structure, ensuring that the completed tunnel still maintains the correct extension trend and structural continuity on a macro scale, solving the problem of global distortion or breakage that is easy to occur in the prior art.

[0048] 2. High geometric regularity of tunnel outline and facilities: By introducing a structured prior guided decoder (SPGD) and structural continuity loss, this invention effectively utilizes the geometric prior knowledge of the surrounding rock and fixed facilities of the tunnel. When filling large-area gaps, it can generate smooth, continuous surfaces that conform to physical laws, significantly improving the realism and engineering usability of the completion results.

[0049] 3. High accuracy and detail fidelity in completion: This invention adopts an improved Point Transformer module, which constructs anisotropic neighborhoods guided by the central axis. It can more intelligently focus on truly relevant neighborhoods in terms of structure. Combined with the intelligent information filtering mechanism of the adaptive feature fusion gate, it can achieve a higher level of fidelity when restoring the natural texture and complex details of the tunnel wall. Attached Figure Description

[0050] Figure 1 This is a flowchart of the present invention;

[0051] Figure 2 This is a schematic diagram of the overall architecture of the dual-branch anisotropic-axial prior network (DAAP-Net) in an embodiment of the present invention;

[0052] Figure 3 This is a comparative diagram of the anisotropic neighborhood construction strategy of the present invention and the standard KNN neighborhood construction strategy, wherein a) is a schematic diagram of the principle of the standard KNN neighborhood construction strategy, and b) is a schematic diagram of the principle of the axis-guided anisotropic neighborhood construction strategy of the present invention.

[0053] Figure 4 This is a schematic diagram illustrating the working principle of the Vertical Structure Convolutional Network (LSCN) in an embodiment of the present invention;

[0054] Figure 5 This is a schematic diagram comparing the completion effect of the roadway point cloud sample in an embodiment of the present invention, where (a) is the input point cloud, (b) is the real point cloud, and (c) is the completed point cloud. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0056] This invention provides a high-fidelity point cloud completion method for roadways based on bi-branch coding and geometric prior guidance, such as... Figure 1 As shown, this method mainly includes data acquisition and preprocessing, bi-branch feature encoding, guided decoding based on geometric priors, and output and optimization; the detailed implementation steps are as follows:

[0057] Step 1, Data Acquisition and Preprocessing;

[0058] An incomplete 3D tunnel point cloud containing missing data regions is taken as input and systematically preprocessed. The key step of this preprocessing is to extract the 3D spatial centerline that can represent the overall orientation of the tunnel and generate standardized training samples.

[0059] By preprocessing the acquired data, the raw and messy tunnel point cloud data is transformed into standardized training samples suitable for input into the deep learning network, and the key prior information required for subsequent steps—the tunnel centerline—is extracted; specifically, the following steps are included:

[0060] Step 11, global outlier removal;

[0061] The original point cloud often contains isolated noise points caused by dust, water mist reflection, or measurement errors. To eliminate their influence, this embodiment employs a statistical outlier removal (SOR) algorithm. For each point in the point cloud, the average distance to its K nearest neighbors is calculated. If the average distance of a point is greater than a threshold determined by the global average distance and standard deviation, the point is identified as an outlier and removed. This effectively reduces noise while preserving the integrity of the main tunnel structure to the greatest extent possible.

[0062] Step 12, density homogenization;

[0063] Because the distance between the laser scanner and the tunnel wall is constantly changing, the original point cloud often exhibits a problem of dense points near the edges and sparse points at distant edges. To address this issue, this embodiment employs a voxel grid downsampling algorithm. The entire 3D space is divided into a cubic voxel grid with a preset side length. For each non-empty voxel, a new point is generated using the arithmetic mean of the 3D coordinates of all points within it (i.e., the centroid) to represent that voxel. In this way, while effectively reducing data redundancy, a relatively uniform point cloud density is obtained, and key geometric contours are preserved.

[0064] Step 13, Structured interference filtering and central axis extraction;

[0065] Temporary work equipment, pipelines, pedestrians, and other non-main structures often exist inside the tunnel, interfering with the completion of the surrounding rock. This embodiment uses a filtering method based on the tunnel's central axis to accurately remove these interferences. The specific steps are as follows:

[0066] First, principal component analysis (PCA) is used to determine the local principal orientation of the roadway point cloud. Then, combined with B-spline curve fitting, a three-dimensional central axis L that smoothly represents the overall orientation of the roadway is robustly extracted from the global point cloud. Next, the roadway is stepped along this three-dimensional central axis L with a fixed step size. At each step position, an adaptively sized inner bounding box is dynamically constructed based on the cross-sectional distribution of the local point cloud around that position (by calculating the bounding box or covariance matrix of the cross-sectional point cloud). This bounding box is designed to tightly enclose the main passage space of the roadway.

[0067] Finally, all points falling into these inner bounding boxes distributed along the three-dimensional central axis L are identified as structured interference and removed. In this way, the internal space of the tunnel can be accurately cleared while the outline of the surrounding rock of the tunnel is well preserved.

[0068] Step 14, Sample generation;

[0069] The clean, continuous lane point cloud processed by the above steps is then segmented into multiple overlapping local point cloud patches using a sliding window strategy. Each point cloud patch is then standardized to generate training samples, including the following steps:

[0070] Step 141: Using the Farthest Point Sampling (FPS) algorithm, each point cloud block is uniformly sampled to a preset number of input points N (e.g., N=2048) to ensure the consistency of the input dimension.

[0071] Step 142: Normalize the sampled point cloud by translating it so that its centroid is located at the origin, and then scaling it so that all points are contained within a single unit sphere. This aims to eliminate differences in spatial location and scale between samples and enhance the network's generalization ability.

[0072] Step 143: Generate missing data for each point cloud patch by simulating single-viewpoint occlusion, using these missing data as standardized training samples. Specifically, a virtual viewpoint is randomly selected outside the unit sphere, and the M nearest points to this viewpoint (e.g., M=512) are removed as the ground truth. The remaining N-M points are then used as the incomplete input point cloud X. in .

[0073] Step 2: By constructing a dual-branch anisotropic-axial prior network (DAAP-Net), a dual-branch feature encoder is used to fuse local geometric features and axially dependent features to obtain the final fused feature vector representing the macroscopic structure and microscopic details of the entire tunnel point cloud.

[0074] like Figure 2 As shown, the standardized incomplete input point cloud generated in step 1 will be... The data is fed into a dual-branch feature encoder (DBE), which interprets and encodes the point cloud in parallel from both local and global dimensions. The dual-branch feature encoder includes a Local Geometry Perception Branch (LGPB) and an Axis Perception Branch (APB).

[0075] (A1) Local Geometric Perceptual Branch (LGPB)

[0076] This branch aims to accurately capture the microscopic geometric details of point clouds. Its core is a multi-layered, improved PointTransformer module, which implements the following functions:

[0077] A11, construct an anisotropic neighborhood guided by the central axis;

[0078] like Figure 3 As shown in a), the standard KNN algorithm uses an isotropic spherical search range, selecting neighboring points solely based on Euclidean distance, and cannot perceive the structural orientation of the point cloud. In contrast, the strategy of this invention introduces the three-dimensional central axis L extracted in step S1 as a macroscopic geometric representation of the "channel extension direction". The core improvement of this invention lies in neighborhood construction, such as... Figure 3 As shown in b), the macroscopic "tunnel extension direction" in this illustration is a geometric representation of the three-dimensional centerline L extracted in step 1. For each center point to be processed... In order to obtain the local extension direction of its location, this embodiment first calculates the... Find the nearest projection point of the point onto the central axis L, and take the tangent vector of the central axis at that projection point, denoted as . This vector It accurately represents the location of the alleyway. The local extension direction in the vicinity, for this vector When searching for K nearest neighbors, this invention does not use the standard Euclidean distance, but instead uses a weighted, direction-aware distance metric. :

[0079] (1)

[0080] in, ,express and Euclidean distance between them; It is a positive hyperparameter (e.g., ), used to control the intensity of anisotropy.

[0081] This distance metric is related to the direction of the roadway extension (by... (This indicates that) a distance penalty is applied to points that deviate significantly from the target area, by using... As a search metric for the K-Nearest Neighbors (KNN) algorithm, the improved Point Transformer module is able to construct a structurally more meaningful ellipsoidal anisotropic neighborhood that extends along the lane direction. .

[0082] A12, Position-aware self-attention computation;

[0083] Anisotropic neighborhoods were constructed. Then, the center point Input features and neighboring points Input features (Initially, the coordinates of a point are used; subsequent layers use features learned by the network.) This is achieved through a learnable linear projection matrix. , , These are mapped to query, key, and value vectors, respectively:

[0084] (2)

[0085] (3)

[0086] (4)

[0087] Simultaneously, calculation and neighboring points relative position coordinate vector And through a separate small Mapped to relative position encoding :

[0088] (5)

[0089] in, This represents a mapping function implemented by a small multilayer perceptron (MLP), which is responsible for mapping the input relative position coordinate vector. Transformed into the corresponding positional encoding vector .

[0090] Ultimately, the center point Update features It is the weighted sum of all value vectors in its neighborhood:

[0091] (6)

[0092] Among them, attention weight The calculation incorporates positional coding, and the formula is as follows:

[0093] (7)

[0094] in, is the dimension of the key vector; "T" indicates matrix transpose; softmax represents the normalization exponential function, which transforms a vector of arbitrary real numbers into a probability distribution vector with a sum of 1.

[0095] Through this improvement, LGPB can more intelligently focus on structurally relevant neighborhoods and extract more accurate local geometric features. Furthermore, this LGPB branch constructs a multi-resolution hierarchical feature extraction architecture by combining multi-layered stacked improved Point Transformer modules with Farthest Point Sampling (FPS), ultimately outputting multi-scale local feature maps. Specifically, starting from the initial input layer, the point cloud undergoes feature learning at each Point Transformer module layer, followed by downsampling using FPS to reduce the number of points and increase the receptive field before entering the next deeper Point Transformer module. This process is repeated, and LGPB outputs a corresponding point-by-point feature map at each resolution level s (e.g., with N, N / 2, N / 4 points). Ultimately, this branch outputs a set of feature maps from all these different resolution levels, represented as a multi-scale local feature map. .

[0096] (A2) Axial Sensing Branch (APB)

[0097] This branch aims to model the macroscopic structural continuity of the tunnel, and its functions are as follows:

[0098] A21, Serialization guided by the central axis;

[0099] For the input point cloud X in Each point in Calculate its orthogonal projection point on the three-dimensional central axis 𝐿, and obtain the arc length of the projection point along the starting point of the central axis. According to distance The size of the points is sorted, transforming the unordered point cloud set into an ordered point cloud feature sequence tensor. Where 𝐶 is the feature dimension, It is the set of real numbers.

[0100] A22 outputs axial dependency features through a Vertically Structured Convolutional Network (LSCN). ;

[0101] like Figure 4 As shown, LSCN takes a sequence tensor S as input, and its core consists of stacked one-dimensional convolutional layers. For the nth point in the sequence, its output feature in the lth convolutional layer is... From a size of convolution kernel The calculation yielded:

[0102] (8)

[0103] in, It represents the output feature of the previous layer. ReLU stands for Corrected Linear Unit activation function. This represents the learnable weight parameters of the convolution kernel at position k in the l-th convolutional layer; This represents the input feature vector at position t+k of the (l-1)th layer (i.e., the previous layer); This represents the learnable bias term of the l-th convolutional layer.

[0104] By stacking multiple layers of 1D-CNN, LSCN can effectively capture the periodic structural patterns and continuous changing trends of the tunnel along the direction of travel, ultimately outputting an axially dependent feature that summarizes macroscopic information. .

[0105] The local geometric features output by the LGPB module at the highest abstraction level (fewest points) ( It is a set One of the elements (i.e., the feature map at the highest abstraction level (with the fewest points) and the axial dependency features of the APB output. The feature channels are concatenated, and then a deep nonlinear fusion is performed through a fusion MLP consisting of multiple fully connected layers to obtain the final comprehensive fused features. :

[0106] (9)

[0107] Here, MLP() represents a multilayer perceptron network for feature fusion, which consists of at least one or more hidden layers composed of fully connected layers and nonlinear activation functions, used to perform nonlinear transformations on input features.

[0108] Step 3, guided decoding based on geometric priors, from the generated fused features The missing point cloud is recovered, and the complete point cloud is obtained. ;

[0109] This step is responsible for extracting the fusion features generated in step 2. The missing point cloud was recovered, such as Figure 2 As shown, the Guided Generator Decoder (SPGD) employs a coarse-to-fine hierarchical generation strategy, generating complete point clouds layer by layer to fill in the missing data regions. During point cloud generation, the Guided Generator introduces and utilizes the inherent geometric continuity prior of the surrounding rock of the tunnel / alley or its fixed structures as a dynamic geometric constraint to ensure that the generated complete point cloud surface is structurally smooth and geometrically logical; the specific implementation is as follows:

[0110] The guided decoder comprises multiple upsampling layers (e.g., increasing the number of points from 64 to 128 to 512). At each upsampling layer, a skip connection is established with the corresponding resolution layer in the LGPB via an adaptive feature fusion gate. This adaptive feature fusion gate is an attention-based gating unit that dynamically calculates and assigns different weights to features passed from the encoder based on the feature context of the current layer in the decoder, thereby intelligently filtering and fusing the most useful information for the current reconstruction task.

[0111] At key levels of the guided decoder (e.g., when generating the final 512 points), the Structured Prior Module (SPM) is activated to ensure the geometric regularity of the generated surface. The SPM first analyzes the normal vectors and curvature information of known points on the boundary of the region to be filled. Based on these boundary conditions, it constructs an implicit Laplacian smoothing field or Poisson field as a geometric constraint. This field guides newly generated points to tend to fall in positions that allow the entire filled surface to transition smoothly with its surroundings, effectively avoiding unreasonable bumps and depressions at the center of large missing areas.

[0112] Step 4: Training, Output, and Optimization;

[0113] This step includes the training process of the dual-branch anisotropic-axial prior network (DAAP-Net) and the output of the final point cloud.

[0114] (1) Output and merging:

[0115] During the inference phase, the guided decoder generates a completed point cloud. Incomplete point cloud from the original input The data is then merged to form the final, complete, and high-fidelity tunnel point cloud.

[0116] (2) Training and composite loss function:

[0117] To achieve end-to-end optimization of DAAP-Net, this invention employs a carefully designed composite loss function:

[0118] (10)

[0119] in:

[0120] The reconstruction loss is calculated using the multi-stage Chamfer Distance (CD). It measures the point cloud generated by the decoder at each level. Compared with the true missing point cloud The bidirectional average distance between them ensures precise alignment of the geometry;

[0121] To mitigate the adversarial loss, a discriminative network, 𝐷, is introduced, which attempts to distinguish between the real complete point cloud and the completed point cloud generated by this method. The generator (i.e., encoder-decoder) aims to generate point clouds that the discriminator 𝐷 cannot distinguish, thereby improving the visual realism and detail of the generated results.

[0122] This invention introduces a structural continuity loss. When used in conjunction with the SPM module, this loss mathematically ensures the smoothness of the generated surface. This embodiment employs a Laplacian smoothing loss, which takes the following form:

[0123] (11)

[0124] in, This is the final generated complete point cloud. It is a point The loss term emphasizes points that are significantly deviated from their neighbor centroids, thus effectively promoting the smoothness of the generated surface.

[0125] , , It is a hyperparameter used to balance the weights of the three loss terms.

[0126] By minimizing this composite total loss function, the network of this invention can learn to generate alleyway point cloud completion results that are both geometrically accurate and visually highly realistic.

[0127] The technical effects that the method of the present invention can achieve are as follows: Figure 5 As shown Figure 5The incomplete point cloud shown in (a) is processed by the method described in this invention to obtain the following: Figure 5 The completion result shown in (c) is geometrically and in detail identical to the real point cloud (e.g., Figure 5 The results show a high degree of agreement with (b) in the figure, thus verifying the effectiveness of the method of the present invention.

[0128] The above descriptions are merely some preferred embodiments of the present invention, intended to illustrate the core technology of the invention rather than to limit it. Those skilled in the art, after understanding the spirit of the present invention, can easily make various modifications, equivalent substitutions, or improvements. For example, adding, deleting, or rearranging method steps, changing the specific implementation of system components, and applying the present invention to other similar scenarios, etc., as long as these changes do not fundamentally deviate from the technical problem to be solved by the present invention and the core technical solution adopted, should all be considered to fall within the protection scope of the present invention.

Claims

1. A high-fidelity completion method for roadway point clouds based on bi-branch coding and geometric prior guidance, characterized in that, The steps include the following: S1, obtain an incomplete 3D tunnel point cloud containing missing data regions as input, and preprocess the incomplete 3D tunnel point cloud to generate standardized training samples. S2. Construct a DAAP-Net network and employ a dual-branch feature encoder to fuse local geometric features and macroscopic structural features of the tunnel. The dual-branch feature encoder includes a local geometry perception branch and an axis perception branch. The local geometry perception branch is used to extract a local geometric feature at the highest abstraction level that can characterize the details of the point cloud. The axial sensing branch is used to extract axially dependent features that characterize the topological continuity and structural patterns of the tunnel. ; local geometric features With axial dependence features The features are spliced ​​and fused along the feature channel dimension to obtain the fused features. ; S3 utilizes a guided decoder to fuse features. A coarse-to-fine strategy is adopted to generate a layered point cloud for filling in the missing data regions; The guided decoder utilizes the inherent geometric continuity of the surrounding rock of the tunnel as a dynamic geometric constraint during the point cloud generation process. S4 merges the completed point cloud generated by the guided decoder with the original incomplete input point cloud to form a structurally complete and high-fidelity alley / tunnel point cloud.

2. The high-fidelity completion method for roadway point clouds based on bi-branch coding and geometric prior guidance as described in claim 1, characterized in that, The detailed steps for generating standardized training samples include: S11 employs a statistical filtering algorithm to identify and remove outliers by analyzing the statistical characteristics of each point and its neighborhood. S12 uses a voxel downsampling algorithm to divide the entire space into cubic voxels of a preset size, and uses the three-dimensional coordinate centroid of all points in each non-empty voxel to represent the voxel, thereby obtaining a point cloud with uniform density and preserving the key geometric contours. S13 employs a filtering method guided by the roadway centerline to eliminate interference from non-main structures; S14 uses a sliding window strategy to cut the point cloud into multiple overlapping local point cloud blocks, and uses the farthest point sampling algorithm to uniformly sample each point cloud block to a preset number of input points N; then, the uniformly sampled point cloud is normalized; finally, by simulating single viewpoint occlusion, missing data is generated for each point cloud block as a standardized training sample.

3. The high-fidelity completion method for roadway point clouds based on bi-branch coding and geometric prior guidance according to claim 2, characterized in that, The detailed implementation steps of the filtering method based on the roadway centerline guidance are as follows: S131, using principal component analysis to determine the local principal direction of the roadway point cloud, and then combining B-spline curve fitting, extracting a three-dimensional central axis L from the global point cloud that can smoothly represent the overall direction of the roadway. S132, proceeds along the three-dimensional central axis L with a fixed step size. At each step position, based on the cross-sectional distribution of the local point cloud around that position, an adaptive-sized inner bounding box is dynamically constructed. This bounding box can tightly enclose the main passage space of the alley. S133 identifies all points falling into these inner bounding boxes distributed along the three-dimensional central axis L as structured interference and removes them.

4. The high-fidelity completion method for roadway point clouds based on bi-branch coding and geometric prior guidance according to claim 1, characterized in that, The local geometry-aware branch outputs multi-scale local feature maps through a multi-layered stacked improved Point Transformer module and FPS downsampling operation. ; The axis-aware branch outputs axis-dependent features through a stacked multi-layer 1D-CNN vertical structure convolutional network. ; Local geometric features With axial dependence features The features are concatenated along the feature channel dimension, and then subjected to deep nonlinear fusion through a fusion MLP consisting of multiple fully connected layers to obtain the final comprehensive fused features. : ， MLP() represents a multilayer perceptron network.

5. The high-fidelity completion method for roadway point clouds based on bi-branch coding and geometric prior guidance according to claim 4, characterized in that, Obtain local feature maps The implementation steps are as follows: SA11, constructing an anisotropic neighborhood guided by the central axis; For each center point to be processed First, calculate the Find the nearest projection point of the point onto the central axis L, and take the tangent vector of the central axis at that projection point, denoted as . The vector The representative alleyway is at the point The local extension direction in the vicinity is this vector. When searching for the K nearest neighbors, a weighted, direction-aware distance metric is used. : ， in, ,express and Euclidean distance between them; It is a positive hyperparameter used to control the intensity of anisotropy; By using As a search metric for the K-nearest neighbor algorithm, the improved Point Transformer module constructs an anisotropic neighborhood extending along the lane direction. ; SA12, computational position-aware self-attention; center point Input features and neighboring points Input features Each is achieved through a learnable linear projection matrix. , , Mapped to query vectors respectively Key vector Sum value vector : ， Simultaneously, calculation and neighboring points relative position coordinate vector And through a separate small Mapped to relative position encoding : ， in, This represents a mapping function implemented by a small multilayer perceptron (MLP); Ultimately, the center point Update features It is the weighted sum of all value vectors in its neighborhood: ， ， in, It represents the dimension of the key vector, "T" indicates matrix transpose, and softmax represents the normalized exponential function.

6. The high-fidelity completion method for roadway point clouds based on bi-branch coding and geometric prior guidance according to claim 4, characterized in that, Obtain axial dependence features The implementation steps are as follows: SA21, for the input point cloud X in Each point in Calculation points Find the orthogonal projection point on the three-dimensional central axis 𝐿, and obtain the arc length distance of this projection point along the starting point of the central axis. According to distance The size of the points is sorted, transforming the unordered point cloud set into an ordered point cloud feature sequence tensor. Where 𝐶 is the feature dimension, It is the set of real numbers; SA22 uses the sequence tensor S as the input to the axial sensing branch. For the nth point in the sequence, the output feature of the lth convolutional layer is... The expression is as follows: ， in, It represents the output features of the previous layer. ReLU stands for Modified Linear Unit Activation Function, and the convolution kernel... The size is ; By stacking multiple layers of 1D-CNN, the final output is axially dependent features. .

Images