A method and system for completing coal mine tunnel point cloud data

Abstract

The application discloses a kind of coal mine roadway point cloud data completion method and system, obtain the original point cloud data of roadway, segment and obtain roadway wall point cloud data after pretreatment;The local density of each point of point cloud is calculated, adaptive downsampling is carried out based on local density, sparse region upsampling, remove its outlier, obtain pure roadway wall point cloud data;Based on roadway central axis, segmented classification processing is carried out to point cloud data, obtain multi-class point cloud data;Based on multi-class point cloud data, construct local coordinate system, project point cloud to two-dimensional plane of local coordinate system, generate hole boundary using improved point cloud reconstruction algorithm;Based on hole boundary, carry out basic missing point sampling, obtain the multi-class point cloud data after completion;Plane fitting is carried out to the multi-class point cloud data after completion, obtain each fitting plane;Based on the intersection line of adjacent face fitting plane, carry out fine completion to corner area, obtain high-quality point cloud data, through the above whole process processing, realize the high-quality completion of complex roadway point cloud.

Description

Technical Field

[0001] This invention relates to the field of coal mine digital technology, and in particular to a method and system for completing point cloud data of coal mine roadways. Background Technology

[0002] The statements in this section are merely background information relating to this disclosure and do not necessarily constitute prior art.

[0003] With the advancement of intelligent coal mining technology, lidar point cloud technology has become a core tool for constructing 3D models of coal mine roadways. However, the underground environment of coal mines is extremely complex, with dust interference, equipment obstruction, and irregular roadway structures leading to widespread large-area gaps in the collected point cloud data. Traditional point cloud processing algorithms (such as geometric feature-based denoising methods and interpolation completion algorithms) are ill-suited to the complex morphology of coal mine roadways, and are prone to feature loss or over-smoothing when processing unstructured data, resulting in insufficient modeling accuracy.

[0004] Existing technologies have significant shortcomings. Point cloud completion lacks targeted segmentation strategies, treating the entire roadway as the completion target and failing to consider the differentiated morphological features of the top, bottom, and left and right walls. The accuracy of boundary extraction for missing regions is insufficient; traditional algorithms such as Alpha-shape are prone to missed or false detections in low-density point cloud scenarios. Furthermore, the geometric continuity between the completed point cloud and the original data is poor, leading to noticeable stitching marks in subsequent modeling. These technical deficiencies make existing solutions unable to meet the requirements of complete point cloud data for high-precision modeling of coal mine roadways, severely limiting the application value of coal mine digital twins. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a method and system for completing point cloud data of coal mine roadways, which solves the problems of point cloud density imbalance, low completion accuracy, rough boundary processing and susceptibility to interference in the prior art. Through the whole process of density control, segmentation and classification, boundary extraction and corner optimization, high-quality completion of point clouds of complex roadways is achieved.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, the present invention provides a method for completing point cloud data of coal mine roadways, comprising: The raw point cloud data of the coal mine roadway is acquired and segmented and preprocessed to obtain the point cloud data of the roadway wall. Calculate the local density of each point in the tunnel wall point cloud data, and perform adaptive downsampling and sparse region upsampling on the tunnel wall point cloud data based on the local density to obtain tunnel wall point cloud data with uniform density and remove outliers to obtain clean tunnel wall point cloud data. The tunnel centerline is generated based on the clean tunnel wall point cloud data. The point cloud data is then segmented based on the tunnel centerline to obtain multiple segment point cloud data. Based on each segment point cloud data, the local curvature of each point is calculated. The neighborhood normal vectors are smoothed and aggregated using the inverse curvature weight. The top, bottom, left, and right walls are classified according to the smoothed normal vector components to obtain multiple types of point cloud data. A local coordinate system is constructed based on the aforementioned multi-type point cloud data. The multi-type point cloud data is projected onto the two-dimensional plane of the local coordinate system, and an improved point cloud reconstruction algorithm is used to generate complete hole boundaries. Based on the hole boundary, basic missing point sampling is performed to generate two-dimensional sampling points. Residual compensation is used to map the two-dimensional sampling points to three-dimensional space, thereby obtaining the completed multi-class point cloud data. Plane fitting is performed on the completed point cloud data of various types to obtain each fitted plane; fine-grained completion is performed on the diagonal region of the intersection line of the fitted planes of adjacent planes to obtain high-quality point cloud data.

[0007] A further technical solution involves performing semantic segmentation on the original point cloud data to identify and separate the tunnel wall point cloud data and the internal component point cloud data.

[0008] A further technical solution involves adaptive downsampling of the tunnel wall point cloud data based on local density, specifically as follows: A density distribution histogram is constructed based on local density to determine the target density; Based on the ratio of local density to target density, the downsampling parameters are automatically adjusted using a dynamic voxel size calculation formula to simplify high-density areas and retain the original point cloud distribution in low-density areas.

[0009] A further technical solution involves upsampling sparse regions of the tunnel wall point cloud data based on local density, specifically as follows: A density distribution histogram is constructed based on local density to determine the target density; Statistical methods are used to identify sparse regions in point clouds; Accelerate neighborhood queries through spatial indexing, and generate new points according to the target density on the local fitting plane of sparse regions.

[0010] Further technical solutions yielded various types of point cloud data, specifically: The point cloud is dynamically divided into segments according to the curvature of the tunnel along the central axis of the tunnel, resulting in multiple segments of point cloud data. Calculate the normal vector of each point in each segment of point cloud data, and extract the X and Y components of the normal vector; Based on the X and Y components, the data is classified according to preset rules to obtain multiple types of point cloud data.

[0011] A further technical solution involves projecting multiple types of point cloud data onto a two-dimensional plane in a local coordinate system, and then using an improved point cloud reconstruction algorithm to generate complete hole boundaries. Specifically: Multiple types of point cloud data are projected onto the XOZ plane of the local coordinate system and transformed into a two-dimensional point set; Calculate the average neighborhood distance of a two-dimensional point set, setting the Alpha parameter to 1.5 times the average neighborhood distance; Triangular meshes are generated using Delaunay triangulation. The circumcircle radius of each triangle is calculated, and the edges of triangles with circumcircle radii smaller than a set value are retained as boundary edges. The boundary edges are combined into an initial boundary, and then a complete hole boundary outline is generated through polygon fitting and smoothing.

[0012] A further technical solution involves refining the point cloud data by using the diagonal region of the intersection line of adjacent fitted planes to obtain high-quality point cloud data. Specifically: Calculate the intersection line of the fitted planes of adjacent surfaces, and generate intersection points uniformly along the intersection line; Reference points are obtained from two adjacent surfaces through dynamic radius search. Supplementary points are generated on the line connecting the intersection point and the reference point to fill the missing corner area, while ensuring that the included angle of the line connecting the reference points meets the geometric constraints.

[0013] Secondly, the present invention provides a system for completing point cloud data of coal mine roadways, comprising: The data acquisition and preprocessing module is configured to: acquire the raw point cloud data of the coal mine roadway, and perform segmentation and preprocessing to obtain the roadway wall point cloud data; The density control module is configured to: calculate the local density of each point in the tunnel wall point cloud data, perform adaptive downsampling and sparse region upsampling on the tunnel wall point cloud data based on the local density, obtain tunnel wall point cloud data with uniform density and remove outliers to obtain clean tunnel wall point cloud data. The point cloud segmentation and classification module is configured as follows: generating the roadway centerline based on clean roadway wall point cloud data; segmenting the point cloud data based on the roadway centerline to obtain multiple segments of point cloud data; calculating the local curvature of each point based on each segment of point cloud data; smoothing and aggregating the neighborhood normal vectors using inverse curvature weights; and classifying the top, bottom, left, and right walls according to the smoothed normal vector components to obtain multiple categories of point cloud data. The boundary extraction module is configured to: construct a local coordinate system based on the multi-type point cloud data, project the multi-type point cloud data onto the two-dimensional plane of the local coordinate system, and generate complete hole boundaries using an improved point cloud reconstruction algorithm; The point filling generation module is configured to: sample basic missing points based on the hole boundary, generate two-dimensional sampling points, and use residual compensation to map the two-dimensional sampling points to three-dimensional space, thereby obtaining the completed multi-class point cloud data; The corner completion module is configured to: perform planar fitting on the completed multi-class point cloud data to obtain each fitted plane; and perform fine-grained completion on the corner region based on the intersection line of the fitted planes of adjacent planes to obtain high-quality point cloud data.

[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for completing point cloud data of coal mine roadways as described in the first aspect.

[0015] Fourthly, the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the method for completing point cloud data of coal mine roadways as described in the first aspect.

[0016] The above one or more technical solutions have the following beneficial effects: This invention addresses the problems of uneven point cloud density and complex missing shapes in coal mine roadways by providing a point cloud completion technology that can adaptively balance density, accurately handle complex boundaries, and effectively resist interference. Through adaptive density adjustment, segmented classification completion, and refined corner processing, it achieves high-precision restoration of missing point clouds and ultimately outputs high-density, low-noise, and high-quality point cloud data without obvious missing parts, providing high-quality data support for 3D modeling of coal mine roadways.

[0017] This invention achieves precise and efficient density balancing. It innovatively employs a density controller and a dynamic voxel downsampling algorithm, combined with Z-score sparse region detection and local planar upsampling, to achieve adaptive point cloud density balancing. Compared to traditional fixed threshold processing, it can both simplify data in dense regions and supplement feature points in sparse regions, improving the uniformity of point cloud distribution and providing a high-quality data foundation for subsequent completion. Simultaneously, through statistical outlier removal technology, it significantly improves data purity.

[0018] This invention significantly improves the accuracy and completeness of point cloud data completion. Refined classification based on normal vector components (top / bottom / left / right wall) and a local coordinate system conforming to geometric features provides a precise spatial reference for completion. An improved Alpha-shape algorithm combined with adaptive Alpha parameters effectively solves the problem of missed detection at low-density point cloud boundaries. An innovative corner region dual-reference point completion strategy (angle > 80°) specifically targets missing triangles and strip-shaped corner regions, greatly improving corner region completeness. The geometric error between the completed region and the original point cloud is controlled within 5cm, significantly outperforming traditional interpolation methods.

[0019] This invention significantly improves the accuracy and completeness of point cloud completion in coal mine roadways through fully automated processing, providing a reliable data foundation for subsequent 3D modeling and digital twin applications.

[0020] This invention provides a fully automated process for handling complex coal mine environments. From adaptive density adjustment to corner area completion, each step is designed for special scenarios such as over- and under-excavation of coal mine roadways, dust interference, and dense component density. The entire process, from raw point cloud to high-quality completed data, can be completed without manual intervention. Compared to general completion algorithms, it improves adaptability to irregular roadway morphologies by 60%, increases completion efficiency by more than 2 times, and the generated point cloud data can directly support high-precision 3D modeling, providing reliable data support for the application of coal mine digital twins in safety monitoring, engineering acceptance, and other scenarios. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is a flowchart of a method for completing point cloud data of coal mine roadways according to an embodiment of the present invention; Figure 2 This is a partial view of the original point cloud of an embodiment of the present invention. Figure 3 This is a partial view of the point cloud completion effect according to an embodiment of the present invention. Detailed Implementation

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0025] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0026] Example 1 like Figure 1As shown in the figure, this embodiment discloses a method for completing point cloud data of coal mine roadways, which includes the following steps: S1: Obtain the raw point cloud data of the coal mine roadway and perform segmentation and preprocessing to obtain the roadway wall point cloud data; In this embodiment, a Hovermap handheld LiDAR is used to scan the coal mine roadway, acquiring raw point cloud data including the roadway walls, internal components, and spatial environment. Scanning parameters are automatically adapted to the roadway environment to ensure complete point cloud coverage. In this embodiment, the scanning frequency is set to 100-200Hz, and the point cloud density is controlled at 20-50 points / ... .

[0027] Segmentation preprocessing involves semantic segmentation of the raw point cloud data. A point cloud segmentation model (PointNet++) is built using a deep learning framework. The model is pre-trained using a coal mine point cloud dataset labeled with roadway structure categories. The network parameters are then fine-tuned using a small number of labeled samples from the target roadway to obtain a trained point cloud segmentation model. The raw point cloud data is then input into the trained point cloud segmentation model for semantic segmentation, identifying and separating the roadway wall point cloud from the point clouds of internal components such as ventilation ducts, power lines, and overhead rails. This provides relatively clean roadway wall point cloud data for subsequent processing.

[0028] S2: Calculate the local density of each point in the tunnel wall point cloud data, and perform adaptive downsampling and sparse region upsampling on the tunnel wall point cloud data based on the local density to obtain tunnel wall point cloud data with uniform density and remove its outliers to obtain pure tunnel wall point cloud data. S201: Calculate the local density of each point in the tunnel wall point cloud data using the sphere radius neighborhood search algorithm, statistically analyze the point cloud density distribution characteristics, determine the target density threshold as a reference benchmark for subsequent processing, and set the density controller based on the target density threshold.

[0029] In the sphere radius neighborhood search algorithm, the sphere radius is set to 0.2-0.5m (adaptively adjusted according to the average point distance in the roadway). The number of points within this radius around each point is counted as the local density value. A density distribution histogram is established, and the target density threshold is determined (usually the 75th percentile of the density distribution is taken).

[0030] S202: Based on the ratio of local density to target density, the downsampling parameters are automatically adjusted through the dynamic voxel size calculation formula. High-density areas are appropriately simplified, while the original point cloud distribution is preserved in low-density areas, thereby reducing data redundancy and avoiding feature loss.

[0031] Based on the local density value provided by the density controller, using the formula Automatically calculate voxel size (where The basic voxel size is 0.05-0.1m. The local density at the current point; (Target density).

[0032] Adaptive downsampling of tunnel wall point cloud data is performed based on density. Specifically, dense areas with local density 1.5 times higher than the target density are considered high-density areas. A voxel grid downsampling algorithm is used for high-density areas to merge point clouds according to the calculated voxel size, reducing data redundancy. For areas with local density lower than the target density (low-density areas), no downsampling operation is performed, and the original point cloud distribution is preserved.

[0033] S203: Statistical methods are used to identify sparse regions in the point cloud, and spatial indexing is used to accelerate neighborhood queries. New points are generated on the local fitting plane of the sparse region according to the target density to balance the overall density distribution of the point cloud.

[0034] The Z-score method is used to detect sparse regions, and the Z-score value of the local density at each point is calculated. ,in For average density, The standard deviation is used to define regions with a Z-score < -1.5 as sparse regions. Sparse regions are upsampled, and a spatial index is built using KDTree to accelerate neighborhood queries. Local plane fitting is performed on the point cloud within the sparse region (using least squares to fit 30-50 neighboring points). New points are generated within the fitted plane according to the target density. The coordinates of the new points are calculated by combining the plane equation with uniform sampling to achieve density balance.

[0035] S204: Calculate the distance deviation from each point to the local plane, and remove outliers that exceed the threshold through multiple rounds of iterative filtering. These points are mainly the internal component points left over from segmentation, ensuring the purity of the retained point cloud data.

[0036] Calculate the distance deviation from each point in the point cloud after upsampling in the sparse region to its local plane, using the 3-standard-deviation criterion (distance deviation > 1). Outliers are identified, primarily from residual internal components such as pipes and cables. Outliers are removed using iterative filtering, with the local plane and standard deviation recalculated after each iteration, until the proportion of outliers falls below 2%, ensuring the purity of the retained point cloud data.

[0037] The above technical solution addresses the challenge of adapting traditional fixed-parameter processing to the uneven density of point clouds in coal mine roadways. This method achieves differentiated processing by dynamically adjusting density to "simplify dense areas and supplement sparse areas," while also specifically removing residual interference points. This not only improves data processing efficiency but also provides high-quality basic data for subsequent completion.

[0038] Compared to traditional methods, this invention abandons the traditional fixed voxel / threshold approach and calculates local density using KDTree. It dynamically adjusts voxel sizes according to the "current / target density cube root ratio," achieving precise simplification of dense areas and reasonable preservation of medium-density areas. It uses the Z-score method (mean minus 1.5 times the standard deviation) to locate sparse areas, which is more accurate in identifying low-density regions than the traditional fixed threshold approach. During upsampling, points are added as needed in the local plane of the sparse area (PCA fitting) to fit the original geometry, avoiding the structural damage caused by traditional global interpolation. The newly added points can also be color-marked for easy verification, resulting in better adaptability and balance accuracy.

[0039] S3: Generate the tunnel centerline based on the clean tunnel wall point cloud data. Segment the point cloud data based on the tunnel centerline to obtain multiple segment point cloud data. Calculate the local curvature of each point based on each segment point cloud data. Use the inverse curvature weight to smooth and aggregate the neighborhood normal vectors. Classify the top, bottom, left, and right walls according to the smoothed normal vector components to obtain multiple types of point cloud data. S301: The direction of roadway extension is determined by the bounding box of the point cloud, the centroid of the segmented point cloud is calculated to generate the initial axis point set, and cubic spline interpolation is used to fit and generate a smooth roadway centerline, which fully expresses the roadway orientation characteristics.

[0040] The minimum bounding box of the uniformly dense point cloud data of the roadway wall is calculated to obtain the length values ​​in the X, Y, and Z axes. The longest axis is determined as the principal axis (usually the roadway extension direction). A point cloud segment is taken every 5m along the principal axis, and the geometric centroid of each segment is calculated to form an initial axis point set. A cubic spline interpolation algorithm is used to fit the initial axis point set to generate a smooth and continuous roadway centerline, with the fitting error controlled within 5cm.

[0041] S302: The point cloud is dynamically divided into segments along the central axis according to the curvature of the roadway. The segment length is shortened in curved areas with greater curvature and increased in straight areas with gentle curvature to ensure that the shape of each point cloud segment is relatively regular.

[0042] Point cloud segments are divided along the central axis of the tunnel in lengths of 5-15m. The specific length is dynamically adjusted according to the curvature of the tunnel: for straight sections with curvature <0.03 / m, segments of 10-15m are used; for curved sections with curvature ≥0.03 / m, segments are shortened to 5-8m to ensure that the shape of each point cloud segment is relatively regular, which is convenient for subsequent classification and processing.

[0043] Furthermore, the steps for calculating the curvature of the tunnel are as follows: extract discrete sampling points of the central axis from the tunnel point cloud (such as the sequence of center points of each tunnel segment), fit a local circular arc or quadratic curve to three or more adjacent sampling points, and if the fitting is a circular arc, the curvature formula is... ,in, For curvature, The radius of the arc; if fitted to a spatial parametric curve, such as the parametric equation. , , Then use the differential form formula ,in, It is a cross product. For dot product, , Curvature with respect to parameters The first derivative, finally obtained The value corresponds to the curvature of that section of the tunnel.

[0044] S303: Calculate the component characteristics of the point cloud normal vector, and divide each segment of the point cloud into four categories: top, bottom, left wall, and right wall according to preset rules. The point cloud of each category can be accurately distinguished by the difference in the direction of the normal vector.

[0045] After segmentation processing in step S302, multiple segments of tunnel wall point cloud data are obtained. The normal vector of each point in each segment is calculated, and the X component (nx) and Y component (ny) of the normal vector are extracted and classified according to the following preset rules: Top point cloud: ny > 0.8 (normal vector upward, angle with gravity direction < 36°) Bottom point cloud: ny < -0.8 (normal vector downwards, angle with the opposite direction of gravity < 36°) Left wall point cloud: nx < 0 (normal vector points to the left side of the tunnel) Right wall point cloud: nx > 0 (normal vector points to the right side of the tunnel).

[0046] Furthermore, the steps for obtaining the normal vector of each point in the point cloud are as follows: First, for each point in each segment of the point cloud, a certain number of neighboring points are searched within a specified radius using KDTree. The plane is then fitted using the least squares method to obtain the initial normal vector. Then, all normal vectors are aligned with the centroid of the point cloud to ensure that the directions are consistent. Finally, the normal vector corresponding to each point is obtained.

[0047] Based on the existing normal vector classification, a dynamic weight smoothing mechanism is introduced to address the problem of abrupt changes in normal vectors on coal mine roadway walls caused by support components (such as anchor bolt trays and localized bulges in shotcrete). (1) Curvature evaluation: For each point cloud segment, the ratio of the eigenvalues ​​of the covariance matrix is ​​calculated. Obtain the local curvature at each point, where, , , These are the three eigenvalues ​​of the point cloud covariance matrix, representing the distribution variance (dispersion) of the point cloud in the three principal directions of the local coordinate system.

[0048] (2) In the k-neighborhood, instead of directly using the single-point normal vector for classification, a weighted smooth normal vector is calculated. Weight With curvature Inversely proportional.

[0049] This dynamic weight smoothing mechanism can effectively filter out normal noise caused by support residues, ensuring clear category classification logic at classification boundaries (such as the junction of sidewall and roof) and avoiding confusion in classification results.

[0050] The aforementioned technical solutions reveal significant differences in the geometric morphology of different areas within coal mine roadways, making overall processing prone to insufficient completion accuracy. By segmenting complex roadways into regular units along the central axis and combining this with normal vector features to achieve refined classification, the completion strategy for each type of point cloud becomes more targeted, effectively improving the consistency between the completion results and the actual structure.

[0051] S4: Construct a local coordinate system based on the multi-type point cloud data, project the multi-type point cloud data onto the two-dimensional plane of the local coordinate system, and use the improved Alpha-shape algorithm to generate complete hole boundaries; S401: Perform principal component analysis on each type of point cloud after classification, determine the three-axis orientation of the local coordinate system based on the distribution characteristics of the point cloud, and ensure that the coordinate system is highly consistent with the geometric features of the point cloud through orthogonalization, so as to provide a unified benchmark for subsequent two-dimensional projection.

[0052] For each type of point cloud (top, bottom, left wall, right wall) after classification, a local coordinate system is constructed using Principal Component Analysis (PCA). Specifically, the eigenvalues ​​and eigenvectors of the point cloud covariance matrix are calculated. The eigenvector corresponding to the largest eigenvalue is used as the Z-axis (the direction with the widest point cloud distribution), and the eigenvector corresponding to the smallest eigenvalue is used as the X-axis (perpendicular to the Z-axis and conforming to the point cloud plane). The Y-axis is calculated by the cross product of the X-axis and Z-axis, forming a local coordinate system. The coordinate system is orthogonalized to ensure that the three axes are pairwise perpendicular and normalized, providing a unified reference for subsequent two-dimensional projection.

[0053] Furthermore, when constructing a local coordinate system for the classified point cloud (top, bottom, left wall, right wall) using PCA, the eigenvalues ​​and eigenvectors need to be centered for this type of point cloud (subtracting the mean of all point coordinates), and then the covariance matrix of the decentralized point cloud matrix is ​​calculated. The eigenvalues ​​and corresponding eigenvectors of the covariance matrix are obtained through matrix operations. The eigenvalues ​​represent the distribution variance of the point cloud in the direction of the corresponding eigenvector (the larger the value, the more dispersed the point cloud is in that direction), and the eigenvectors are mutually orthogonal unit vectors (representing the main direction of the point cloud distribution). Therefore, the eigenvector corresponding to the largest eigenvalue is taken as the Z-axis with the widest distribution of the point cloud, and the eigenvector corresponding to the smallest eigenvalue is taken as the X-axis with the most concentrated distribution. The Y-axis is obtained by cross product of the X-axis and Z-axis. After orthogonalization and normalization, a local coordinate system is formed, providing a reference for subsequent two-dimensional projection.

[0054] S402: Project the point cloud onto a two-dimensional plane of the local coordinate system, adaptively set algorithm parameters according to the point cloud density, generate complete hole boundary contours through triangulation and boundary filtering, and eliminate small isolated boundary segments to ensure accurate definition of missing areas.

[0055] An improved point cloud reconstruction algorithm (improved Alpha-shape algorithm) is used to extract the hole boundary. Specifically, multiple types of point cloud data are projected onto the XOZ plane of the local coordinate system and transformed into a two-dimensional point set. The average neighborhood distance of the two-dimensional point set is calculated, and the Alpha parameter is set to 1.5 times the average neighborhood distance. A triangular mesh is generated through Delaunay triangulation, and the circumcircle radius of each triangle is calculated. Triangle edges with circumcircle radii less than a set value (1 / Alpha in this embodiment) are retained as boundary edges. The boundary edges are combined into an initial boundary, and a complete hole boundary contour is generated through polygon fitting and smoothing (removing isolated boundary segments with a length less than 0.5m).

[0056] With the above technical solution, traditional boundary extraction algorithms have fixed parameters, which are prone to missing or false detection of boundaries in low-density or noisy point cloud scenarios. This invention significantly improves the recognition accuracy of hole boundaries by dynamically adjusting parameters based on point cloud density and combining boundary optimization processing, and is especially suitable for complex point cloud environments in coal mine roadways.

[0057] Compared to the original open-source Alpha-shape boundary extraction algorithm, the improved Alpha-shape algorithm of this invention calculates the average neighbor distance of the point cloud using K-nearest neighbors and uses its reciprocal as the alpha value, achieving adaptive parameter adjustment and avoiding the experience-based reliance on manually preset alpha in traditional algorithms. A protection mechanism is added to the area calculation during triangle selection, avoiding numerical anomalies caused by minimal or collinear triangles by taking the maximum value, thus enhancing computational stability. Multiple extracted polygons are merged to generate a single geometric object, simplifying subsequent processing. Furthermore, a complete functional chain from parameter calculation and boundary extraction to difference set region sampling is constructed, expanding the algorithm's application in real-world scenarios. These improvements make the code more adaptable and practical when processing complex point clouds.

[0058] S5: Based on the hole boundary, sample the basic missing points to generate two-dimensional sampling points, extract the three-dimensional depth information of the known points of the hole boundary, construct the depth gradient field, and use residual compensation to map the two-dimensional sampling points to three-dimensional space so that the supplemented points fit the undulating shape of the tunnel wall, thereby obtaining the completed multi-type point cloud data. S501: Within the extracted hole boundary range, two-dimensional sampling points are generated uniformly according to the target density. Collision detection is used to remove sampling points that overlap with residual interference points to ensure the effectiveness of the supplementary points.

[0059] Based on the hole boundary contour, basic missing point sampling is performed. Specifically, uniform sampling is conducted within the difference region between the extracted hole boundary polygon and the point cloud bounding box, with the sampling interval set to 0.05-0.2m according to the target density. Two-dimensional candidate points (two-dimensional sampling points) are generated using a random sampling algorithm. Collision detection is used to remove candidate points that overlap with residual points of internal components (a small number of points that have not been completely removed), ensuring the effectiveness of the sampling points.

[0060] The classification of 2D sampling points depends on the "classification processing logic": instead of generating all 2D sampling points first and then determining the category, the process of "local coordinate system construction → hole boundary extraction → 2D sampling" is executed separately for each type of point cloud (such as top, left wall, etc.). For example, when processing the left wall point cloud, the generated 2D sampling points themselves belong to the "left wall category". When calculating the Y coordinate later, the category attribute of the left wall point cloud can be directly used (taking the average Y coordinate of the left wall). This ensures that the category of each 2D sampling point is consistent with the category of the corresponding point cloud being processed, and the supplementary points can accurately fit the geometric shape of the point cloud of that type.

[0061] S502: Construct a coordinate transformation matrix based on local coordinate system parameters to map two-dimensional sampling points to three-dimensional space. Calculate the three-dimensional coordinates of the supplementary points according to the point cloud category characteristics to ensure that the spatial position of the supplementary points (two-dimensional sampling points) is continuous with that of the original point cloud.

[0062] Based on the constructed local coordinate system parameters, a coordinate transformation matrix is ​​established from two-dimensional sampling points (x,z) to three-dimensional space (X,Y,Z). Matrix operations are then used to map the two-dimensional points to the three-dimensional space. Specifically, the Y-coordinate is calculated based on the point cloud category characteristics: the top point cloud uses the maximum value of the Y-axis in the local coordinate system, the bottom point cloud uses the minimum value, and the left and right wall point clouds use the average Y-coordinate of the corresponding plane, ensuring that the supplemented points conform to the geometry of the original point cloud.

[0063] Furthermore, the local coordinate system parameters include three-axis unit vectors (direction vectors of the local X, Y, and Z axes, which have been orthogonalized and normalized) and the origin of the local coordinate system. When establishing the coordinate transformation matrix, the spatial relationship between the local coordinate system and the global coordinate system (the coordinate system where the original point cloud resides) is used as the basis: first, a 3×3 rotation matrix is ​​constructed, with each column corresponding to the unit vectors of the local X, Y, and Z axes in the global coordinate system, used to achieve direction transformation; then, a 1×3 translation vector (i.e., the coordinates of the local coordinate system origin in the global coordinate system) is added to achieve position translation; finally, a 4×4 homogeneous coordinate transformation matrix is ​​formed (adapting to linear transformations of points in three-dimensional space). For a two-dimensional sampling point (x, z), it is first completed as a three-dimensional coordinate (x, Y, z, where Y is a fixed value calculated according to the category) in the local coordinate system, and then mapped to a three-dimensional coordinate (X, Y, Z) in the global coordinate system through the operation of "local three-dimensional coordinates × transformation matrix," completing the transformation from two-dimensional to three-dimensional.

[0064] Furthermore, the Y-coordinate is calculated based on the point cloud category characteristics. The "characteristics" essentially refer to the inherent geometric attributes of each type of point cloud: for example, the local Y-axis direction of the top point cloud corresponds to the upward direction of the tunnel height, and its maximum Y-axis value is the actual height position of the top wall; the average Y-coordinate of the left and right wall point clouds corresponds to the middle height of the wall. These attributes are the core geometric features that distinguish this type of point cloud from other types, and determine the spatial position where the supplementary points should fit.

[0065] Furthermore, taking the average value of the Y coordinate (depth) can lead to the completed area being too flat and inconsistent with the actual over- or under-excavation pattern of the tunnel. This invention introduces residual compensation mapping, which makes the completed point cloud macroscopically present a "micro-arc" or "undulation" consistent with the original tunnel, completely solving the geometric splicing traces between the completed area and the original data.

[0066] The residual compensation mapping is as follows: Extract the 3D depth attribute of the hole boundary line in the two-dimensional projection plane XOZ; By using radial basis functions (RBF), the depth variation trend at the boundary is diffused towards the center of the hole to construct a local depth residual field. ; Coordinates are reconstructed based on the local depth residual field. The 3D coordinate mapping formula for the supplementary points is modified as follows:

[0067] Among them, the complete point cloud data is shown. The reference plane height of the point cloud (e.g., the average sidewall height). This is the correction amount calculated based on the wall undulation trend.

[0068] S6: Perform plane fitting on the completed multi-class point cloud data to obtain each fitted plane; perform fine-grained completion on the diagonal area of ​​the intersection line of the fitted planes of adjacent planes to obtain high-quality point cloud data.

[0069] S601: Perform planar fitting on the completed point cloud, and use a pass-through filter to remove overflow points outside the planar threshold range. Crop the excess area and optimize the point cloud boundary to ensure that the point cloud boundary is consistent with the actual tunnel structure.

[0070] For each type of point cloud after completion, plane fitting is performed (using the RANSAC algorithm) to determine the plane equation; a pass-through filtering method is used to remove points outside the threshold range on both sides of the plane (the threshold is set to ±0.3m), eliminating boundary overflow points caused by over-excavation and under-excavation, and ensuring that the point cloud boundary is consistent with the actual roadway structure.

[0071] S602: Calculate the intersection line of the fitted planes of adjacent faces, generate intersection points uniformly along the intersection line, obtain reference points from the two adjacent faces through dynamic radius search, generate supplementary points on the line connecting the intersection point and the reference point to fill in the missing corner area, and at the same time ensure that the included angle of the line connecting the reference points meets the geometric constraints.

[0072] Calculate the intersection line of the fitted plane of adjacent faces (such as top and left wall, left wall and bottom, etc.), and generate intersection points at 0.1m intervals within the current point cloud segment. For each intersection point, search for neighboring points with an initial radius of 0.1m. If the number of points on both adjacent faces within this radius is less than 3, gradually expand the search radius (maximum 10m) until at least one reference point is obtained from each of the two faces, and the angle between the lines connecting the two points is greater than 80°. Perform fine-grained completion on the corner area, generating supplementary points at 0.05m intervals on the lines connecting the intersection point and the two reference points to fill in the triangular or strip-shaped missing corner areas. Traverse all intersection points to complete the full corner area completion, ensuring the complete geometric shape of the corner area.

[0073] Through the above technical solutions, the corner areas of coal mine roadways often have special missing shapes due to scanning blind spots. Traditional interpolation methods are difficult to restore their true geometric features. This invention ensures that the supplementary points fit the shape of the corner area by constraining the included angle of the two reference points, and solves the problem of sparse neighboring points by dynamic radius search. It is specifically designed for the missing corner areas of coal mine roadways and significantly improves the completeness and accuracy of the supplementation results.

[0074] Compared to traditional corner completion methods, this invention overcomes the limitations of traditional fixed threshold judgment of missing points. It uses a two-dimensional approach of "normal vector direction + distance" to determine the true missing points in corner areas, avoiding misjudgments caused by uneven density. It abandons fixed radius search and adopts a dynamically expanded radius (maximum 5m) and limits the search to adjacent faces, prioritizing distant reference points to solve the problems of not finding points in sparse areas and finding the wrong points in dense areas. When completing points, an "80° angle constraint" is added to fit the right-angle characteristics of the alley corner area, avoiding the excessive smoothing of traditional interpolation. At the same time, segmented interpolation ensures seamless connection between the completed points and the original point cloud, resulting in better completion accuracy and scene adaptability.

[0075] In summary, through fully automated processing, the transformation of coal mine roadway point clouds from raw data to high-quality completion results was achieved. The geometric continuity between the completed area and the original point cloud is good, and the features of complex parts such as corners are completely preserved, providing reliable data support for subsequent 3D modeling and digital twin applications. This invention fully leverages the advantages of multi-algorithm collaboration, effectively solving the problems of low accuracy and poor adaptability in coal mine roadway point cloud completion, and is suitable for point cloud processing scenarios in various complex coal mine environments.

[0076] This invention uses a Hovermap handheld LiDAR to collect raw point cloud data of coal mine roadways. It achieves adaptive balance of point cloud density through a density controller and removes residual points of internal components by combining statistical filtering. It achieves fine partitioning based on point cloud segmentation and normal vector features and uses an improved Alpha-shape algorithm to accurately extract hole boundaries. Through a strategy that combines basic point filling with corner area-specific completion, it achieves complete restoration of missing areas and finally outputs high-density, low-noise, and high-quality point cloud data without obvious missing points.

[0077] Traditional methods of constructing coal mine roadway meshes using the Poisson reconstruction algorithm often result in mesh distortion, missing holes, and coarse boundary treatment. This invention proposes a completely new reconstruction process. Compared to traditional solutions, this invention addresses the point cloud density imbalance caused by dust and occlusion in coal mine roadways by innovatively using dynamic density control (high-density downsampling and sparse area upsampling) to achieve uniform density, avoiding redundancy or feature loss associated with traditional fixed thresholds. It breaks through the limitations of traditional overall processing by segmenting by roadway curvature and classifying by normal vectors, making the completion more closely match the differentiated shapes of the top, bottom, and walls. The improved Alpha-shape algorithm (dynamic Alpha parameters) solves the problem of missed and false detections at traditional boundaries. An innovative corner completion strategy (intersection guidance + angle constraint) accurately restores corner missing areas that traditional interpolation cannot handle, and the entire process is automated with a completion error of ≤5cm. It improves adaptability by 60% and efficiency by 2 times, making it more suitable for complex coal mine scenarios.

[0078] Example 2 This embodiment discloses a system for completing point cloud data of coal mine roadways, including: The data acquisition and preprocessing module is configured to: acquire the raw point cloud data of the coal mine roadway, and perform segmentation and preprocessing to obtain the roadway wall point cloud data; The density control module is configured to: calculate the local density of each point in the tunnel wall point cloud data, perform adaptive downsampling and sparse region upsampling on the tunnel wall point cloud data based on the local density, obtain tunnel wall point cloud data with uniform density and remove outliers to obtain clean tunnel wall point cloud data. The point cloud segmentation and classification module is configured as follows: generating the roadway centerline based on clean roadway wall point cloud data; segmenting the point cloud data based on the roadway centerline to obtain multiple segments of point cloud data; calculating the local curvature of each point based on each segment of point cloud data; smoothing and aggregating the neighborhood normal vectors using inverse curvature weights; and classifying the top, bottom, left, and right walls according to the smoothed normal vector components to obtain multiple categories of point cloud data. The boundary extraction module is configured to: construct a local coordinate system based on the multi-type point cloud data, project the multi-type point cloud data onto the two-dimensional plane of the local coordinate system, and generate complete hole boundaries using an improved point cloud reconstruction algorithm; The point filling generation module is configured to: sample basic missing points based on the hole boundary, generate two-dimensional sampling points, and use residual compensation to map the two-dimensional sampling points to three-dimensional space, thereby obtaining the completed multi-class point cloud data; The corner completion module is configured to: perform planar fitting on the completed multi-class point cloud data to obtain each fitted plane; and perform fine-grained completion on the corner region based on the intersection line of the fitted planes of adjacent planes to obtain high-quality point cloud data.

[0079] Example 3 The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of Embodiment 1.

[0080] Example 4 The purpose of this embodiment is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the method of Embodiment 1.

[0081] The steps and methods involved in the apparatuses of Embodiments 3 and 4 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0082] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0084] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for completing point cloud data of coal mine roadways, characterized in that, include: The raw point cloud data of the coal mine roadway is acquired and segmented and preprocessed to obtain the point cloud data of the roadway wall. Calculate the local density of each point in the tunnel wall point cloud data, and perform adaptive downsampling and sparse region upsampling on the tunnel wall point cloud data based on the local density to obtain tunnel wall point cloud data with uniform density and remove outliers to obtain clean tunnel wall point cloud data. The tunnel centerline is generated based on the clean tunnel wall point cloud data. The point cloud data is then segmented based on the tunnel centerline to obtain multiple segment point cloud data. Based on each segment point cloud data, the local curvature of each point is calculated. The neighborhood normal vectors are smoothed and aggregated using the inverse curvature weight. The top, bottom, left, and right walls are classified according to the smoothed normal vector components to obtain multiple types of point cloud data. A local coordinate system is constructed based on the aforementioned multiple types of point cloud data. The multiple types of point cloud data are projected onto the two-dimensional plane of the local coordinate system, and an improved point cloud reconstruction algorithm is used to generate complete hole boundaries. Based on the hole boundary, basic missing point sampling is performed to generate two-dimensional sampling points. Residual compensation is used to map the two-dimensional sampling points to three-dimensional space, thereby obtaining the completed multi-class point cloud data. Plane fitting is performed on the completed point cloud data of various types to obtain each fitted plane; fine-grained completion is performed on the diagonal region of the intersection line of the fitted planes of adjacent planes to obtain high-quality point cloud data.

2. The method for completing point cloud data of coal mine roadways as described in claim 1, characterized in that, The segmentation preprocessing involves semantic segmentation of the original point cloud data to identify and separate the point cloud data of the tunnel wall and the point cloud data of the internal components.

3. The method for completing point cloud data of coal mine roadways as described in claim 1, characterized in that, Adaptive downsampling of tunnel wall point cloud data based on local density is performed as follows: A density distribution histogram is constructed based on local density to determine the target density; Based on the ratio of local density to target density, the downsampling parameters are automatically adjusted using a dynamic voxel size calculation formula to simplify high-density areas and retain the original point cloud distribution in low-density areas.

4. The method for completing point cloud data of coal mine roadways as described in claim 1, characterized in that, Sparse regions of the tunnel wall point cloud data are upsampled based on local density, specifically as follows: A density distribution histogram is constructed based on local density to determine the target density; Statistical methods are used to identify sparse regions in point clouds; Accelerate neighborhood queries through spatial indexing, and generate new points according to the target density on the local fitting plane of sparse regions.

5. The method for completing point cloud data of coal mine roadways as described in claim 1, characterized in that, Multiple types of point cloud data were obtained, specifically: The point cloud is dynamically divided into segments according to the curvature of the tunnel along the central axis of the tunnel, resulting in multiple segments of point cloud data. Calculate the normal vector of each point in each segment of point cloud data, and extract the X and Y components of the normal vector; Based on the X and Y components, the data is classified according to preset rules to obtain multiple types of point cloud data.

6. The method for completing point cloud data of coal mine roadways as described in claim 1, characterized in that, Multiple types of point cloud data are projected onto a two-dimensional plane in a local coordinate system, and an improved point cloud reconstruction algorithm is used to generate complete hole boundaries, specifically: Multiple types of point cloud data are projected onto the XOZ plane of the local coordinate system and transformed into a two-dimensional point set; Calculate the average neighborhood distance of a two-dimensional point set, setting the Alpha parameter to 1.5 times the average neighborhood distance; Triangular meshes are generated using Delaunay triangulation. The circumcircle radius of each triangle is calculated, and the edges of triangles with circumcircle radii smaller than a set value are retained as boundary edges. The boundary edges are combined into an initial boundary, and then a complete hole boundary outline is generated through polygon fitting and smoothing.

7. The method for completing point cloud data of coal mine roadways as described in claim 1, characterized in that, Fine-grained completion is performed on the diagonal region of the intersection line of adjacent fitted planes to obtain high-quality point cloud data, specifically: Calculate the intersection line of the fitted planes of adjacent surfaces, and generate intersection points uniformly along the intersection line; Reference points are obtained from two adjacent surfaces through dynamic radius search. Supplementary points are generated on the line connecting the intersection point and the reference point to fill the missing corner area, while ensuring that the included angle of the line connecting the reference points meets the geometric constraints.

8. A system for completing point cloud data of coal mine roadways, characterized in that, include: The data acquisition and preprocessing module is configured to: acquire the raw point cloud data of the coal mine roadway, and perform segmentation and preprocessing to obtain the roadway wall point cloud data; The density control module is configured to: calculate the local density of each point in the tunnel wall point cloud data, perform adaptive downsampling and sparse region upsampling on the tunnel wall point cloud data based on the local density, obtain tunnel wall point cloud data with uniform density and remove outliers to obtain clean tunnel wall point cloud data. The point cloud segmentation and classification module is configured as follows: generating the roadway centerline based on clean roadway wall point cloud data; segmenting the point cloud data based on the roadway centerline to obtain multiple segments of point cloud data; calculating the local curvature of each point based on each segment of point cloud data; smoothing and aggregating the neighborhood normal vectors using inverse curvature weights; and classifying the top, bottom, left, and right walls according to the smoothed normal vector components to obtain multiple categories of point cloud data. The boundary extraction module is configured to: construct a local coordinate system based on the multi-type point cloud data, project the multi-type point cloud data onto the two-dimensional plane of the local coordinate system, and generate complete hole boundaries using an improved point cloud reconstruction algorithm; The point filling generation module is configured to: sample basic missing points based on the hole boundary, generate two-dimensional sampling points, and use residual compensation to map the two-dimensional sampling points to three-dimensional space, thereby obtaining the completed multi-class point cloud data; The corner completion module is configured to: perform planar fitting on the completed multi-class point cloud data to obtain each fitted plane; and perform fine-grained completion on the corner region based on the intersection line of the fitted planes of adjacent planes to obtain high-quality point cloud data.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the method for completing point cloud data of coal mine roadways as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the method for completing point cloud data of coal mine roadways as described in any one of claims 1-7.

Images