A three-dimensional point cloud-based slope surface crack quantification analysis system and method

Abstract

This invention relates to the field of crack image recognition and analysis technology, and discloses a quantitative analysis system and method for slope surface cracks based on three-dimensional point clouds. The method includes acquiring and preprocessing three-dimensional point cloud image data of the slope; analyzing local geometric features to determine attribute levels; introducing a depth continuity verification mechanism and combining it with normal vector consistency analysis to identify real cracks; extracting a three-dimensional skeleton using a geodesic distance iterative refinement algorithm and calculating all-element dimensions such as length, opening, and depth; and conducting risk assessment based on volumetric damage and distribution density to generate a refined risk distribution heatmap. This application can effectively eliminate interference from gravel and misjudgment of rock bedding, achieving topological reconstruction and three-dimensional quantitative analysis of complex interwoven crack networks, thereby improving the accuracy of slope crack identification and providing precise support for real-time early warning and reinforcement of slope stability.

Description

Technical Field

[0001] This invention relates to the field of crack image recognition and analysis technology, and more specifically, to a quantitative analysis system and method for slope surface cracks based on three-dimensional point clouds. Background Technology

[0002] With the continuous deepening of the demand for refined management of geological disaster prevention and control and safe operation and maintenance of infrastructure throughout its entire life cycle, especially in high-risk scenarios such as water conservancy hubs, high slopes of highways and mine spoil heaps, the need for large-scale, high-timeliness, and comprehensive accurate quantification of the geometric parameters of surface cracks on slopes is becoming increasingly urgent.

[0003] Traditional crack analysis methods based on two-dimensional images and discrete detection have made some progress in crack pixel segmentation, surface geometric parameter extraction, and multi-platform data acquisition. Even with the introduction of high-density three-dimensional point cloud technology, systematic identification errors caused by the interruption of point cloud depth due to the filling of small gravel particles in the crack depth region, as well as mis-extraction problems caused by the high overlap between the original stone bedding texture and mineral particle boundaries in the three-dimensional point cloud geometric properties and cracks, have not been effectively solved. These problems make it difficult to meet the generalized application needs of slope scenarios with different lithologies, scales, and morphological complexities.

[0004] Therefore, how to integrate high-density 3D point cloud video stream data, automatically respond to particle filling interference and stone texture misidentification, and realize an end-to-end closed-loop analysis method from automatic crack extraction, 3D geometric modeling to multi-dimensional quantitative index generation has become a key problem that urgently needs to be solved in the field of intelligent slope monitoring. Summary of the Invention

[0005] This invention provides a quantitative analysis system and method for slope surface cracks based on three-dimensional point clouds, which solves the technical problems in the prior art of integrating high-density three-dimensional point cloud video stream data, automatically dealing with particle filling interference and misidentification of stone texture, and realizing an end-to-end closed-loop analysis method from automatic crack extraction, three-dimensional geometric modeling to multi-dimensional quantitative index generation.

[0006] This invention provides a system and method for quantitative analysis of surface cracks on slopes based on three-dimensional point clouds, including:

[0007] Firstly, a method for quantitative analysis of surface cracks on slopes based on three-dimensional point clouds includes:

[0008] High-density three-dimensional point cloud video stream data of the slope surface is acquired, and the high-density three-dimensional point cloud data is preprocessed to obtain a standardized point cloud image set.

[0009] Analyze the local geometric features of each frame of the point cloud in the point cloud image set to obtain the point cloud geometric attribute type;

[0010] Based on the point cloud geometric attribute type, the corresponding attribute level is set to obtain the morphological features of the crack candidate region;

[0011] Based on the morphological features, identify real crack entities on the slope surface in the point cloud image set;

[0012] Among them, a depth continuity verification mechanism is introduced for the candidate crack region. By analyzing the comparison between the local fluctuation standard deviation of the point cloud depth value in the candidate region and the particle size threshold of the filling material, pseudo crack regions with discontinuous depth abruptness caused by particle accumulation are eliminated. Combining point cloud normal vector consistency analysis and multi-scale curvature difference characteristics, texture structures with periodic repetition patterns and real cracks with random extension direction and local depression features are distinguished. Double filtering is performed by setting normal vector deflection angle threshold and curvature abruptness gradient threshold to accurately obtain real crack entities.

[0013] Extract the three-dimensional skeleton of the real crack entity, and calculate the three-dimensional geometric dimensions and crack distribution density of the real crack entity;

[0014] The algorithm employs an iterative refinement method based on geodesic distance to peel off crack voxels layer by layer, preserving topological connectivity while suppressing redundant branches. Cross-node detection is performed on the extracted skeleton, and the continuity of local tangent directions is used as the criterion for classifying each branch path, decomposing the skeleton into several independent crack path units with clear orientations. The crack network topology is reconstructed based on the endpoint connections of each path unit, and the hierarchical relationship between the main crack and secondary branch cracks is statistically analyzed to calculate the three-dimensional geometric dimensions (length, width, and depth) and crack distribution density of each crack path.

[0015] Based on the three-dimensional geometric dimensions and the crack distribution density, a risk assessment is conducted on the slope surface to obtain the slope stability risk assessment results.

[0016] Furthermore, high-density three-dimensional point cloud video stream data of the slope surface is acquired, and the high-density three-dimensional point cloud video stream data is preprocessed to obtain a standardized point cloud image set, including:

[0017] A three-dimensional laser scanning device was used to perform a full-coverage scan of the slope surface to obtain raw point cloud data containing spatial three-dimensional coordinates and reflection intensity.

[0018] Spatial pass-through filtering is performed on the raw point cloud data, and a preset range of coordinate axes is set to remove background point clouds located outside the monitoring area;

[0019] A statistical filtering algorithm is used to denoise the remaining point cloud. The average distance between each point and a preset number of points in its neighborhood is calculated, and points whose average distance exceeds a preset deviation threshold are identified as outlier noise points and removed.

[0020] The voxel downsampling algorithm is applied to resample the denoised point cloud, dividing the point cloud space into tiny voxels of a preset size, and replacing the original points in each voxel with the centroid of all points in that voxel.

[0021] The point cloud data of each station, after the above filtering and resampling processing, is mapped to a preset unified global coordinate system by the iterative nearest-point algorithm, thereby completing the alignment of the spatial reference and generating the standardized point cloud image set.

[0022] Furthermore, the local geometric features of each frame of the point cloud in the point cloud image set are analyzed to obtain the point cloud geometric attribute type, including:

[0023] For each point to be processed in the point cloud image set, search its neighborhood point set within a preset search radius;

[0024] Construct the covariance matrix of the neighborhood point set, and calculate the eigenvalues ​​of the covariance matrix, wherein the eigenvalues ​​include a first eigenvalue, a second eigenvalue, and a third eigenvalue;

[0025] The local surface change rate is calculated by dividing the third characteristic value by the sum of the three characteristic values.

[0026] The local tangent plane of the neighborhood point set is fitted using the least squares method, and the angle deviation between the normal vector of the point to be processed and the normal vector of the local tangent plane is calculated.

[0027] Based on the local surface change rate and the included angle deviation, the geometric properties of the point cloud are divided into multiple geometric property types.

[0028] Further, based on the point cloud geometric attribute type, corresponding attribute levels are set and morphological features of the crack candidate region are obtained, including:

[0029] Based on various geometric attribute types, attribute levels are set according to edge abruptness and indentation depth. The attribute levels include: flat region level with low edge abruptness and shallow indentation depth, transition region level with medium edge abruptness and moderate indentation depth, suspected crack region level with high edge abruptness and deep indentation, and high-confidence crack region level with extremely high edge abruptness and continuous indentation extension. Further, by combining the connectivity, length-to-width ratio, and curvature change trend of crack candidate regions, high-confidence crack regions that meet preset thresholds are selected, and false crack segments caused by noise or local disturbances are eliminated, finally generating a crack spatial distribution map with clear structure and semantics.

[0030] Based on the attribute level, points belonging to the suspected crack region level and the high confidence crack region level are filtered to obtain a candidate point set that meets the crack geometric features. The crack geometric features include: abrupt change in local curvature of the point cloud, normal vector deflection angle exceeding a preset angle threshold, linear continuous length along the main extension direction exceeding a preset length threshold, and cross-sectional width perpendicular to the extension direction within a preset width range.

[0031] Using candidate points with high curvature as seed points, the candidate point set is extended in all directions according to preset normal vector angle thresholds and curvature deviation thresholds, and points that satisfy geometric continuity are clustered into crack clusters.

[0032] Principal component analysis was performed on the crack clusters to extract the principal direction vector, secondary direction vector, and normal vector of each crack cluster, so as to obtain the morphological features.

[0033] Furthermore, based on the morphological features, identifying actual crack entities on the slope surface within the point cloud image set includes:

[0034] A depth continuity verification mechanism is introduced for the crack candidate region. The local fluctuation standard deviation of the point cloud depth value in the candidate region is calculated. The local fluctuation standard deviation is compared with the preset filling particle size threshold. False crack regions with discontinuous depth caused by the accumulation of small gravel particles are eliminated, and crack candidate regions with depth continuity verification are obtained.

[0035] The particle size threshold of the filler is determined according to the lithology of the slope. For fractured slopes, the value range is 3-5 times the average spacing of the point cloud. In this invention, it is set to 0.01dm-0.01m to ensure that centimeter-level particle jumps can be identified.

[0036] For the candidate crack regions verified by deep continuity, a geometric co-occurrence matrix reflecting the local spatial correlation of the point cloud is constructed based on the normal vector in the morphological features. By calculating the energy, entropy, contrast and correlation of the geometric co-occurrence matrix, feature vectors reflecting the difference between crack texture and background surface are extracted.

[0037] Based on the aforementioned feature vectors, and combining point cloud normal vector consistency analysis and multi-scale curvature difference characteristics, a normal vector deflection angle threshold and a curvature abrupt change gradient threshold are set to perform dual filtering on stone texture structures with periodic repetition patterns and real cracks with random extension directions and local depression characteristics.

[0038] The feature vector is input into a preset classification model, and the discrimination threshold determined by the dual filtering is used to distinguish between plant shading, surface soil and real cracks, eliminate false crack point clouds, and identify the real crack entities.

[0039] Further, the three-dimensional skeleton of the real crack entity is extracted, and the three-dimensional geometric dimensions and crack distribution density of the real crack entity are calculated, including:

[0040] A three-dimensional solid model is created for the identified real crack entities, and a mesh model of the crack surface is generated using the Delaunay triangulation algorithm.

[0041] An iterative refinement algorithm based on geodesic distance is used to peel off the crack voxels in the mesh model layer by layer, suppressing redundant branches while preserving topological connectivity, and generating a three-dimensional skeleton line composed of single-point chains.

[0042] The three-dimensional skeleton lines are subjected to intersection node detection. The branch paths are assigned based on the continuity of local tangent direction, and the skeleton is decomposed into several independent crack path units with clear directions.

[0043] The crack network topology is reconstructed based on the endpoint connection relationship of the independent crack path unit, the hierarchical relationship between the main crack and the secondary branch is statistically analyzed, and the three-dimensional skeleton line is smoothed by the Laplace smoothing algorithm to remove burrs on the skeleton.

[0044] Based on the processed 3D skeleton line, the true 3D length of the crack is calculated using the path integral method.

[0045] Search for the edge points of the real crack entity in the normal plane of the three-dimensional skeleton line, and calculate the average Euclidean distance between the edge point pairs to obtain the crack opening value.

[0046] Along the crack extension direction indicated by the three-dimensional skeleton line, calculate the projected displacement of the actual crack entity in the direction perpendicular to the slope surface to obtain the crack depth value.

[0047] The total number of real crack entities within the preset monitoring area is counted, and the crack coverage ratio per unit area is calculated to obtain the crack distribution density.

[0048] Furthermore, a refined analysis is performed on the node regions with interwoven cracks and complex branches in the three-dimensional skeleton lines, including:

[0049] At the intersection node, extract the local tangent vectors of all converging branches, cluster the local tangent vectors according to angular similarity, and determine the candidate potential continuous crack paths passing through the intersection node.

[0050] For each candidate path, calculate the curvature change and tangent direction deflection angle of the skeleton line before and after crossing the intersection node. The path with the smallest curvature change and the smallest tangent direction deflection angle is determined as the main extension direction, and the remaining converging branches are determined as secondary branches.

[0051] Record the intersection angles of the secondary branches and the main crack paths, as well as the spatial coordinates of the connecting nodes, to construct a crack intersection network that includes primary and secondary hierarchical relationships and intersection position information;

[0052] A global topology consistency check is performed on the fracture cross network to eliminate path loops and isolated short paths caused by misjudgment of local nodes, resulting in a fracture path network with a complete topology and clear hierarchy.

[0053] Furthermore, based on the three-dimensional geometric dimensions and the crack distribution density, a risk assessment is performed on the slope surface to obtain the slope stability risk assessment results, including:

[0054] The calculated true three-dimensional length, the opening value, and the depth value are multiplied to obtain the geometric volume damage of the crack.

[0055] Based on the geometric volume damage and the crack distribution density, a slope health index model is constructed using a weighted scoring method.

[0056] The values ​​output by the slope health index model are compared with multiple preset risk classification thresholds to determine whether the slope is in any of the following states: safe, under monitoring, under early warning, or in danger, thus obtaining the risk assessment result.

[0057] Furthermore, through inter-frame dynamic analysis and spatial hazard mapping, a spatiotemporally combined refined risk distribution map of the slope surface is generated, including:

[0058] Inter-frame difference detection is performed on the continuously acquired high-density three-dimensional point cloud video stream data, the spatial displacement between adjacent frame point clouds is calculated, and key frames whose surface deformation exceeds a preset dynamic threshold are selected.

[0059] The point cloud data corresponding to the key frame is differentially processed with the initial reference frame to extract the expansion increment of the crack region in the time dimension and construct a crack dynamic evolution sequence.

[0060] Based on the dynamic evolution sequence of the cracks, the three-dimensional size expansion rate of the cracks and the frequency of new cracks appearing within a preset monitoring period are calculated to obtain the crack activity level of each area on the slope surface.

[0061] Based on the crack network topology, the intersection nodes of multiple crack paths and the high-density clustering areas of cracks are identified, and the intersection nodes and high-density clustering areas are marked as potential priority damage points.

[0062] The spatial coordinates of the potential priority failure points are superimposed and mapped with the corresponding real three-dimensional length, opening value and depth value to generate a crack hazard distribution heat map indexed by the spatial location of the slope surface.

[0063] Based on the aforementioned hazard distribution heatmap and the aforementioned crack activity level, a refined risk distribution map of the slope surface, including safety, monitoring, early warning, and hazard zones, is output to assist in the formulation of differentiated slope reinforcement schemes and key monitoring area deployment strategies.

[0064] Secondly, a slope surface crack quantification analysis system based on three-dimensional point clouds includes:

[0065] Data acquisition and preprocessing module: used to acquire high-density three-dimensional point cloud video stream data of the slope surface and perform preprocessing to obtain a preprocessed point cloud image set;

[0066] Feature analysis module: used to analyze the local geometric features of each frame of point cloud in the point cloud image set to obtain the point cloud geometric attribute type; according to the point cloud geometric attribute type, set the corresponding attribute level to obtain the morphological features of the crack candidate region.

[0067] Crack identification module: used to identify real crack entities on the slope surface in the point cloud image set based on the morphological features;

[0068] Crack statistics module: used to extract the three-dimensional skeleton of the real crack entity and calculate the three-dimensional geometric dimensions and crack distribution density of the real crack entity;

[0069] Risk assessment module: used to conduct risk assessment on the slope surface based on the three-dimensional geometric dimensions and the crack distribution density, so as to obtain the slope crack risk assessment result.

[0070] The beneficial effects of this invention are as follows: By introducing a depth continuity verification mechanism, this invention calculates the local fluctuation standard deviation of the point cloud depth value within the crack candidate region and compares it with a preset infill particle size threshold, thereby automatically eliminating false crack regions caused by particle filling and effectively eliminating missed detections and depth quantization errors. Addressing the problem of mis-extraction caused by the high overlap between the original stone bedding texture and mineral particle boundaries and the geometric properties of cracks, a dual filtering mechanism based on normal vector consistency analysis and multi-scale curvature difference features accurately distinguishes between stone texture structures with periodic repetition patterns and real cracks with random extension directions and local depressions, improving the accuracy and reliability of crack identification. Regarding the problem of overlapping skeleton extraction caused by multiple interwoven cracks and complex branch structures, an iterative method based on geodesic distance is used... The algorithm is refined and the intersection node classification is used to decompose the complex skeleton into independent path units with clear directions, and the crack network topology is reconstructed to achieve hierarchical management of main cracks and secondary branches. On this basis, based on the path integral of the 3D skeleton line and the edge search of the normal plane, the true 3D length, opening value and depth of each crack are accurately calculated. Combined with the crack distribution density, a slope health index model is constructed to support the quantitative determination of four levels of risk status: safe, monitoring, early warning and dangerous. The overall solution uses high-density 3D point cloud video stream as a unified data source to realize end-to-end closed-loop analysis from automatic crack extraction, 3D geometric modeling to multi-dimensional quantitative index generation and risk assessment. It does not require manual intervention or prior assumptions and meets the generalized application needs of slope scenarios with different lithology, scale and morphological complexity. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of a slope surface crack quantification analysis method based on three-dimensional point cloud provided in an embodiment of the present invention;

[0072] Figure 2 This is a schematic diagram of a slope surface crack quantification analysis system module based on three-dimensional point cloud provided in an embodiment of the present invention. Detailed Implementation

[0073] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0074] At least one embodiment of the present invention discloses a slope surface crack quantification analysis system and method based on three-dimensional point clouds, including:

[0075] like Figure 1 As shown, a method for quantitative analysis of slope surface cracks based on three-dimensional point clouds includes the following steps:

[0076] Step 1: Acquire high-density 3D point cloud video stream data of the slope surface and perform preprocessing to obtain a preprocessed point cloud image set;

[0077] Step 2: Analyze the local geometric features of each frame of the point cloud in the point cloud image set to obtain the point cloud geometric attribute type; based on the point cloud geometric attribute type, set the corresponding attribute level to obtain the morphological features of the crack candidate region.

[0078] Step 3: Based on the morphological features, identify the actual crack entities on the slope surface in the point cloud image set;

[0079] Step 4: Extract the three-dimensional skeleton of the real crack entity, and calculate the three-dimensional geometric dimensions and crack distribution density of the real crack entity;

[0080] Step 5: Based on the three-dimensional geometric dimensions and the crack distribution density, conduct a risk assessment of the slope surface to obtain the slope crack risk assessment results.

[0081] The method of this invention is based on system module execution. This system consists of five functional modules working together: a data acquisition and preprocessing module, a feature analysis module, a crack identification module, a crack statistics module, and a risk assessment module. Specifically, the data acquisition and preprocessing module acquires high-density 3D point cloud video stream data of the slope surface and performs preprocessing to obtain a standardized point cloud image set; the feature analysis module analyzes the local geometric features of each frame of the point cloud image set to obtain the point cloud geometric attribute type, and sets the corresponding attribute level according to the geometric attribute type, thereby obtaining the morphological features of the crack candidate region; the crack identification module identifies real crack entities on the slope surface in the point cloud image set based on the morphological features; the crack statistics module extracts the 3D skeleton of the real crack entities and calculates the 3D geometric dimensions and crack distribution density of the real crack entities; the risk assessment module performs a risk assessment of the slope surface based on the 3D geometric dimensions and crack distribution density to obtain the slope crack risk assessment result. These five modules are sequentially connected through a high-speed data bus, forming an end-to-end processing pipeline from point cloud acquisition to risk assessment result output.

[0082] The first step is data acquisition and preprocessing. In this embodiment, a 3D laser scanning device is used to perform a full-coverage scan of the target slope surface. This 3D laser scanning device is typically mounted on a stable platform opposite the slope. It acquires raw point cloud data containing spatial 3D coordinates and reflection intensity by emitting laser pulses and receiving reflected signals. Due to the complex terrain of the slope, multi-site scanning is usually required to eliminate occlusions and shadows. After receiving this raw data, the data acquisition and preprocessing module first performs spatial pass-through filtering. This process truncates and removes background point clouds located outside the slope monitoring area, such as distant mountains, vegetation, or sky points, by setting a preset range for the coordinate axes in the software system, thereby reducing the data load for subsequent calculations.

[0083] Subsequently, the data acquisition and preprocessing module employs a statistical filtering algorithm to denoise the remaining point cloud. In actual operation, the system traverses each point in the point cloud, quickly finding its 20 nearest neighbors using a KD-tree index, and calculating the average Euclidean distance from that point to these 20 neighbors. Based on the distance distribution of the entire point cloud set, the global average distance μ and standard deviation σ are calculated. If the average Euclidean distance of a point exceeds the deviation threshold formed by μ + 2σ, it is identified as an outlier noise point caused by ambient light interference or electronic noise from equipment, and a physical removal operation is performed. To further optimize data density and preserve feature information, a voxel downsampling algorithm is applied to resample the denoised point cloud. The system divides the point cloud space into tiny voxel grids with a side length of 0.02m, calculates the centroid coordinates of all points within each voxel, and uses these centroid coordinates to replace the original point cloud points within that voxel grid, thereby significantly reducing the amount of data while preserving the geometric contour of the slope surface. Finally, by iterating the nearest point algorithm, the rotation and translation matrix between the cloud data of each site is calculated, and the processed multi-site cloud data is mapped to a preset unified global coordinate system to complete the alignment of the spatial reference and generate a standardized point cloud image set.

[0084] The system then proceeds to the local geometric feature extraction and morphological analysis stage. For each point in the point cloud image set, the feature analysis module uses a KD-tree search algorithm to find its neighborhood point set within a search radius of 0.15m, requiring at least five effective neighborhood points to ensure the statistical stability of the covariance matrix construction. This step is fundamental to all subsequent geometric analyses. The system constructs the covariance matrix of the neighborhood point set, which reflects the spatial dispersion of the local point cloud. By performing eigenvalue decomposition on the covariance matrix, the first, second, and third eigenvalues ​​are calculated and ordered. These three eigenvalues ​​have clear physical meanings: when the first and second eigenvalues ​​are much larger than the third eigenvalue, it indicates a planar distribution of the local point cloud; when the three eigenvalues ​​are close in magnitude, it indicates a spherical distribution. The system performs a division operation using the third eigenvalue as the numerator and the sum of the three eigenvalues ​​as the denominator to obtain the local surface change rate, which directly reflects the surface roughness or curvature. Simultaneously, the least squares method is used to fit the local tangent plane of the neighborhood point set, and the angular deviation between the normal vector of the point to be processed and the normal vector of the local tangent plane is calculated. The feature analysis module uses the local surface change rate and the angular deviation as input parameters, and classifies the geometric attributes of the point cloud into flat, edge, concave, and abrupt types according to a preset geometric feature classification table. These attribute types provide key clues for subsequent identification of crack edges and grooves.

[0085] After extracting local geometric features, the feature analysis module further sets corresponding attribute levels based on the obtained geometric attribute types. Specifically, the system extracts edge abruptness data and indentation depth data for each point. Edge abruptness is determined by the gradient of the normal vector, and indentation depth is determined by the distance from the point to the fitted plane. Based on these data values, four attribute levels are set: the flat region level corresponds to points with a local surface change rate below 0.10 and a normal vector deflection angle less than 5°; the transition region level corresponds to points with a local surface change rate of 0.10-0.30 and a normal vector deflection angle of 5°-15°; the suspected crack region level corresponds to points with a local surface change rate of 0.30-0.60 and a normal vector deflection angle of 15°-30°; and the high-confidence crack region level corresponds to points with a local surface change rate greater than 0.60 and a normal vector deflection angle greater than 30°. The system filters out points belonging to the suspected crack region level and the high-confidence crack region level, and extracts local curvature data, normal vector deflection angle data, linear continuous length data, and cross-sectional width data from the point cloud. This will satisfy the condition that the local principal curvature is greater than 0.50m. -1Candidate points with a normal vector deflection angle greater than 15°, a linear continuous length greater than 0.05m along the main extension direction, and a cross-sectional width perpendicular to the extension direction within the range of 0.002m-0.10m are converged to obtain a candidate point set that conforms to the geometric characteristics of a crack. Specifically, the crack geometric characteristics are: abrupt changes in local curvature of the point cloud (principal curvature greater than 0.50m). -1 The criteria for cracks include: a normal vector deflection angle exceeding 15°; a linear continuous length exceeding 0.05m along the main extension direction; and a cross-sectional width perpendicular to the extension direction within the range of 0.002m-0.10m. Furthermore, the system performs a secondary screening by combining the spatial connectivity, length-to-width ratio, and curvature variation trend of the candidate regions to eliminate pseudo-crack segments caused by noise or local disturbances, ultimately generating a spatially distributed crack map with clear structure and semantics. To form a complete crack structure, a region growing algorithm is executed on the candidate point set. The system uses a principal curvature exceeding 0.50m. -1 Candidate points are used as initial seed points, and topological extension is performed in all directions based on a 15° threshold for the angle between the neighboring normal vectors and a 30% threshold for the relative curvature deviation. If the angle between the normal vectors of a neighboring point and the seed point is within 15° and the relative curvature deviation does not exceed 30%, it is incorporated into the current crack cluster. Finally, principal component analysis is performed on the formed crack clusters to extract the principal direction vector, secondary direction vector, and normal vector of each crack cluster to obtain morphological features reflecting the crack extension trend.

[0086] The crack identification module is the core verification link of this invention. Its main task is to eliminate interference. The whole adopts a three-stage serial logic of rule pre-screening, multi-dimensional feature extraction and model fine judgment. The three stages are executed sequentially, and the output of the previous stage is directly used as the input of the next stage. There is no feature feedback relationship between the dual filtering and the random forest classification model. The filtering threshold is not used as the model input feature, but as a hard decision boundary to complete the rapid elimination.

[0087] The first stage involves depth continuity verification and adaptive dual-rule filtering. The system first performs depth continuity verification: calculating the standard deviation of the depth values ​​in the candidate region relative to the local fitted plane. If the standard deviation of depth fluctuation exceeds the infill particle size threshold of 0.01m and the abrupt distribution exhibits random discontinuous characteristics (i.e., the variance of the depth difference between adjacent points is greater than the normal fluctuation range of a continuous concave crack), then pseudo-crack regions with discontinuous depth value abrupt changes due to gravel accumulation are identified and eliminated, resulting in crack candidate regions verified for depth continuity. Subsequently, the system automatically selects a reference block with a flat region attribute level in the current frame point cloud (ensuring the selected block does not contain cracks or vegetation), and calculates the mean μref and standard deviation σref of the distribution of the normal vector deflection angle and curvature gradient of all points within this reference block. The system then dynamically determines the normal vector deflection angle threshold and curvature abrupt gradient threshold under the current scanning conditions based on μref+2σref. The aforementioned adaptive mechanism enables the threshold to be automatically adjusted according to lithology and point cloud accuracy: under standard granite-like hard lithology and a scanning accuracy of approximately 0.01m point spacing, the dynamically calculated reference values ​​are approximately 25° and 0.80m, respectively. -1 / m; For soft lithologies such as schist or sandstone with well-developed bedding, the overall statistical distribution of the reference block shifts, and the dynamic threshold is usually adjusted upwards to 30°-35°; When the point cloud point spacing increases to more than 0.03m, the local stability of curvature estimation decreases, and the curvature gradient threshold needs to be adjusted upwards by 20%-30% accordingly. The system performs dual filtering on candidate regions verified by depth continuity: for regions where the normal vector deflection angle is lower than the dynamic threshold and the curvature gradient is lower than the dynamic threshold, they are identified as stone textures with periodic arrangement characteristics and are removed; regions exceeding the dynamic threshold are retained as candidate samples for the second stage.

[0088] The second stage involves multi-dimensional feature vector extraction. For the crack candidate regions retained after dual-rule filtering, a geometric co-occurrence matrix is ​​constructed based on their normal vector distribution data, and four texture parameters—energy, entropy, contrast, and correlation—are calculated. Simultaneously, the directional entropy of the neighborhood point cloud normal vectors, the curvature difference between the coarse (0.05m) and fine (0.02m) scales, the depth fluctuation ratio (the ratio of the local fluctuation standard deviation to the depth mean), and the linear continuity score (continuously satisfying a principal curvature greater than 0.50m along the main extension direction) are calculated. -1 The number of condition points together constitute an 8-dimensional feature vector. This feature vector fully encodes the texture structure and geometric extension characteristics of the candidate region, providing sufficient discriminative information for the third-stage classification.

[0089] The third stage is fine-grained classification using random forests. The aforementioned 8-dimensional feature vectors are input into a pre-trained random forest classification model to achieve three-class classification and labeling of vegetation occlusion, surface soil, and real cracks. Typical feature vectors of vegetation-occupied areas exhibit high depth undulation ratios (set to be greater than 0.30), low linear continuity scores (the number of points continuously satisfying the condition along the main extension direction is usually less than 3), and high geometric co-occurrence matrix entropy values, reflecting irregular surface randomness. Based on this, the model distinguishes it from real cracks with continuous depth. Typical features of surface soil areas include low curvature differences (curvature difference between coarse and fine scales is close to zero) and high normal vector consistency (extremely low directional entropy, close to a flat surface). However, its geometric co-occurrence matrix correlation parameters show discriminatory differences from both the rock base and real cracks, leading the model to remove it from the crack candidate pool. Real crack areas exhibit high curvature differences, directional entropy within a specific range (reflecting consistency along the extension direction rather than random scattering), and a continuous concave distribution of depth undulations. Ultimately, the model classifies these as real crack entities and outputs the labeling results.

[0090] The crack statistics module is responsible for performing high-precision 3D measurements on identified real crack entities. First, the system models the real crack entities in 3D, generating a triangular mesh model of the crack surface using the Delaunay triangulation algorithm. This step transforms the discrete point cloud into a continuous topological surface. Then, an iterative refinement algorithm based on geodesic distance is used to extract the skeleton. This algorithm differs from traditional Euclidean distance stripping; it calculates the shortest path distance from each point to the boundary along the mesh surface, i.e., the geodesic distance. Based on the geodesic distance field, the system strips crack voxels within the mesh model layer by layer inward until a single-point chain reflecting the crack center direction is retained, generating a 3D skeleton line. Subsequently, the system iteratively smooths the 3D skeleton line using the Laplace smoothing algorithm, removing burrs caused by discrete noise on the skeleton and ensuring the geometric continuity of the skeleton line. For the common phenomenon of multiple interwoven cracks and complex branching in complex slopes, the crack statistics module performs refined intersection node detection. Local tangent vectors of all converging branches are extracted at the intersection nodes. An angle similarity clustering algorithm is used to determine potential continuous crack path candidates passing through the intersection nodes. For each path candidate, the curvature change of the skeleton line before and after crossing the intersection node and the tangent direction deflection angle are calculated. The path with the smallest curvature change and the smallest tangent direction deflection angle is determined as the main extension direction, and the remaining converging branches are determined as secondary branches. The intersection angles of the secondary branches with the main crack path and the spatial coordinates of the connecting nodes are recorded to construct a crack intersection network containing the primary and secondary hierarchical relationships and intersection position information. The crack intersection network is subjected to global topology consistency verification to eliminate path loops and isolated short paths caused by local node misjudgments, resulting in a crack path network and topology graph with complete topology and clear hierarchy.

[0091] After obtaining a refined skeleton and topological relationships, the crack statistics module begins to calculate the full-element three-dimensional geometric dimensions of the cracks. Using a path integral calculation algorithm, the Euclidean distances between adjacent points on the processed three-dimensional skeleton line are accumulated to obtain the true three-dimensional length of the crack. A normal tangent plane is generated at each node of the skeleton line, and the edge boundary points of the crack entity are searched within this tangent plane. The average Euclidean distance between relative edge point pairs is calculated to obtain the crack opening value. Simultaneously, the corresponding depth value is calculated for each node on the skeleton line. Because the slope is a spatial curved surface, the slope normal vector at each location continuously changes with the terrain undulations. A global horizontal plane or a single fixed normal vector cannot be used as a unified projection reference; otherwise, elevation differences caused by terrain undulations such as protruding rock blocks and local folds will be mixed into the crack depth measurement results. To this end, the system independently performs the following local reference plane extraction and projection calculation process for each skeleton node: First, with the current skeleton node as the center, all background points not identified as crack points (i.e., points whose attribute level belongs to the flat area level or transition area level, which represent complete slopes without cracking on both sides) are searched within a neighborhood of 0.20m. The least squares method is used to fit a local tangent plane to the above set of background points. The normal vector nlocal of the resulting plane is the local slope reference normal vector at the node, which can adaptively reflect the actual orientation of the slope at the current position without being affected by distance. The system addresses several issues: First, it considers the interference from topography. Second, it searches for the edge and bottom points of crack entities within the normal tangent plane (a section determined by the tangent direction of the skeleton and nlocal) at the skeleton node. The bottom point Pbottom, furthest from the local tangent plane along the nlocal direction, is used as the starting point for depth measurement, and the vertical projection of Pbottom onto the local tangent plane is used as the ending point. Finally, the vertical distance from Pbottom to the local tangent plane is calculated, which is the depth value D at that skeleton node. Its physical meaning is the true indentation depth of the crack cavity bottom relative to the surrounding intact slope. Through this node-by-node local reference plane adaptive method, the elevation difference caused by slope topographic undulations is explicitly absorbed by the local tangent plane fitting process. The depth value only reflects the true cavity depth of the crack relative to its adjacent background slope, thus achieving accurate quantification of crack depth on complex undulating rock slopes. Furthermore, the system statistically analyzes the total number of real crack entities within the preset monitoring area and calculates the crack coverage ratio per unit area to obtain the crack distribution density.

[0092] The risk assessment module performs in-depth slope stability analysis based on the aforementioned geometric quantification results. First, the true three-dimensional length Li, opening value Wi, and depth Di of the i-th crack are multiplied to obtain the geometric volume damage Vi = Li × Wi × Di (unit: m³), ​​which serves as a physical indicator for directly quantifying the cavity volume of a single crack. Subsequently, the slope health index Hidx is calculated using the following four-factor normalized weighted scoring method, the formula consisting of three steps: normalization, weighted summation, and health mapping.

[0093] The first step is to normalize each factor. Using reference critical length L0=2.0m, reference critical aperture W0=0.05m, reference critical depth D0=0.10m, and reference critical crack coverage ratio ρ0=5% as benchmarks (each reference critical value is determined based on domestic industry standards for slope geological disaster prevention and control and statistical analysis of historical typical instability cases), the normalized mean values ​​of the three-dimensional geometric parameters of all N cracks in the monitoring area are calculated for each dimension: normalized mean length L=(1 / N)×Σmin(Li / 2.0,1), normalized mean aperture W=(1 / N)×Σmin(Wi / 0.05,1), normalized mean depth D=(1 / N)×Σmin(Di / 0.10,1); normalized crack coverage ratio ρ=min(ρ / 5%,1); when the measured value exceeds the corresponding reference critical value, it is truncated to 1.0 to prevent a single extreme crack or extreme local area from dominating the score.

[0094] The second step is weighted summation. The comprehensive damage index S is calculated using the following formula: S = wL×L + wW×W + wD×D + wρ×ρ, where the weights of each factor are: wL = 0.15, wW = 0.25, wD = 0.30, and wρ = 0.30, with a sum of 1.00. The weight allocation follows the following engineering judgment criteria: the depth value has the highest weight (0.30), because the crack depth directly determines the thickness loss of the rock mass structure and is the most direct driving factor for instability; the crack distribution density has the same weight as the depth value (0.30), because a large-area crack group, even if the depth of a single crack is limited, may lead to overall slope instability; the crack opening has a moderate weight (0.25), reflecting the degree of rock mass separation and the risk of hydraulic fracturing; and the length has the lowest weight (0.15) to avoid excessively lowering the score due to long and shallow surface cracks.

[0095] The third step is health index mapping. The comprehensive damage index is mapped to a health index of 100 points using Hidx = 100 × (1-S): when there are no cracks in the monitored area, W = D = ρ = 0, Hidx = 100, indicating a healthy slope; when all parameters reach the reference critical value, S = 1.00, Hidx = 0, indicating a severely damaged slope. The system logically compares Hidx with the risk classification threshold: a health index greater than 80 points indicates a safe state; a health index between 60 and 80 points indicates a monitoring state; a health index between 40 and 60 points indicates a warning state; and a health index below 40 points indicates a dangerous state, automatically triggering a warning and pushing an alarm notification. The initial values ​​of the aforementioned reference critical values ​​and weighting coefficients are determined based on industry standards. During the engineering deployment phase, it is recommended to collect measured parameters and actual stability assessment conclusions of the target slope during the initial monitoring period, and to perform online correction of wL, wW, wD, wρ and each reference critical value through regression analysis to adapt to specific lithology, geological structure and engineering background.

[0096] To achieve dynamic monitoring, the risk assessment module also performs inter-frame dynamic analysis. The system performs inter-frame differential detection on continuously acquired high-density 3D point cloud video stream data and calculates the spatial displacement vector between point clouds in adjacent time frames. If the displacement modulus of a certain area exceeds the dynamic threshold of 2mm, the time frame is marked as a critical time frame. The system performs spatial subtraction on the point cloud of the critical time frame and the initial reference frame to extract the deformation expansion increment of the crack area on the time axis and construct a crack dynamic evolution time series. Based on this series, the system calculates the 3D size expansion rate of the crack and the frequency of new crack occurrence within a preset monitoring period (usually 7 days as an analysis period): an expansion rate below 0.1mm / day is judged as stable; 0.1-0.5mm / day is judged as active; and more than 0.5mm / day is judged as accelerated, directly triggering an early warning response; thus, the crack activity level of each area on the slope surface is obtained. Based on the crack network topology map, the system identifies the intersection nodes and high-density clustering areas of multiple crack paths and marks the intersection nodes and high-density clustering areas as potential priority failure points. Subsequently, the spatial coordinates of potential priority failure points are overlaid and mapped with their corresponding actual three-dimensional length, aperture value, and depth value to generate a crack hazard distribution heatmap indexed by the spatial location of the slope surface. Finally, the risk assessment module performs a spatiotemporal joint analysis based on the hazard distribution heatmap and the aforementioned crack activity levels, outputting a refined risk distribution map of the slope surface that includes safety, monitoring, early warning, and hazard zoning, providing engineers with an intuitive decision-making basis that considers both spatial distribution and dynamic evolution trends.

[0097] In practical applications, the system operates as follows: A 3D laser scanner scans mine or highway slopes at preset time intervals. The data acquisition and preprocessing module receives the data stream in real time and performs automatic registration and noise reduction. The feature analysis module extracts local geometric features and performs morphological analysis on millions of point cloud data points to quickly locate suspected crack areas. The crack identification module automatically filters out debris falling into the cracks and interference from the rock's texture through depth continuity verification and geometric co-occurrence matrix analysis. The crack statistics module then intervenes, dissecting the complex crack network to extract the length, opening, and depth of each crack. The risk assessment module continuously compares the scan results at different time points. If it detects an abnormally rapid opening speed of a crack or a trend of multiple cracks connecting to each other, the system immediately renders the area as a red danger level on the risk distribution heatmap and triggers an early warning signal.

[0098] The modules of this invention communicate sequentially in series via a high-speed data bus. The output of the data acquisition and preprocessing module is connected to the input of the feature analysis module; after completing local geometric feature extraction and morphological analysis, the feature analysis module outputs the morphological features to the crack identification module; the output of the actual crack identification results from the crack identification module is connected to the crack statistics module; and the output of the three-dimensional geometric dimensions and crack distribution density from the crack statistics module is connected to the risk assessment module. This five-module pipeline structure cascades clearly defines the responsibilities of each processing stage and the data flow direction, enabling efficient processing of large-scale three-dimensional point cloud video stream data and output of slope crack risk assessment results.

[0099] The geodesic distance-based iterative refinement algorithm in this embodiment has advantages in handling slope surfaces with complex undulations. Since slope surfaces are not planar, traditional Euclidean distance ignores the surface undulations when calculating the skeleton, causing the skeleton to deviate from the actual crack center. Geodesic distance, however, is calculated along the triangular mesh surface, perfectly matching the natural morphology of the slope. During the iterative refinement process, the system constructs a distance field based on the geodesic distance from each voxel point to the crack boundary. Higher-priority boundary points are peeled off layer by layer, while points on the ridge of the distance field are retained. This method not only ensures the topological connectivity of the skeleton but also greatly suppresses redundant burr branches caused by noise, making the final extracted crack length closer to the physical truth.

[0100] Furthermore, the introduction of a depth continuity verification mechanism solves the long-standing problem of gravel effect in point cloud crack identification. In actual slopes, weathered gravel often falls into cracks, which appears as local depth bulges in the 3D point cloud. If judged solely by depth information, these bulges would be considered crack interruptions. This invention identifies these unnatural jumps caused by discrete particles by analyzing the standard deviation of local undulations, and restores crack continuity through interpolation or smoothing algorithms, thereby ensuring the integrity of crack identification.

[0101] The random forest classification model used in the crack identification module underwent systematic offline training and validation. The training process was as follows: The training dataset consisted of labeled samples from various lithological slopes (covering typical engineering lithologies such as granite, schist, limestone, and sandstone). At least 5000 crack samples and at least 5000 non-crack samples (including interference categories such as vegetation obstruction, surface soil, and stone texture) were collected and divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The input feature dimensions were energy, entropy, contrast, and correlation calculated based on the geometric co-occurrence matrix, superimposed with normal vector direction entropy and multi-scale curvature differences (coarse scale 0.05m, fine scale 0.02m), resulting in an 8-dimensional feature vector. The model consisted of 100 decision trees, with a maximum depth of 15 layers per tree and a minimum number of samples for node splitting of 5. A 5-fold cross-validation strategy was used to fine-tune the hyperparameters to prevent overfitting. On the independent test set, the model's crack category recognition accuracy is no less than 92% and recall is no less than 88%. After meeting the accuracy requirements for engineering applications, the trained model parameters are solidified and saved for use as pre-trained weights during system deployment.

[0102] This embodiment also provides a computing device, including one or more processors and a storage device. The storage device stores a computer program, which, when executed by the processor, enables the implementation of all the above-described method steps. This device can be integrated into the field processing unit of a slope monitoring station to achieve on-site data processing and real-time early warning. Simultaneously, this invention also provides a computer-readable storage medium, on which a program stored, when executed, can drive a hardware system to complete the entire process from point cloud acquisition to risk heat map generation.

[0103] In the system's risk assessment module, the slope health index model fully considers the geometric evolution characteristics of cracks. The model assigns weights of 0.15, 0.25, 0.30, and 0.30 to the four factors of length, aperture, depth, and distribution density, respectively, ensuring that each factor contributes to the final health index through a uniquely determined calculation path, without ambiguity or gaps in implementation. Because the model independently calculates normalized parameters for each crack frame-by-frame and updates Hidx in real time, it can capture the evolutionary trend of cracks over time: for example, a crack with a small absolute aperture but a rapidly increasing W value between two monitoring periods will see its contribution to W×W rise rapidly, thus lowering Hidx; while an older crack with a large aperture that has stabilized will have a stable impact on Hidx because W no longer changes. This weighting mechanism based on the normalized component mean allows the model to assess both static damage levels and the dynamic evolution risk through continuous frame-by-frame comparisons. Spatial hazard mapping technology transforms the output of Hidx per spatial grid cell into a color gradient heatmap, clearly showing the weak points on the slope surface and guiding the precise implementation of reinforcement projects.

[0104] This invention, through the synergistic cooperation of the aforementioned modules, achieves a fully automated process from macroscopic identification to microscopic quantification of slope cracks. By introducing key technologies such as depth continuity verification, geometric co-occurrence matrix texture filtering, and geodesic distance-based skeleton extraction, the accuracy of crack quantification in complex field environments is improved. The system not only provides static crack parameters but also reveals the dynamic evolution of slope stability through spatiotemporal joint analysis, providing strong technical support for the early detection and scientific prevention of geological disasters.

[0105] To enable those skilled in the art to fully understand and implement this invention, the following supplements the specific implementation principle of this invention with a specific application scenario for monitoring steep slopes in open-pit mines.

[0106] Step 1, Data Acquisition and Refined Preprocessing, involves first setting up a 3D laser scanner on a monitoring platform opposite the mine slope. The data acquisition and preprocessing module controls the scanner to perform multi-station scans of the rock face at risk of cracking. Due to the numerous protruding rock blocks and depressions on the slope surface, single-station scans generate significant occlusion shadows. Fusing multi-station point cloud data compensates for these missing geometric information. After the data enters the system, the data acquisition and preprocessing module utilizes spatial pass-through filtering principles to physically crop irrelevant point clouds (such as ore transport roads and distant mountains) outside the scan range based on preset slope boundary coordinates, thus concentrating computational resources on the target rock face area. Subsequently, by executing an iterative nearest-point algorithm, the system finds overlapping corresponding feature points in the point clouds of adjacent stations. By minimizing the sum of squared Euclidean distances between corresponding point pairs, a precise rotation and translation matrix is ​​calculated, achieving high-precision alignment of multi-source point clouds in a unified spatial coordinate system. This process eliminates positional errors caused by multi-station splicing, laying a high-precision spatial reference for subsequent millimeter-level crack width extraction.

[0107] Step 2, Local Geometric Feature Extraction and Candidate Region Locking: The feature analysis module performs microscopic geometric attribute analysis on the aligned point cloud image set. The system constructs a local covariance matrix for each sampling point and obtains three eigenvalues ​​reflecting the local spatial distribution through eigenvalue decomposition. In flat rock areas, the third eigenvalue approaches zero, while when the scan line sweeps across the crack edge, the abrupt change in depth causes a sharp increase in the local surface change rate. The feature analysis module utilizes this principle to calculate the angular deviation between the normal vector and the local fitting plane, thereby marking points with high curvature and high deflection characteristics as suspected crack points. Subsequently, the feature analysis module further utilizes the region growing principle, using these suspected points as seeds to search outwards based on the consistency of the normal vectors. When the normal vector deflection gradient of neighboring points remains within a preset evolution threshold, the system merges them into the same crack cluster. This growth method based on geometric topology can connect discrete suspected points into crack candidates with continuous physical morphology, effectively distinguishing isolated noise points from potential cracks with linear extension characteristics.

[0108] Step 3, Complex Interference Removal and Real Crack Verification: The crack identification module uses a three-stage sequential process—pre-screening according to rules, multi-dimensional feature extraction, and model refinement—to progressively eliminate complex interference from the mine slope. In real cracked cavities, the depth value exhibits a regular concavity with spatial location, while fallen gravel within the crack causes chaotic and discontinuous jumps in the point cloud depth. In the first stage, the crack identification module calculates the standard deviation of the depth value fluctuation within the candidate region and logically compares it with the gravel particle size threshold of 0.01m, automatically removing pseudo-crack regions with random and discontinuous depth jumps, thus completing the initial removal of gravel interference. Subsequently, the system automatically selects a flat reference block in the current frame point cloud, calculates the mean and standard deviation of its normal vector deflection angle and curvature gradient distribution, and dynamically determines a dual filtering threshold based on the mean plus twice the standard deviation—under the condition of a granite rock wall and a point spacing of 0.01m in this mine, the dynamic thresholds are approximately 25° and 0.80m, respectively. -1 In the first stage, a dual-rule filtering process is performed on candidate regions verified for depth continuity. Regions with both normal vector deflection angle and curvature gradient below the dynamic threshold are identified as stone textures and removed. In the second stage, an 8-dimensional feature vector is extracted from the candidate regions retained by rule filtering. This vector includes texture parameters of geometric co-occurrence matrix (energy, entropy, contrast, correlation), normal vector direction entropy, multi-scale curvature difference, depth fluctuation ratio, and linear continuity score. The dual-filter threshold is not used as the feature input model but only as a pre-selective hard decision boundary. In the third stage, the aforementioned 8-dimensional feature vectors are input into the pre-trained random forest classification model: areas with plant shading are eliminated due to their high depth fluctuation ratio (set to be greater than 0.30), low linear continuity score (less than 3 points that continuously meet the conditions), and high geometric co-occurrence matrix entropy value; areas with surface loose soil are accurately excluded because their curvature difference is close to zero, their directional entropy is extremely low, but their geometric co-occurrence matrix correlation parameters are distinguishable from those of real cracks; and real crack areas are accurately labeled and output by the model due to their high curvature difference, regular continuous concave depth distribution, and specific directional entropy range, ensuring that the object of quantitative analysis is structural cracks that pose a disaster threat, and significantly reducing the false alarm rate of the system.

[0109] In step four, 3D quantification calculation and dynamic risk assessment, the crack statistics module first uses the Delaunay triangulation algorithm to convert the identified crack point cloud into a continuous triangular mesh model, and then uses an iterative refinement algorithm based on geodesic distance to extract the 3D skeleton. Unlike traditional algorithms, the geodesic distance is calculated along the undulating path of the rock surface, which ensures that the extracted skeleton lines can accurately reproduce the winding path of the crack on the irregular rock wall. After the skeleton is generated, the system removes burrs on the skeleton using the Laplace smoothing algorithm and performs a fine analysis of the intersection nodes: clustering by angle similarity to determine the main extension direction of each node, and classifying the remaining converging branches as secondary branches, thus constructing a crack path network with a complete topological structure and clear hierarchy. Subsequently, the system performs path integration along the skeleton line to obtain the true three-dimensional length of the crack. By constructing a normal tangent plane at the skeleton node to measure the Euclidean distance between edge points, the opening value is obtained. The local slope tangent plane normal vector nlocal, obtained by least-squares fitting of background points at each skeleton node, is used as the projection reference direction. The vertical distance from the crack bottom point Pbottom to the local tangent plane within the normal section is taken as the depth value. This allows the measurement results to adapt to the slope undulations and accurately reflect the true indentation depth of the crack cavity relative to the surrounding intact slope, eliminating the interference of terrain undulations on depth quantization. After obtaining the aforementioned quantitative parameters, the risk assessment module multiplies the true three-dimensional length, aperture value, and depth of each crack to obtain the geometric volume damage of each crack. It then weights the normalized mean of each dimension with the crack distribution density using weights of wL=0.15, wW=0.25, wD=0.30, and wρ=0.30, and calculates Hidx=100×(1-S). Simultaneously, it performs inter-frame differential detection based on historical monitoring data to calculate the crack propagation rate and the frequency of new crack occurrence within a preset monitoring period, obtaining the crack activity level for each area of ​​the slope surface. When the crack aperture increase rate exceeds the warning threshold of 0.5 mm / day, it indicates that the rock mass at that location is in an accelerated deformation stage. The risk assessment module performs spatiotemporal joint analysis of the crack hazard distribution heatmap and the crack activity level, outputting a refined risk distribution map including safety, monitoring, early warning, and hazard zoning. In this way, managers can clearly see which cracks on the slope are expanding rapidly, thus achieving a technological leap from seeing cracks to predicting risks, and providing scientific decision support for precise mine support and personnel evacuation.

[0110] like Figure 2 As shown, a slope surface crack quantification analysis system based on three-dimensional point clouds includes:

[0111] Data acquisition and preprocessing module: used to acquire high-density three-dimensional point cloud video stream data of the slope surface and perform preprocessing to obtain a preprocessed point cloud image set;

[0112] Feature analysis module: used to analyze the local geometric features of each frame of point cloud in the point cloud image set to obtain the point cloud geometric attribute type; according to the point cloud geometric attribute type, set the corresponding attribute level to obtain the morphological features of the crack candidate region.

[0113] Crack identification module: used to identify real crack entities on the slope surface in the point cloud image set based on the morphological features;

[0114] Crack statistics module: used to extract the three-dimensional skeleton of the real crack entity and calculate the three-dimensional geometric dimensions and crack distribution density of the real crack entity;

[0115] Risk assessment module: used to conduct risk assessment on the slope surface based on the three-dimensional geometric dimensions and the crack distribution density, so as to obtain the slope crack risk assessment result.

[0116] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for quantitative analysis of surface cracks on slopes based on three-dimensional point clouds, characterized in that, include: High-density 3D point cloud video stream data of the slope surface is acquired and preprocessed to obtain a preprocessed point cloud image set. Analyze the local geometric features of each frame of the point cloud in the point cloud image set to obtain the point cloud geometric attribute type; Based on the point cloud geometric attribute type, the corresponding attribute level is set to obtain the morphological features of the crack candidate region; Based on the morphological features, identify real crack entities on the slope surface in the point cloud image set; Specifically, by analyzing the local fluctuation standard deviation of the point cloud depth value within the crack candidate region and comparing it with the filler particle size threshold, pseudo-crack regions with discontinuous depth abrupt changes caused by particle accumulation are eliminated. Combining point cloud normal vector consistency analysis and multi-scale curvature difference characteristics, texture structures with periodic repetition patterns are distinguished from real cracks with random extension directions and local depressions. Double filtering is performed by setting normal vector deflection angle threshold and curvature abrupt change gradient threshold to accurately obtain real crack entities. Extract the three-dimensional skeleton of the real crack entity, and calculate the three-dimensional geometric dimensions and crack distribution density of the real crack entity; Based on the three-dimensional geometric dimensions and the crack distribution density, a risk assessment is conducted on the slope surface to obtain the slope crack risk assessment results.

2. The method for quantitative analysis of slope surface cracks based on three-dimensional point clouds according to claim 1, characterized in that, Acquire high-density 3D point cloud video stream data of the slope surface, and preprocess the high-density 3D point cloud video stream data to obtain a standardized point cloud image set, including: A full-coverage scan of the slope surface was performed to acquire high-density three-dimensional point cloud video stream data; Based on the point cloud video stream data, obtain raw point cloud data containing spatial three-dimensional coordinates and reflection intensity; Spatial pass-through filtering is performed on the raw point cloud data, and a preset range of coordinate axes is set to remove background point clouds located outside the monitoring area; The remaining point cloud after the removal process is denoised. The average distance between each point cloud and a preset number of neighboring point clouds is calculated. Point clouds whose average distance exceeds a preset deviation threshold are identified as outlier noise point clouds and removed. The denoised point cloud is resampled, the point cloud space is divided into tiny voxels of a preset size, and the original points in each voxel are replaced by the centroid of all points in that voxel. The filtered and resampled point cloud data from each site are mapped to a pre-defined unified global coordinate system to align the spatial reference and generate a standardized point cloud image set.

3. The method for quantitative analysis of slope surface cracks based on three-dimensional point clouds according to claim 1, characterized in that, Analyze the local geometric features of each frame of the point cloud in the point cloud image set to obtain the point cloud geometric attribute types, including: Obtain each point to be processed in the point cloud image set, and search the neighborhood point set of each point to be processed within a preset search radius; Construct the covariance matrix of the neighborhood point set, and calculate the eigenvalues ​​of the covariance matrix, wherein the eigenvalues ​​include a first eigenvalue, a second eigenvalue, and a third eigenvalue; The local surface change rate is calculated by dividing the third characteristic value by the sum of the first, second, and third characteristic values. By fitting the local tangent plane of the neighborhood point set, the angular deviation between the normal vector of the point to be processed and the normal vector of the local tangent plane is obtained; Based on the local surface change rate and the included angle deviation, the geometric properties of the point cloud are divided into multiple geometric property types.

4. The method for quantitative analysis of slope surface cracks based on three-dimensional point clouds according to claim 3, characterized in that, Based on the point cloud geometric attribute type, the corresponding attribute level is set and the morphological features of the crack candidate region are obtained, including: Based on the various geometric attribute types, the attribute levels corresponding to the various geometric attribute types are set according to the edge abruptness and the indentation depth; The attribute levels include flat region level with low edge abruptness and shallow depression depth, transition region level with medium edge abruptness and moderate depression depth, suspected crack region level with high edge abruptness and deep depression, and high confidence region level with extremely high edge abruptness and continuous depression extension. Based on the attribute levels, points belonging to the suspected crack region level and the high confidence crack region level are filtered to obtain a candidate point set that conforms to the crack geometric characteristics. Using candidates with high curvature as seed points, the candidate point set is extended outwards according to preset normal vector angle thresholds and curvature deviation thresholds, and points that satisfy geometric continuity are clustered into crack clusters. The morphological parameters of length, width, orientation angle, and curvature continuity of each crack cluster are extracted to obtain the morphological features.

5. The method for quantitative analysis of slope surface cracks based on three-dimensional point clouds according to claim 1, characterized in that, Based on the morphological features, the identification of real crack entities on the slope surface in the point cloud image set includes: Calculate the local fluctuation standard deviation of the point cloud depth value within the crack candidate region, compare the local fluctuation standard deviation with the preset filler particle size threshold, and eliminate pseudo crack regions with discontinuous depth caused by the accumulation of small gravel particles to obtain the first crack candidate region verified by depth continuity. Based on the first crack candidate region and the normal vector in the morphological features, a geometric co-occurrence matrix reflecting the local spatial correlation of the point cloud is constructed. By calculating the energy, entropy, contrast and correlation threshold of the geometric co-occurrence matrix, feature vectors reflecting the difference between crack texture and background surface are extracted. Based on the aforementioned feature vectors, combined with point cloud normal vector consistency analysis and multi-scale curvature difference characteristics, a normal vector deflection angle threshold and a curvature abrupt change gradient threshold are set to perform dual filtering on slope stone texture structures with periodic repetition patterns and real cracks with random extension directions and local depression characteristics. The feature vector is input into a preset classification model, and the discrimination threshold determined by the dual filtering is used to distinguish between plant shading, surface soil and real cracks, eliminate false crack point clouds, and identify the real crack entities.

6. The method for quantitative analysis of slope surface cracks based on three-dimensional point clouds according to claim 1, characterized in that, Extract the three-dimensional skeleton of the real crack entity, and calculate the three-dimensional geometric dimensions and crack distribution density of the real crack entity, including: A three-dimensional solid model is created for the identified real crack entities to generate a mesh model of the crack surface; The crack voxels within the mesh model are peeled off layer by layer, and redundant branches are suppressed while preserving topological connectivity, generating a three-dimensional skeleton line composed of single-point chains. The three-dimensional skeleton lines are subjected to intersection node detection. The branch paths are assigned based on the continuity of local tangent direction, and the skeleton is decomposed into several independent crack path units with clear directions. The crack network topology is reconstructed based on the endpoint connection relationship of the independent crack path units, the hierarchical relationship between the main crack and the secondary branches is statistically analyzed, and the three-dimensional skeleton lines are smoothed to remove burrs on the skeleton. The true three-dimensional length of the crack is obtained based on the smoothed three-dimensional skeleton lines. Search for the edge points of the real crack entity in the normal plane of the three-dimensional skeleton line, and calculate the average Euclidean distance between the edge point pairs to obtain the crack opening value. Along the crack extension direction indicated by the three-dimensional skeleton line, calculate the projected displacement of the actual crack entity in the direction perpendicular to the slope surface to obtain the crack depth value. The total number of actual crack entities within the preset slope monitoring area is counted, and the crack coverage ratio per unit area is calculated to obtain the crack distribution density.

7. The method for quantitative analysis of slope surface cracks based on three-dimensional point clouds according to claim 6, characterized in that, A detailed analysis of the nodal regions with interwoven cracks and complex branches in the three-dimensional skeleton line is performed, including: At the intersection node, extract the local tangent vectors of all converging branches, cluster the local tangent vectors according to angular similarity, and determine the candidate potential continuous crack paths passing through the intersection node. For each candidate path, calculate the curvature change and tangent direction deflection angle of the skeleton line before and after crossing the intersection node. Preset the curvature change threshold and tangent direction deflection angle threshold. When the node meets the curvature change threshold and tangent direction deflection angle threshold, it is determined as the main extension object. The nodes that do not meet the threshold are aggregated and determined as secondary branches. Record the intersection angles of the secondary branches and the main extended crack paths, as well as the spatial coordinates of the connecting nodes, to construct a crack intersection network that includes primary and secondary hierarchical relationships and intersection position information; A global topology consistency check is performed on the fracture cross network to eliminate path loops and isolated short paths caused by misjudgment of local nodes, resulting in a fracture path network with a complete topology and clear hierarchy.

8. The method for quantitative analysis of slope surface cracks based on three-dimensional point clouds according to claim 6, characterized in that, Based on the stated three-dimensional geometric dimensions and the stated crack distribution density, a risk assessment is conducted on the slope surface to obtain the slope crack risk assessment results, including: The calculated true three-dimensional length, the opening value, and the depth value are multiplied to obtain the geometric volume damage of the crack. Based on the geometric volume damage and the crack distribution density, a slope health index model is constructed. The slope health index model is input based on the current slope crack parameters to output the corresponding slope health index value. The slope health index value is compared with multiple preset risk classification thresholds to determine whether the slope is in a safe, monitored, early warning, or dangerous state, thus obtaining the risk assessment result.

9. The method for quantitative analysis of slope surface cracks based on three-dimensional point clouds according to claim 8, characterized in that, Through inter-frame dynamic analysis and spatial hazard mapping, a spatiotemporally combined refined risk distribution map of the slope surface is generated, including: Inter-frame difference detection is performed on the continuously acquired high-density three-dimensional point cloud video stream data, the spatial displacement between adjacent frame point clouds is calculated, and key frames whose surface deformation exceeds a preset dynamic threshold are selected. The point cloud data corresponding to the key frame is differentially processed with the initial reference frame to extract the expansion increment of the crack region in the time dimension and construct a crack dynamic evolution sequence. Based on the dynamic evolution sequence of the cracks, the three-dimensional size expansion rate of the cracks and the frequency of new cracks appearing within a preset monitoring period are calculated to obtain the crack activity level of each area on the slope surface. Based on the crack network topology, the intersection nodes of multiple crack paths and the high-density clustering areas of cracks are identified, and the intersection nodes and high-density clustering areas are marked as potential priority damage points. The spatial coordinates of the potential priority failure points are superimposed and mapped with the corresponding real three-dimensional length, opening value and depth value to generate a crack hazard distribution heat map indexed by the spatial location of the slope surface. Based on the aforementioned hazard distribution heatmap and the aforementioned crack activity level, a refined risk distribution map of the slope surface, including safety, monitoring, early warning, and hazard zones, is output to assist in the formulation of differentiated slope reinforcement schemes and key monitoring area deployment strategies.

10. A slope surface crack quantification analysis system based on three-dimensional point clouds, used to execute the slope surface crack quantification analysis method based on three-dimensional point clouds as described in any one of claims 1-9, characterized in that, include: Data acquisition and preprocessing module: used to acquire high-density three-dimensional point cloud video stream data of the slope surface and perform preprocessing to obtain a preprocessed point cloud image set; Feature analysis module: used to analyze the local geometric features of each frame of point cloud in the point cloud image set to obtain the point cloud geometric attribute type; according to the point cloud geometric attribute type, set the corresponding attribute level to obtain the morphological features of the crack candidate region. Crack identification module: used to identify real crack entities on the slope surface in the point cloud image set based on the morphological features; Crack statistics module: used to extract the three-dimensional skeleton of the real crack entity and calculate the three-dimensional geometric dimensions and crack distribution density of the real crack entity; Risk assessment module: used to conduct risk assessment on the slope surface based on the three-dimensional geometric dimensions and the crack distribution density, so as to obtain the slope crack risk assessment result.

Images