A method and system for real-time identification of particle size distribution and gap distribution in rockfill based on three-dimensional laser scanning

By combining three-dimensional laser scanning with multi-source feature fusion and manifold constraint segmentation technology, the problems of insufficient segmentation and gap identification in rockfill quality inspection have been solved. Real-time and quantitative detection of rockfill particle size distribution and gap distribution has been achieved, supporting real-time control of construction quality in large-scale rockfill areas.

CN123119087APending Publication Date: 2026-11-17CHINA CONSTR EIGHT ENG DIV CORP LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610901793.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-22
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the construction of existing rockfill concrete dams, the rockfill quality inspection methods are highly subjective and inefficient, and cannot achieve quantitative online inspection of the entire surface. In particular, the rock contact boundary is not sufficiently segmented and the internal gap distribution cannot be identified.

Method used

By employing a three-dimensional laser scanning method, and combining multi-source heterogeneous feature fusion, manifold geometric constraint supervoxel segmentation, and three-dimensional morphological gap quantitative identification technology with two-dimensional image data for spatial registration, a comprehensive, quantitative, and online detection of particle size distribution and gap distribution in rockfill can be achieved.

Benefits of technology

It enables comprehensive, quantitative, and online detection of particle size distribution and gap distribution in rockfill, providing real-time data support and reliable data support for construction quality control. It can generate test reports within minutes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123119087A_ABST
    Figure CN123119087A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for real-time identification of particle size distribution and gap distribution in rockfill based on three-dimensional laser scanning. The method includes: acquiring synchronously collected three-dimensional point cloud data and two-dimensional image data; preprocessing the point cloud data; spatially registering the image data and point cloud data to generate a fused feature vector containing three-dimensional geometric features and two-dimensional texture features; performing supervoxel clustering based on the fused feature vector, using geodesic distance as a manifold constraint, to segment the point cloud into supervoxel regions corresponding to each stone material; extracting representative particle size values ​​for each supervoxel region and statistically generating gradation information; marking the supervoxel regions as occupied and performing three-dimensional morphological analysis on the remaining space to identify the equivalent pore size and volume distribution of particle gaps; generating and outputting a detection report. This invention achieves comprehensive, quantitative, and online detection of particle size distribution and gap distribution in rockfill.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering construction quality inspection technology, specifically to a method and system for real-time identification of particle size distribution and gap distribution in rockfill based on three-dimensional laser scanning. This invention is particularly suitable for online, non-contact, and quantitative detection of particle size composition, gradation curves, and particle gap distribution of rockfill in the dam surface during construction. Background Technology

[0002] The construction quality of rockfill concrete dams largely depends on the particle size distribution and packing density of the rockfill. Engineering design documents typically specify clear technical requirements for the particle size range, gradation curve shape, and rockfill ratio of the rockfill. When the proportion of oversized rocks in the rockfill is too high or the content of fine particles is insufficient, it will lead to the formation of excessively large voids within the rockfill, affecting the filling density of the concrete and the overall impermeability of the dam. When the gradation of the rockfill deviates significantly from the design curve, it will directly reduce the mechanical properties and long-term durability of the dam.

[0003] Currently, the quality inspection of rockfill in the construction of rockfill concrete dams mainly relies on manual visual inspection and random sampling and weighing, which has drawbacks such as strong subjectivity, inability to quantify, difficulty in covering the entire area, and long time consumption.

[0004] In recent years, some scholars have attempted to introduce machine vision and 3D laser scanning technology into the field of rockfill quality inspection. Patent CN110363855B discloses a method for transparent modeling of rockfill dams, using 3D laser scanning technology to acquire 3D point clouds of the rockfill surface, and then using point cloud processing technology to identify and spatially locate large stones. However, this method is limited to point cloud acquisition and basic processing, and does not involve quantitative statistical methods for particle size distribution or the identification of gap distribution.

[0005] Patent application CN117237395A discloses a method for analyzing the bulk density of a rock pile. This method uses a supervoxel segmentation algorithm to segment the point cloud data of the rock pile under test, optimizing the segmentation effect by setting noise removal and discrete thresholds. However, this method still only performs supervoxel segmentation based on the spatial information of the point cloud, and undersegmentation problems still occur in the boundary areas where rocks contact each other due to insufficient point cloud density. Furthermore, this method does not address the identification of gaps within the rock pile.

[0006] Existing particle size identification methods based on two-dimensional images are severely affected by lighting conditions and mutual occlusion of stones; existing local three-dimensional reconstruction schemes based on structured light or lidar have limited coverage and cannot meet the detection efficiency requirements of continuous construction scenarios in large warehouses.

[0007] In summary, existing methods for testing the quality of rockfill are highly subjective, inefficient, and unable to achieve quantitative online testing across the entire rockfill surface. Summary of the Invention

[0008] To address the problems of existing rockfill quality testing technologies, such as high subjectivity, low efficiency, and inability to achieve quantitative online detection of the entire rockfill surface, the present invention aims to provide a method and system for real-time identification of rockfill particle size distribution and gap distribution based on three-dimensional laser scanning. This method utilizes multi-source heterogeneous feature fusion, manifold geometrically constrained supervoxel segmentation, and three-dimensional morphological gap quantitative identification techniques to achieve comprehensive, quantitative, and online detection of rockfill particle size distribution and gap distribution, providing real-time data support for the construction quality of rockfill concrete dams.

[0009] To address the aforementioned technical problems, this invention provides a real-time identification method for the particle size distribution and gap distribution of rockfill based on three-dimensional laser scanning. This method is executed by an edge computing unit, which is communicatively connected to a data acquisition terminal. The method includes: (1) Acquire the three-dimensional point cloud data and two-dimensional image data of the rockfill surface synchronously collected by the data acquisition terminal; (2) Preprocess the three-dimensional point cloud data to obtain standardized point cloud data; (3) Spatial registration is performed between the two-dimensional image data and the normalized point cloud data to generate a fusion feature vector for each spatial point. The fusion feature vector includes the three-dimensional geometric features and two-dimensional texture features of the spatial point. (4) Based on the fusion feature vector of each spatial point, supervoxel clustering is performed with geodesic distance as manifold constraint condition to divide the point cloud on the surface of the rockfill into multiple supervoxel regions corresponding to each stone. (5) Extract three-dimensional size parameters for each of the supervoxel regions and determine the representative particle size value of the stone according to the preset particle size definition rules; (6) Statistically analyze the distribution of representative particle size values ​​in all supervoxel regions to generate particle size distribution information; (7) After marking the space occupied by each supervoxel region as occupied, perform three-dimensional morphological analysis on the remaining space to identify the distribution of interparticle gaps inside the rockfill and calculate the equivalent pore size and volume parameters of each gap region. (8) Generate and output a test report based on the particle size distribution information and the gap distribution information.

[0010] Furthermore, the three-dimensional geometric features include the local normal vector direction and the local surface variation of the spatial point; the two-dimensional texture features are local texture descriptors extracted based on the corresponding pixel position of the spatial point in the two-dimensional image data.

[0011] Further, the geodesic distance in step (4) is obtained by calculating the shortest path on the graph structure constructed from the point cloud data. The weight of the edge used for geodesic distance calculation in the graph structure is the Euclidean distance between the two endpoints in three-dimensional space. The manifold constraint condition is: the geodesic distance between the candidate point and the seed point does not exceed a preset threshold. The physical meaning of the preset threshold is the expected minimum boundary interval between adjacent stones.

[0012] Furthermore, the three-dimensional morphological analysis in step (7) includes: The space occupied by the riprap is voxelized into a binary 3D mesh; Perform connected component analysis on the binary 3D mesh to identify regions that are empty and spatially connected as gap regions; Perform a three-dimensional distance transformation on each gap region; The local maxima in the three-dimensional distance transformation result are used as the equivalent aperture of the gap region.

[0013] Furthermore, the representative particle size value in step (5) is the equivalent diameter of the volumetric sphere of the stone corresponding to the supervoxel region, which is calculated by the geometric mean of the major axis, middle axis and minor axis dimensions of the oriented bounding box of the supervoxel region.

[0014] To address the aforementioned technical problems, this invention also provides a real-time identification system for the particle size distribution and gap distribution of rockfill based on three-dimensional laser scanning, comprising: A data acquisition terminal is equipped with a 3D laser scanner and at least one industrial camera; the data acquisition terminal is configured to move along a preset path above the surface of a rockfill concrete dam, the 3D laser scanner is configured to acquire 3D point cloud data of the rockfill body on the dam surface, and the industrial camera is configured to acquire 2D image data of the surface of the rockfill body synchronously with the 3D laser scanner. An edge computing unit, which is communicatively connected to the data acquisition terminal; the edge computing unit is equipped with: A data preprocessing module is configured to perform denoising and spatial normalization processing on the three-dimensional point cloud data to obtain normalized point cloud data. A multi-source feature fusion module is configured to spatially register the two-dimensional image data with the normalized point cloud data, and generate a fused feature vector for each spatial point, wherein the fused feature vector includes the three-dimensional geometric features and two-dimensional texture features of the spatial point. The manifold constraint segmentation module is configured to perform supervoxel clustering based on the fused feature vectors of each spatial point, using geodesic distance as the manifold constraint condition, to segment the point cloud on the surface of the rockfill into multiple supervoxel regions corresponding to each stone. The particle size extraction module is configured to extract three-dimensional size parameters for each of the supervoxel regions and determine the representative particle size value of the stone according to a preset particle size definition rule. A gradation statistics module is configured to statistically analyze the distribution of representative particle size values ​​in all hypervoxel regions and generate particle size gradation information. The gap identification module is configured to mark the space occupied by each supervoxel region as occupied, perform three-dimensional morphological analysis on the remaining space, identify the distribution of particle gaps inside the rockfill, and calculate the equivalent pore size and volume parameters of each gap region. A report generation module is configured to generate and output a test report based on the particle size distribution information and the gap distribution information.

[0015] Furthermore, the three-dimensional geometric features include the local normal vector direction and the local surface variation of the spatial point; the two-dimensional texture features are local texture descriptors extracted based on the corresponding pixel position of the spatial point in the two-dimensional image data.

[0016] Furthermore, in the manifold constraint segmentation module, the geodesic distance is obtained by calculating the shortest path on the graph structure constructed from the point cloud data. The weight of the edge used for geodesic distance calculation in the graph structure is the Euclidean distance between the two endpoints in three-dimensional space. The manifold constraint condition is: the geodesic distance between the candidate point and the seed point does not exceed a preset threshold. The physical meaning of the preset threshold is the expected minimum boundary interval between adjacent stones.

[0017] Furthermore, in the gap recognition module, the three-dimensional morphological analysis includes: The space occupied by the riprap is voxelized into a binary 3D mesh; Perform connected component analysis on the binary 3D mesh to identify regions that are empty and spatially connected as gap regions; Perform a three-dimensional distance transformation on each gap region; The local maxima in the three-dimensional distance transformation result are used as the equivalent aperture of the gap region.

[0018] Furthermore, the representative particle size value is the equivalent diameter of the volumetric sphere of the stone corresponding to the supervoxel region, which is calculated by the geometric mean of the major axis, middle axis and minor axis dimensions of the oriented bounding box of the supervoxel region.

[0019] The present invention enables comprehensive, quantitative, and online detection of particle size distribution and interstitial distribution in rockfill. Compared with existing technologies, the present invention further offers the following advantages.

[0020] This invention simultaneously outputs particle size distribution information and gap distribution information. The particle size distribution information is used to evaluate whether the particle composition of the rockfill meets the design gradation requirements, while the gap distribution information (including the effective equivalent pore size and volume parameters of each gap) is used to evaluate whether the internal void state of the rockfill is conducive to the subsequent filling and flow of self-compacting concrete. These two types of information together constitute a complete digital profile of the rockfill quality. Existing technologies only focus on a single indicator such as particle size or rockfill ratio. This invention fills the gap in the detection of the crucial quality dimension of gap distribution for the first time.

[0021] This invention introduces a fused feature vector in the supervoxel segmentation step. This fused feature vector simultaneously contains three-dimensional geometric features (normal vectors, surface variations) and two-dimensional texture features (dense SIFT descriptors), enabling the segmentation algorithm to make distinctions based on both geometric and texture boundaries. Even if the point cloud density at the stone contact boundary is insufficient, texture differences can still provide an effective basis for segmentation, thus effectively overcoming the undersegmentation problem caused by insufficient point cloud density at the stone boundary in single three-dimensional point cloud data.

[0022] This invention, through a four-step process of binary voxel mesh construction, 3D connected domain analysis, 3D distance transformation, and local maximum extraction, achieves for the first time a quantitative description of the interparticle gaps within a rockfill mass. In particular, it defines an "effective equivalent pore size" (the minimum of all local maxima), which directly reflects the size of the narrowest gap and is a key parameter for determining whether self-compacting concrete can pass through smoothly. This quantitative data can be directly used to assess the infillability of the rockfill mass, providing precise input conditions for subsequent mix design and optimization of pouring process parameters for self-compacting concrete.

[0023] Each processing step of this invention is executed in a pipeline manner on the edge computing unit. After the data acquisition terminal completes the scan, the data is processed in real time at the edge, without needing to be uploaded to the cloud. The end-to-end latency of the entire process, from point cloud data input to detection report output, is controlled within minutes. This real-time online processing capability enables this invention to perform synchronous detection along with the construction progress of the concrete slab, issuing timely quality warnings before concrete pouring, and providing reliable data support for real-time control of construction quality. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0025] Figure 1 This is a schematic diagram illustrating the deployment principle of the real-time identification system for rockfill particle size distribution and gap distribution based on three-dimensional laser scanning in this invention.

[0026] Figure 2 This is an example diagram illustrating the structure of the multi-source feature fusion module in this invention.

[0027] Figure 3 This is an example diagram illustrating the structure of the manifold constraint segmentation module in this invention.

[0028] Figure 4 This is an example diagram illustrating the structure of the gap recognition module in this invention.

[0029] Figure 5 This is a schematic diagram of the overall process of the real-time identification method for rockfill particle size distribution and gap distribution based on three-dimensional laser scanning in this invention.

[0030] Figure 6 This is a schematic diagram of the multi-source feature fusion process in this invention.

[0031] Figure 7 This is a schematic diagram illustrating the principle of manifold-constrained supervoxel segmentation and particle size extraction in this invention.

[0032] Figure 8 This is a schematic diagram illustrating the principle of three-dimensional morphological gap recognition in this invention. Detailed Implementation

[0033] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0034] Analysis of existing technologies reveals three main technical deficiencies: First, existing point cloud-based methods for identifying the particle size of piled stones only utilize spatial geometric information, resulting in severe undersegmentation at the contact boundaries of the stones due to insufficient point cloud density. Secondly, existing methods cannot quantitatively identify the interparticle gaps inside the rockfill, while the gap distribution is a key indicator for evaluating the fillability of the rockfill. Third, existing detection methods are either static local scanning with limited coverage or post-construction sampling, which cannot meet the needs of real-time online detection for continuous construction of large warehouses.

[0035] Accordingly, this invention provides a real-time identification scheme for the particle size distribution and gap distribution of rockfill based on three-dimensional laser scanning. This scheme effectively solves the problems existing in the prior art by adopting the following innovative technical means.

[0036] This invention constructs a deep fusion mechanism for multi-source heterogeneous features. Addressing the problem of insufficient point cloud density at stone boundaries in pure point cloud methods, this invention precisely maps texture information from a 2D image to every spatial point in a 3D point cloud through spatial registration, constructing a fused feature vector containing both geometric features (normal vectors, surface variations) and texture features. This allows the segmentation algorithm to simultaneously judge based on both geometric and texture boundaries; even with insufficient point cloud density, texture differences can still provide effective criteria for segmentation.

[0037] This invention constructs a manifold geometric constraint clustering mechanism based on geodesic distance. Addressing the problem of erroneously merging adjacent stones by crossing gaps in supervoxel segmentation, this invention introduces geodesic distance as a manifold constraint during clustering. Clustering is prohibited if the geodesic distance between a candidate point and a seed point exceeds a preset threshold. This threshold is set as the expected minimum boundary interval between adjacent stones, ensuring that the segmentation boundary is strictly confined within the physical boundaries of the stones. In the calculation of the geodesic distance, the edge weights of the graph structure use spatial Euclidean distance to accurately reflect the geometric path length on the point cloud surface.

[0038] The present invention provides a three-dimensional morphological quantitative identification method for gap distribution. Addressing the technological gap that existing technologies cannot detect gaps inside rockfill bodies, the present invention achieves, for the first time, quantitative calculation of the equivalent pore size and volume of particle gaps inside rockfill bodies through a four-step process of binary voxel mesh construction, three-dimensional connected domain analysis, three-dimensional distance transformation, and local maximum extraction.

[0039] The present invention adopts a real-time online computing architecture with "end-edge" collaboration. In view of the requirements for detection efficiency in continuous construction of large warehouses, the present invention adopts an architecture of collaboration between data acquisition terminal and edge computing unit. The acquisition terminal is responsible for full coverage scanning, and the edge computing unit executes all calculation steps online in a pipeline manner. The end-to-end latency is controlled at the minute level, so that the detection can be carried out synchronously with the construction progress.

[0040] Based on the organic combination of the above-mentioned technical means, the present invention can achieve comprehensive, quantitative, and online detection of the particle size distribution and gap distribution of rockfill.

[0041] See Figure 1 The example shown is an example of a real-time identification system for rockfill particle size distribution and gap distribution based on three-dimensional laser scanning, which is formed by the above-mentioned technical solution mechanism of the present invention.

[0042] Based on the prompt, this real-time identification system for rockfill particle size distribution and gap distribution based on three-dimensional laser scanning is mainly composed of two parts: a data acquisition terminal 100 and an edge computing unit 200.

[0043] The data acquisition terminal 100 is configured to perform full-coverage three-dimensional laser scanning and optical image acquisition of the rockfill concrete dam surface under a unified spatiotemporal reference, providing raw multi-source data for subsequent analysis.

[0044] Furthermore, the data acquisition terminal 100 adopts a gantry-type mobile platform to achieve full coverage scanning of the warehouse surface. On this basis, a 3D laser scanner 110 and multiple industrial cameras 120 are further configured, thereby overcoming the limitation of the limited coverage of traditional static scanning.

[0045] Furthermore, the hardware-level synchronous trigger circuit 130 enables precise synchronous acquisition between the 3D laser scanner 110 and multiple industrial cameras 120, ensuring that the point cloud data and image data are strictly aligned in time.

[0046] Furthermore, the data acquisition terminal 100 integrates differential GPS and an inertial measurement unit to provide a pose reference for the spatial coordinate standardization of point cloud data.

[0047] The data acquisition terminal 100 and the edge computing unit 200 form a data flow transmission relationship through a high-speed data interface. The acquisition terminal is responsible for acquiring and initially organizing the raw data, while the edge computing unit is responsible for in-depth data processing and analysis.

[0048] The edge computing unit 200 in this system is configured to receive raw data streams from the data acquisition terminal 100, perform all calculation steps online from data preprocessing to test report generation, and transform the raw data into quality inspection information usable in the project.

[0049] Furthermore, the edge computing unit 200 adopts an "end-edge" computing architecture, which pushes computing tasks down to the construction site, avoiding bandwidth bottlenecks and latency issues when uploading massive amounts of raw data to the cloud.

[0050] The edge computing unit 200 and the data acquisition terminal 100 form a closed loop, with the former receiving the latter's raw data and the latter receiving the former's control commands and feedback information.

[0051] In this system, the data acquisition terminal 100 and the edge computing unit 200 are connected via a high-speed data interface to form a complete closed loop from data acquisition to result output.

[0052] The following section provides a further explanation of the specific configuration of each component in this system solution, as well as the equipment or parts that may be involved.

[0053] like Figure 1As shown, the main body of the data acquisition terminal 100 is a gantry-type frame structure 101 that can automatically move on tracks laid on both sides of the warehouse surface. The span of the gantry beam is designed according to the width of the warehouse surface to ensure that the scanning range can completely cover the entire warehouse surface. A cluster of core sensing devices is installed below the beam, specifically including: The 3D laser scanner 110 in this data acquisition terminal 100 preferably employs a line-scanning LiDAR with a laser emission wavelength of 905 nanometers (within the human eye-safe band). The scanning line frequency is 200 contour lines per second, with each contour line containing 2048 measurement points. The point measurement accuracy is ±3 millimeters at a measurement distance of 5 meters. The laser scanner continuously scans the surface of the storage area directly below from a top-down angle, generating dense 3D point cloud data covering the entire surface. The combination of high scanning frequency and high point accuracy in this scanner allows for the acquisition of sufficiently dense point cloud data even under moving scanning conditions, providing a data foundation for subsequent stone boundary identification.

[0054] This data acquisition terminal 100 is equipped with multiple high-resolution industrial cameras 120, preferably three, which are mounted around the 3D laser scanner 110 at 120-degree intervals. Each camera has a resolution of 2048×2048 pixels and a frame rate of 30 frames per second. This arrangement of multiple cameras effectively overcomes the occlusion problem caused by the 3D undulations of the rockfill surface caused by a single camera. For example, when a rock is obscured by another rock, cameras at other angles can still capture its surface texture information, ensuring that at least one camera at each spatial point can provide effective image data.

[0055] Based on this, the data acquisition terminal 100 is also configured with a high-precision synchronous trigger circuit 130, a positioning and attitude sensing unit 140, a walking drive system 150, and a data preprocessing and transmission unit 160.

[0056] The high-precision synchronization trigger circuit 130 is configured to connect and control three industrial cameras 120 and a 3D laser scanner 110. This circuit ensures that the three industrial cameras 120 maintain hardware-level synchronization with the scanning pulses of the 3D laser scanner 110, guaranteeing that each frame of image data is precisely aligned with the corresponding laser point cloud data in terms of timestamps. This synchronization mechanism differs from software-level timestamp alignment; the hardware-level synchronization accuracy can reach the microsecond level, eliminating time deviations caused by data transmission and processing delays in software synchronization. This provides a precise time reference for subsequent spatial registration of point clouds and images.

[0057] The positioning and attitude sensing unit 140 includes a Differential Global Positioning System (DGPS) receiver and an Inertial Measurement Unit (IMU) mounted at both ends of the gantry. The DGPS is used to record the terminal's geographic location coordinates in real time, while the IMU is used to record the terminal's attitude angles (roll, pitch, and yaw) in real time. The combined use of these two components enables precise determination of the spatial pose of each frame of data during the moving scan process, providing a spatial reference for subsequent point cloud data coordinate normalization and multi-source data registration.

[0058] The walking drive system 150 drives the gantry frame, which is connected to servo motor-driven walking wheels 102 at both ends, moving along tracks on both sides of the warehouse surface. The walking speed can be adjusted from 0.05 m / s to 0.15 m / s, and the scanning path is a preset serpentine path to ensure full coverage of the warehouse surface.

[0059] The data preprocessing and transmission unit 160 is configured to connect the industrial camera 120 and the 3D laser scanner 110 to perform preliminary processing and compression of the raw point cloud data and image data, and then transmit the data in real time to the edge computing unit 200 via a fiber optic Ethernet interface (10Gbps bandwidth). For example, the data preprocessing and transmission unit 160 can be constructed from an embedded data preprocessing board integrated within the acquisition terminal.

[0060] In its specific implementation, the edge computing unit 200 in this system is mainly composed of an industrial-grade edge server equipped with a multi-core central processing unit (CPU), a large-capacity random access memory (RAM), and a dedicated graphics processing unit (GPU). It is connected to the data acquisition terminal 100 through a fiber optic Ethernet interface and has the ability to operate stably in industrial field environments.

[0061] In terms of software, this edge computing unit 200 has seven software functional modules deployed internally. Each module is executed in series in a pipeline manner, with the output of the previous module serving as the input of the next module, thereby completing the real-time identification of the particle size distribution and gap distribution of the rockfill.

[0062] Combination Figure 2 As shown, the edge computing unit 200 internally deploys a data preprocessing module 210, a multi-source feature fusion module 220, a manifold constraint segmentation module 230, a particle size extraction module 240, a gradation statistics module 250, a gap identification module 260, and a report generation module 270.

[0063] The data preprocessing module 210 in this unit is configured to perform quality purification and spatial normalization on the raw 3D point cloud data. The data input end of this module is communicatively connected to the data preprocessing and transmission unit 160 of the data acquisition terminal 100. The data output end of this module is connected to the data input end of the multi-source feature fusion module 220.

[0064] Furthermore, the data preprocessing module 210 specifically comprises a statistical outlier removal submodule and a coordinate normalization submodule that work together.

[0065] The outlier removal submodule receives the raw point cloud data and performs statistical analysis on the neighborhood of each spatial point. It calculates the mean and standard deviation of the distances between the point and its K nearest neighbors. If the average distance between a point and its neighbors exceeds a preset multiple of the global average distance standard deviation, it is marked as an outlier and filtered out. The denoised point cloud data then enters the coordinate normalization submodule.

[0066] The coordinate normalization submodule is configured to interact with the statistical outlier removal submodule, using DGPS and IMU data from the acquisition terminal to transform the denoised point cloud from the sensor local coordinate system to the unified construction coordinate system of the dam.

[0067] After denoising and coordinate normalization are completed, the data preprocessing module 210 outputs the normalized point cloud data to the multi-source feature fusion module 220. The statistical filtering method in the data preprocessing module 210 does not require a preset absolute threshold parameter and can adaptively identify and eliminate isolated noise points caused by laser multipath reflection or dust scattering.

[0068] The data input terminal of the multi-source feature fusion module 220 in this unit is communicatively connected to the data output terminal of the data preprocessing module 210 (receiving normalized point cloud data) and the data preprocessing and transmission unit 160 of the data acquisition terminal 100 (receiving synchronous two-dimensional image data); simultaneously, the data output terminal of this module is connected to the data input terminal of the manifold constraint segmentation module 230. This multi-source feature fusion module 220 is configured to spatially register two-dimensional image data with three-dimensional point cloud data, generating a fused feature vector for each spatial point.

[0069] like Figure 2 As shown, the multi-source feature fusion module 220 is specifically composed of four sub-modules: spatial registration sub-module 221, three-dimensional geometric feature extraction sub-module 222, two-dimensional texture feature extraction sub-module 223, and fusion feature vector assembly sub-module 224.

[0070] The spatial registration submodule 221 is configured to project each spatial point in the 3D point cloud onto the image pixel coordinate system of each industrial camera 120 using the camera-LiDAR extrinsic parameter matrix obtained in advance through Zhang Zhengyou calibration method and point cloud-image corresponding point matching algorithm, based on the input normalized point cloud data and synchronous 2D image data. For points that fall within the overlapping fields of view of multiple cameras, the camera closest to the point is selected as the main projection camera based on the distance from the point to the optical center of each camera to ensure optimal projection accuracy. This spatial registration submodule 221 generates and outputs the 3D coordinates of each spatial point and its corresponding image pixel coordinates, which are then transmitted to the 3D geometric feature extraction submodule 222 and the 2D texture feature extraction submodule 223, respectively.

[0071] The 3D geometric feature extraction submodule 222 is configured to determine the neighborhood range for each spatial point in the normalized point cloud dataset using a radius search method, calculate the covariance matrix of the neighborhood point set, and solve for the eigenvalues ​​and eigenvectors. The eigenvector corresponding to the smallest eigenvalue is used as the local normal vector of that point; the ratio of the smallest eigenvalue to the sum of the three eigenvalues ​​is used as the local surface variation of that point. The closer this ratio is to 0, the closer the region where the point is located is to a plane; the larger the ratio, the more drastic the surface variation. This eigenvalue is configured as an engineering approximation of curvature, used to distinguish between the stone surface (gradual variation) and the stone boundary or gap (drastic variation). Based on this, the 3D geometric feature extraction submodule 222 generates and outputs the geometric feature vector of each spatial point, and transmits it to the feature vector fusion assembly submodule 224.

[0072] The 2D texture feature extraction submodule 223 is configured to calculate the local texture descriptor around each pixel at the image pixel coordinates corresponding to each 3D spatial point. It should be noted that this scheme uses the Dense Scale Invariant Feature Transform (Dense SIFT) method, which calculates the SIFT descriptor at each pixel location, rather than only at the keypoints detected in scale space as in the standard SIFT algorithm. Dense SIFT can provide corresponding texture features for each spatial point in the point cloud, making it suitable for 3D-2D fusion scenarios requiring point-by-point feature alignment. Specifically, gradient direction histograms of 4×4 sub-regions in 8 directions are statistically analyzed within a 16×16 pixel neighborhood window, forming a 128-dimensional feature vector. Based on this, the 2D texture feature extraction submodule 223 generates and outputs the texture feature vector for each spatial point, and transmits it to the fusion feature vector assembly submodule 224.

[0073] The fusion feature vector assembly submodule 224 is configured to synchronously receive data input from the 3D geometric feature extraction submodule 222 and the 2D texture feature extraction submodule 223. For each spatial point, it concatenates and splices its 3D spatial coordinates (3D), local normal vector (3D), local surface variation (1D), and texture descriptor (128D) to form a 135-dimensional fusion feature vector. Before splicing, each component is normalized to eliminate dimensional differences. The normalization method can be Z-score standardization, which involves subtracting the mean from each feature component and then dividing by the standard deviation. Specifically, the fusion feature vector assembly submodule 224 generates and outputs a point-by-point corresponding fusion feature vector matrix, which is transmitted to the manifold constraint segmentation module 230 along with the normalized point cloud data.

[0074] The multi-source feature fusion module 220 formed based on the above scheme realizes point-by-point fusion of point cloud geometric features and image texture features at the spatial point level, rather than simple data-level stitching or decision-level fusion.

[0075] The data input of the manifold constraint segmentation module 230 in this unit is connected to the data output of the multi-source feature fusion module 220. The data output of this module is also connected to the data inputs of the particle size extraction module 240 and the gap recognition module 260. The manifold constraint segmentation module 230 is configured to use the fused feature vector as the segmentation basis, the normalized point cloud as the segmentation object, and perform supervoxel clustering based on the fused feature vector and geodesic distance as the manifold constraint to segment the point cloud into independent regions corresponding to individual stone blocks.

[0076] See Figure 3 The manifold constraint segmentation module 230 consists of four sub-modules: graph structure construction sub-module 231, seed point initialization sub-module 232, manifold constraint clustering assignment sub-module 233, and iterative optimization sub-module 234.

[0077] The graph structure construction submodule 231 is configured to, for the received fused feature vectors and normalized point cloud data, use each spatial point in the normalized point cloud dataset as a node, search for the K nearest neighbors of each node in the feature space, and establish edge connections between adjacent nodes. Each edge in the graph is assigned two sets of weights: one is the feature space weight, defined as the weighted Euclidean distance between the fused feature vectors of the two endpoints, used for similarity measurement during cluster assignment; the other is the spatial weight, defined as the Euclidean distance between the two endpoints in three-dimensional space, used for geodesic distance calculation. Based on this, the submodule generates an output K nearest neighbor graph structure with dual weights and transmits it to the seed point initialization submodule 232 and the manifold constraint cluster assignment submodule 233.

[0078] The seed point initialization submodule 232 is configured to uniformly sample initial seed points in the point cloud space for the received graph structure data. The sampling interval is set according to half of the minimum detection particle size, and each seed point is assigned a unique supervoxel label. Based on this, the submodule generates an output initial seed point set and its labels, and transmits it to the manifold-constrained clustering allocation submodule 233. The manifold-constrained clustering allocation submodule 233 is configured to, for the received graph structure data and initial seed point set, expand outwards layer by layer along the edges of the graph from each seed point, assigning candidate nodes to the seed point that minimizes the feature distance cost function. Specifically, the feature distance cost function is defined as the weighted Euclidean distance between the fused feature vector F(p) of candidate point p and the fused feature vector F(s) of seed point s: d_feature(p,s) = [Σ_{i=1}^{135} w_i · (F_i(p) - F_i(s)) 2 ]^(1 / 2); Where w_i is the weight of each feature component. The candidate point p is assigned to the seed point s that minimizes d_feature(p,s), but must also satisfy the geodesic distance constraint.

[0079] Unlike standard supervoxel segmentation methods, in this scheme, a candidate node is allowed to be assigned to a certain sub-point only if the shortest path length (i.e., geodesic distance) from the candidate node to the seed point along the graph edge does not exceed a preset threshold. The geodesic distance is calculated by running Dijkstra's shortest path algorithm on the graph, using the aforementioned spatial weights (i.e., Euclidean distance in three-dimensional space) as edge weights. The physical meaning of the preset threshold is the expected minimum boundary interval between adjacent stones, for example, it can be set to 30 mm. Generally, the geodesic distance threshold should be no less than half of the seed point interval and no greater than the seed point interval to ensure that clustering can cover the stone surface without crossing gaps. Based on this, the submodule generates the initial clustering assignment results and transmits them to the iterative optimization submodule 234.

[0080] The iterative optimization submodule 234 is configured to iteratively optimize the initial clustering assignment. In each iteration, the center position of each supervoxel is recalculated, and edge nodes are reassigned to the nearest supervoxel center. Specifically, for each supervoxel region, its center position is defined as the mean vector of the fused feature vectors of all spatial points within that region (i.e., the centroid in feature space), rather than the geometric center in 3D space. The iteration termination condition can be set to a percentage of nodes whose supervoxel boundaries have moved below 0.1% or the number of iterations reaches 20. The output of this submodule is the final supervoxel segmentation result—each spatial point is assigned a supervoxel label, and the set of point clouds with the same label corresponds to an independent stone.

[0081] Based on the above scheme, the geodesic distance constraint in the manifold constraint segmentation module 230 ensures that the clustering process follows the manifold geometry of the point cloud surface: on the stone surface, the geodesic distance is close to the Euclidean distance; however, at the gaps between stones, due to the lack of connection points, the geodesic distance increases sharply and exceeds the threshold, thus preventing clustering from crossing the gaps. This is fundamentally different from the standard method that only uses Euclidean distance or feature space distance: the standard method cannot perceive the topological structure of the point cloud surface and is prone to mistakenly merging stones on both sides of the gap; while the manifold constraint of this invention strictly restricts the segmentation boundary within the physical boundary of the stones, fundamentally preventing mistaken merging across gaps. Iterative optimization causes the segmentation boundary to converge to the geometric position of a local curvature maxima or abrupt change in normal vector, which precisely corresponds to the contact boundary between adjacent stones.

[0082] The particle size extraction module 240 in this unit has its data input terminal connected to the data output terminal of the manifold constraint segmentation module 230, receiving the supervoxel segmentation results and normalized point cloud data; its data output terminal is connected to the data input terminal of the gradation statistics module 250. The particle size extraction module 240 is configured to fit an oriented bounding box to each supervoxel region and extract representative particle size values.

[0083] Specifically, the particle size extraction module 240 is mainly composed of three sub-modules connected and coordinated in sequence: the main direction estimation sub-module, the size extraction sub-module, and the representative particle size calculation sub-module.

[0084] The supervoxel segmentation results and normalized point cloud data first enter the principal orientation estimation submodule. This submodule is configured to calculate the covariance matrix for the coordinates of all points within each supervoxel region, and solve for the eigenvalues ​​and eigenvectors. The three eigenvectors correspond to the major axis, median axis, and minor axis directions, respectively. This submodule generates and outputs the three principal orientation vectors for each supervoxel region.

[0085] The principal orientation estimation results are then fed into the size extraction submodule. This submodule is configured to rotate the point coordinates to a local coordinate system aligned with the principal orientation, calculate the coordinate spans along the three axes, and obtain the major axis dimension L, the median axis dimension M, and the minor axis dimension S. It should be noted that since the eigenvector sign (positive or negative) of the covariance matrix is ​​arbitrary, but the coordinate spans along the three axes (maximum minus minimum) are unaffected by the eigenvector sign, this rotation process has no impact on the subsequent particle size extraction results. This submodule generates the three-dimensional size parameters (L, M, S) for each supervoxel region.

[0086] The three-dimensional dimensional parameters are then fed into the representative particle size calculation submodule. This submodule is configured to calculate the representative particle size value based on the definition of the equivalent diameter of an equal-volume sphere: d_eq = (L × M × S)^(1 / 3). This submodule generates and outputs the representative particle size value for each stone and its corresponding spatial location information.

[0087] The data input terminal of the gradation statistics module 250 in this unit is connected to the data output terminal of the particle size extraction module 240, receiving representative particle size values ​​from all supervoxel regions; the data output terminal of this module is connected to the first data input terminal of the report generation module 270. The gradation statistics module 250 is configured to summarize the particle size information of individual stones into gradation statistics information for the entire bin surface.

[0088] Specifically, the gradation statistics module 250 is mainly composed of three sub-modules connected and coordinated in sequence: a particle size interval division sub-module, a volume weighted accumulation sub-module, and a gradation curve generation sub-module.

[0089] The representative particle size values ​​of all stone materials are first entered into the particle size range segmentation submodule. This submodule is configured to group the representative particle size values ​​according to preset particle size range boundaries. For example, the particle size range boundaries can be set to 100mm, 150mm, 200mm, 300mm, 400mm, 500mm, 600mm, 800mm, 1000mm, etc. Based on this, the submodule generates and outputs the grouping results for each particle size range.

[0090] The grouping results are then fed into a volume-weighted accumulation submodule. This submodule is configured to accumulate the volume of all supervoxel regions falling within each interval (approximately using the bounding box volume L×M×S). Based on this, the submodule generates the volume accumulation value for each particle size interval.

[0091] The volume accumulation results are then fed into the gradation curve generation submodule. This submodule is configured to plot a particle size distribution histogram and a gradation curve with the particle size range as the x-axis and the cumulative volume percentage as the y-axis. Based on this, the submodule generates output gradation statistics, such as particle size distribution histogram data + gradation curve data, which is then passed to the report generation module 270.

[0092] The data input terminal of the gap identification module 260 in this unit is connected to the data output terminal of the manifold constraint segmentation module 230, receiving the supervoxel segmentation results and the orientation bounding box parameters of each supervoxel region. The data output terminal of this module is connected to the second data input terminal of the report generation module 270. The gap identification module 260 is configured to perform three-dimensional morphological quantitative identification of the gaps between particles inside the rockfill.

[0093] See Figure 4 The gap identification module 260 specifically includes four sub-modules that are executed in sequence: binary voxel mesh construction sub-module 261, three-dimensional connected domain analysis sub-module 262, three-dimensional distance transformation sub-module 263, and equivalent aperture and volume extraction sub-module 264.

[0094] Specifically, the super-voxel segmentation results and the oriented bounding box parameters of each super-voxel region are first fed into the binary voxel mesh construction submodule 261. This submodule is configured to regularly divide the space occupied by the rockfill into a 3D mesh, and determine whether the center point of each voxel falls inside the oriented bounding box of any super-voxel region—a value of 1 is assigned if it falls in (rock occupancy), and a value of 0 is assigned if it does not fall in (candidate gap). The spatial range of voxelization is defined as follows: in the horizontal direction, a preset margin is extended outward from the silo surface design boundary (e.g., 100 mm outward from the silo surface boundary) to ensure coverage of the entire rockfill area; in the vertical direction, the lower limit is the silo surface reference plane, and the upper limit is the maximum rockfill height. This submodule excludes voxels below the silo surface reference plane and those exceeding the maximum rockfill height. The maximum rockfill height is obtained by statistically analyzing the difference between the maximum coordinate value of the point cloud on the rockfill surface in the vertical direction and the elevation of the silo surface reference plane. Based on this, the submodule generates the output binary 3D mesh.

[0095] The generated binary 3D mesh then enters the 3D connected component analysis submodule 262. This submodule is configured to perform a 26-neighborhood 3D connected component analysis on all voxels assigned a value of 0, clustering spatially adjacent 0-value voxels into independent connected components. This submodule filters out tiny connected components with volumes smaller than a preset threshold (considered noise artifacts). For example, when the voxel resolution is 10 mm, the volume of a single voxel is 1000 cubic millimeters, and a volume threshold of 100 voxels (i.e., 100,000 cubic millimeters) can be set to filter out connected components smaller than this threshold. Based on this, the submodule generates the output set of each gap connected component and its voxel coordinates.

[0096] The generated interstitial connected domains are then fed into a 3D distance transformation submodule 263. This submodule is configured to calculate the shortest Euclidean distance from each 0-value voxel within each retained interstitial connected domain to the nearest 1-value voxel (stone boundary). This submodule employs a fast distance transformation algorithm with two passes, the time complexity of which is linearly related to the total number of voxels. Based on this, the submodule generates the distance transformation result for each interstitial connected domain, with each voxel carrying the distance value to the boundary.

[0097] The generated distance transformation results are then fed into the equivalent aperture and volume extraction submodule 264. This submodule is configured to search for local maxima in the distance transformation results of each gap's connected domain, using them as the equivalent aperture at that location (physically meaning the largest sphere diameter that can freely pass through that point). The submodule takes the minimum of all local maxima distance values ​​as the effective equivalent aperture for that gap; this directly determines whether self-compacting concrete can smoothly pass through the gap—the narrowest point is the key to fluidity. The submodule obtains the gap volume by multiplying the total number of voxels within the gap's connected domain by the volume of a single voxel. Based on this, the submodule generates the effective equivalent aperture and volume parameters for each gap, which are then passed to the report generation module 270.

[0098] Based on the above scheme, this gap identification module 260 discretizes the continuous space through voxelization to make it computationally processable; it clusters scattered zero-value voxels into physically meaningful independent gaps through connected component analysis; and it assigns the geometric quantity of "distance to the boundary" to each voxel through distance transformation; the local maximum of the distance transformation corresponds precisely to the "widest point" of the gap—that is, the equivalent aperture. Through the organic cooperation between the four sub-modules, a quantitative description of the interparticle gaps inside the rockfill is achieved.

[0099] The first data input terminal of the report generation module 270 in this unit is connected to the data output terminal of the gradation statistics module 250, and the second data input terminal is connected to the data output terminal of the gap identification module 260; the data output terminal of this module is connected to the construction quality control system through a standardized data interface. The report generation module 270 is further configured to present the analysis results in an engineering-usable format.

[0100] The report generation module 270 mainly includes four sub-modules that are executed in sequence: rockfill ratio calculation sub-module, gradation deviation evaluation sub-module, report structure generation sub-module, and data output sub-module.

[0101] The aforementioned gradation statistics and gap identification results are first fed into the rockfill ratio calculation submodule. This submodule is configured to calculate the rockfill ratio of the entire storage area using the formula: Rockfill Ratio = (V_stone / V_total) × 100%, where V_stone is the cumulative volume of all supervoxel regions, and V_total is obtained by inputting the storage area design boundary and design rockfill height into the storage area reference plane obtained from the ground filtering step. If the actual rockfill has overfill or underfill conditions, V_total uses the design value instead of the actual value to ensure that the evaluation benchmark for the rockfill ratio is consistent with the design requirements. For example, if the design rockfill profile of a storage area is 15m × 8m × 3m, then V_total = 15 × 8 × 3 = 360 cubic meters. Regardless of whether the top of the actual rockfill is higher or lower than 3 meters, 360 cubic meters is used as the denominator to calculate the rockfill ratio. Based on this, the submodule generates the output rockfill ratio calculation value.

[0102] The generated calculated riprap ratio is then fed into the gradation deviation evaluation submodule along with the gradation statistics. This submodule is configured to compare the actual gradation curve with the upper and lower limit envelopes of the design gradation curve and quantify the deviation. This submodule generates and outputs the deviation values ​​and warning levels for each particle size range.

[0103] All the above results are then processed by the report structuring generation submodule. This submodule is configured to summarize the above information into a structured inspection report, which includes a particle size distribution histogram, a gradation curve overlay plot, a calculated rockfill ratio, a void volume distribution histogram, a void equivalent pore size distribution plot, and a gradation deviation warning list. The output of this submodule is a complete structured inspection report.

[0104] The generated structured inspection report data then enters the data output submodule. This submodule is configured to output the report to the construction quality control system via a standardized data interface (such as a RESTful API based on the HTTP protocol and JSON data format).

[0105] Based on the established real-time identification system scheme for rockfill particle size distribution and gap distribution, this invention further provides a real-time identification method for rockfill particle size distribution and gap distribution based on this system.

[0106] See Figure 5 The real-time identification method for particle size distribution and gap distribution of rockfill provided by this invention is mainly achieved through the following steps.

[0107] Step S1: Acquisition of multi-source data.

[0108] The data acquisition terminal 100 moves along a preset path above the silo surface, while the 3D laser scanner 110 continuously acquires 3D point cloud data of the rockfill surface. The industrial camera 120 synchronously acquires 2D image data of the same area with the laser scanner, and the data is transmitted in real-time to the edge computing unit 200 via the data preprocessing and transmission unit 160. The function of this step is to provide multi-source raw data under a unified spatiotemporal reference for subsequent analysis.

[0109] Step S2: Data preprocessing.

[0110] The data preprocessing module 210 performs statistical outlier removal and coordinate normalization on the received raw 3D point cloud data. This step eliminates noise in the raw data and unifies the point cloud into a normalized spatial coordinate system, providing a high-quality data foundation for subsequent feature extraction.

[0111] Step S3: Multi-source feature fusion.

[0112] The multi-source feature fusion module 220 spatially registers the two-dimensional image data with the normalized point cloud data and calculates a fused feature vector for each spatial point. The function of this step is to uniformly represent information from different sensors and different data modalities at the spatial point level, so that each point carries dual attribute information of geometry and texture.

[0113] Step S4: Manifold-constrained hypervoxel segmentation.

[0114] The manifold constraint segmentation module 230 takes the fused feature vectors of each spatial point as input, constructs a graph structure on the point cloud, and uses geodesic distance as a manifold constraint for supervoxel clustering. The function of this step is to discretize the continuous point cloud data into independent regions that physically correspond to a single stone, which is the basis for subsequent particle size statistics.

[0115] Step S5: Particle size parameter extraction.

[0116] The particle size extraction module 240 fits an oriented bounding box to each supervoxel region, extracts the major, median, and minor axis dimensions, and calculates representative particle size values. This step quantifies the geometry of each segmented region into statistically significant particle size parameters.

[0117] Step S6: Grading statistics.

[0118] The gradation statistics module 250 groups and statistically analyzes the representative particle size values ​​of all ultravoxel regions according to preset intervals, generating a particle size distribution histogram and gradation curve. This step summarizes the particle size information of individual stones into statistical information reflecting the overall gradation status of the entire storage area.

[0119] Step S7: Gap identification.

[0120] After marking the space occupied by the supervoxel region as occupied, the gap identification module 260 performs voxelization, connected component analysis, distance transformation, and local maximum extraction on the remaining space to identify gap regions and calculate equivalent pore size and volume. The function of this step is to quantitatively describe the distribution of voids inside the rockfill mass.

[0121] Step S8: Output the results.

[0122] The report generation module 270 integrates the statistical results of gradation and the gap identification results to generate a structured inspection report and outputs it to the construction quality control system. The function of this step is to present the analysis results in an engineering-usable format, providing data support for construction decisions.

[0123] The specific technical implementation of each step is described in detail below. The sub-modules involved in each step and their connection and cooperation relationship have been explained in the above system structure.

[0124] In some embodiments of the present invention, the specific implementation of step S1 is as follows: The data acquisition terminal 100 in the system adopts a gantry-type mobile platform, equipped with a line-scanning 3D laser scanner 110 and multiple high-resolution industrial cameras 120. The 3D laser scanner 110 has a scanning line frequency of 200 contour lines per second, with each contour line containing 2048 measurement points. The industrial cameras 120 maintain hardware-level synchronization with the scanning pulses of the 3D laser scanner 110 through a high-precision synchronization trigger circuit 130, ensuring that each frame of image data is precisely aligned with the corresponding laser point cloud data in timestamps. As an example, the data acquisition terminal 100 can move at a constant speed of 0.05 m / s to 0.15 m / s along a serpentine path to achieve full coverage scanning of the entire warehouse surface.

[0125] This step ensures strict temporal alignment of multi-source data through hardware-level synchronization, providing a foundation for subsequent spatial registration and feature fusion; and achieves full coverage of the large warehouse area through mobile scanning, overcoming the limitation of limited coverage of static scanning.

[0126] In some embodiments of the present invention, the specific implementation of step S2 is as follows: The data preprocessing module 210 in the system performs the following processing on the raw 3D point cloud data in sequence: First, outlier removal is performed statistically. This submodule performs statistical analysis on the neighborhood of each spatial point, calculating the mean and standard deviation of the distances between that point and its K nearest neighbors. For example, the K value can be set to 50. If the average distance between a point and its neighbors exceeds a preset multiple of the global average distance (e.g., 2 standard deviations), that point is marked as an outlier and filtered out. This operation eliminates isolated noise points caused by laser multipath reflection or dust scattering.

[0127] Second, coordinate normalization. This submodule uses the data from the positioning and attitude sensing unit 140 of the data acquisition terminal 100 to transform the denoised point cloud from the sensor's local coordinate system to the unified construction coordinate system of the dam, thereby obtaining a normalized point cloud dataset.

[0128] This step eliminates noise points in the original data through statistical filtering and eliminates the impact of terminal pose changes on data consistency during the moving scan process through coordinate normalization, providing high-quality, spatially consistent data input for subsequent processing.

[0129] In some embodiments of the present invention, the specific implementation of step S3 is as follows: like Figure 6 As shown, the multi-source feature fusion module 220 performs the following sub-steps: Sub-step S3-1: Camera-LiDAR spatial registration. The spatial registration submodule 221 uses a pre-calibrated camera-LiDAR extrinsic parameter matrix to project each spatial point in the 3D point cloud onto the image pixel coordinate system of the industrial camera 120. For points that simultaneously fall within the overlapping fields of view of multiple cameras, the camera closest to the point is selected as the main projection camera based on the distance from the point to the optical center of each camera.

[0130] Sub-step S3-2: 3D geometric feature extraction. The 3D geometric feature extraction sub-module 222 uses a radius search method to determine the neighborhood range for each spatial point in the normalized point cloud dataset. For example, the search radius can be set to 50 mm. It calculates the covariance matrix of the neighborhood point set and solves for the eigenvalues ​​and eigenvectors. The eigenvector corresponding to the smallest eigenvalue is used as the local normal vector of that point; the ratio of the smallest eigenvalue to the sum of the three eigenvalues ​​is used as the local surface change of that point. This ratio is a commonly used curvature engineering approximation in point cloud processing—a ratio close to 0 indicates that the area where the point is located is close to a plane, while a larger ratio indicates drastic surface changes, effectively distinguishing between the stone surface (gentle changes) and stone boundaries or gaps (drastic changes).

[0131] Sub-step S3-3: Two-dimensional texture feature extraction. The two-dimensional texture feature extraction sub-module 223 calculates the local texture descriptor around each pixel at the image pixel coordinates corresponding to each three-dimensional spatial point. This scheme uses the Dense Scale Invariant Feature Transform (Dense SIFT) method, which calculates the SIFT descriptor at each pixel location, instead of only calculating it at keypoints detected in scale space as in the standard SIFT algorithm. Dense SIFT can provide corresponding texture features for each spatial point in the point cloud, making it suitable for 3D-2D fusion scenarios requiring point-by-point feature alignment. Specifically, within a 16×16 pixel neighborhood window, gradient orientation histograms of 4×4 sub-regions in 8 directions are statistically analyzed to form a 128-dimensional feature vector.

[0132] Sub-step S3-4: Fusion Feature Vector Assembly. The fusion feature vector assembly sub-module 224 concatenates and splices the 3D spatial coordinates (3D), local normal vector (3D), local surface variation (1D), and texture descriptor (128D) of each spatial point to form a 135-dimensional fusion feature vector. Before splicing, each component is normalized to eliminate dimensional differences. The normalization method uses Z-score standardization, which involves subtracting the mean from each feature component and then dividing by the standard deviation.

[0133] This step precisely maps the rich texture information of the 2D image to every spatial point in the 3D point cloud through spatial registration, enabling each point to simultaneously carry geometric attributes (position, orientation, curvature) and texture attributes (surface pattern, color distribution). This deep fusion of multi-source heterogeneous features allows subsequent segmentation algorithms to make judgments based on both geometric and texture boundaries. Even if the point cloud density at the contact boundary of the stone is insufficient, texture differences can still provide an effective basis for segmentation, fundamentally overcoming the under-segmentation problem of pure point cloud methods.

[0134] In some embodiments of the present invention, the specific implementation of step S4 is as follows: like Figure 7 As shown, the manifold constraint partitioning module 230 performs the following sub-steps: Sub-step S4-1: Graph Structure Construction. The graph structure construction sub-module 231 uses each spatial point in the normalized point cloud dataset as a node, searches for the K nearest neighbors of each node in the feature space (for example, K can be set to 20), and establishes edges connecting adjacent nodes. Each edge in the graph is assigned two sets of weights: one is the feature space weight, defined as the weighted Euclidean distance between the fused feature vectors of the two endpoints, used for similarity measurement during clustering assignment; the other is the spatial weight, defined as the Euclidean distance between the two endpoints in three-dimensional space, used for calculating geodesic distance. The edge weight for geodesic distance uses spatial Euclidean distance rather than feature distance because the physical meaning of geodesic distance is the geometric path length on the point cloud surface, and it should reflect the actual distance in space rather than feature similarity.

[0135] Sub-step S4-2: Seed point initialization. The seed point initialization submodule 232 uniformly samples initial seed points in the point cloud space, with the sampling interval set according to half of the minimum detection particle size. For example, if the minimum detection particle size is set to 100 mm, the sampling interval is set to 50 mm. Each seed point is assigned a unique hypervoxel label.

[0136] Sub-step S4-3: Manifold-constrained clustering assignment. The manifold-constrained clustering assignment submodule 233 expands outwards layer by layer along the edges of the graph from each seed point, assigning candidate nodes to the seed point that minimizes the feature distance cost function. The feature distance cost function is defined as the weighted Euclidean distance between the fused feature vector F(p) of candidate point p and the fused feature vector F(s) of seed point s: d_feature(p,s) = [Σ_{i=1}^{135} w_i · (F_i(p) - F_i(s)) 2 ]^(1 / 2); Candidate point p is assigned to the seed point s that minimizes d_feature(p,s), but must also satisfy the geodesic distance constraint.

[0137] Unlike standard supervoxel segmentation, in this step, a candidate node is allowed to be assigned to a certain sub-node only if the shortest path length (i.e., geodesic distance) from the candidate node to the seed node along the graph edge does not exceed a preset threshold. The geodesic distance is calculated by running Dijkstra's shortest path algorithm on the graph, using the aforementioned spatial weights (i.e., Euclidean distance in 3D space) as edge weights. The physical meaning of this preset threshold is the expected minimum boundary interval between adjacent stones, for example, it can be set to 30 mm. Generally, the geodesic distance threshold should be no less than half and no more than the seed node interval to ensure that clustering covers the stone surface without crossing gaps.

[0138] As an alternative, the manifold-constrained clustering allocation submodule 233 can also use an adaptive threshold based on local point cloud density instead of the fixed threshold mentioned above. Specifically: for the current candidate point p, the number of point clouds N(p) in the spherical neighborhood centered at p with a preset radius R (e.g., R=50 mm) as the search radius is counted, and N(p) is used as the local point cloud density value of the point; then, the adaptive manifold constraint threshold T(p) of the candidate point is calculated based on the local point cloud density N(p).

[0139] There is a positive correlation between the adaptive threshold T(p) and the local point cloud density N(p): T(p) = T_base × (N(p) / N_avg), Where T_base is the baseline threshold (e.g., 30 mm) and N_avg is the average density of the point cloud across the entire cloud area.

[0140] When the local density is higher than the average density, the threshold is appropriately increased to allow for geodesic distance fluctuations within the same stone due to dense point clouds; when the local density is lower than the average density, the threshold is appropriately decreased to prevent false connections across gaps due to sparse point clouds. A candidate point p is only allowed to be assigned to the supervoxel containing the seed point s if the geodesic distance d_geodesic(p,s) from the candidate point p to the seed point s is ≤ T(p).

[0141] This adaptive threshold mechanism enables the segmentation algorithm to dynamically adjust the constraints based on the actual point cloud density distribution on the surface of the riprap. In areas with dense point clouds on the stone surface, the constraints are appropriately relaxed to maintain the integrity of the stone's interior; in areas with sparse point clouds between the stones, the constraints are strictly tightened to prevent erroneous merging across gaps.

[0142] Sub-step S4-4: Iterative optimization. The iterative optimization submodule 234 iteratively optimizes the initial clustering assignment. In each iteration, the center position of each supervoxel is recalculated and the edge nodes are reassigned to the nearest supervoxel center. Specifically, for each supervoxel region, its center position is defined as the mean vector of the fused feature vectors of all spatial points within the supervoxel region (i.e., the centroid in the feature space), rather than the geometric center in three-dimensional space. For example, if a supervoxel region contains 100 spatial points, and the fused feature vectors of each point are F(1), F(2), ..., F(100), then the center position of the supervoxel is C = (F(1) + F(2) + ... + F(100)) / 100. The iteration termination condition can be set to the proportion of nodes whose supervoxel boundaries have moved being less than 0.1% or the number of iterations reaching 20.

[0143] As a further optimization, after iterative optimization, the manifold constraint segmentation module 230 can also perform a supervoxel boundary refinement step based on an energy function. Specifically: all spatial points located in the supervoxel boundary region are identified, and the criterion for determining a boundary point is that at least two of its K nearest neighbors belong to different supervoxel labels. Then, an energy function E(x, l) is constructed for each boundary point x, representing the global energy cost of assigning the boundary point x to the supervoxel label l. The energy function consists of the following three weighted sums: E(x, l) = α·E_feature(x, l) + β·E_smooth(x, l) + γ·E_prior(x, l); Wherein: E_feature(x, l) is the feature consistency term, which measures the weighted Euclidean distance between the fused feature vector of boundary point x and the centroid feature vector of supervoxel l; E_smooth(x, l) is the smoothness term, which measures the degree of label consistency between boundary point x and its neighboring points after the boundary point x is assigned to supervoxel l. Specifically, E_smooth(x, l) = Σ_{y∈N(x)} [l(x) ≠ l(y)] × w(x,y), where N(x) is the set of neighboring points of x, and w(x,y) is the weight based on the spatial distance and feature similarity between x and y; E_prior(x, l) is the prior constraint term, whose physical meaning is: when the boundary point x is located at the intersection of two supervoxels, x is preferentially assigned to the supervoxel on the side with a more gradual curvature change. This prior constraint is based on the geometric characteristics of the riprap, where although the surface of a single stone is rough, its surface change is generally smaller than the abrupt change at the gaps between stones.

[0144] The weighting coefficients α, β, and γ satisfy α + β + γ = 1, with recommended values ​​in the range of α ∈ [0.3, 0.5], β ∈ [0.25, 0.4], and γ ∈ [0.15, 0.3], preferably α = 0.4, β = 0.35, and γ = 0.25. Each weighting coefficient can be obtained through grid search optimization on typical rockfill point cloud samples. For example, several sets of manually labeled rockfill point cloud samples can be prepared, and combinations of α = 0.3, 0.4, 0.5, β = 0.25, 0.35, 0.4, and γ = 0.15, 0.25, 0.3 can be tried respectively. The optimal combination is selected based on the segmentation accuracy (e.g., the consistency rate of boundary point annotations).

[0145] This boundary refinement method, based on the energy function of all boundary points, employs a graph cut algorithm or belief propagation algorithm to globally optimize the energy function of all boundary points. For each boundary point, a label assignment scheme that minimizes the total energy ΣE(x, l) is selected, and the supervoxel labels of the boundary points are updated according to the solution results. By comprehensively considering three dimensions—feature consistency, boundary smoothness, and prior geometric constraints—this method achieves globally optimal adjustment of the supervoxel boundaries, effectively overcoming the problem of local similarity measurement failure caused by the roughness of the rockfill surface.

[0146] In this step of the scheme, the geodesic distance constraint ensures that the clustering process follows the manifold geometry of the point cloud surface. On the stone surface, the geodesic distance is close to the Euclidean distance; however, at the gaps between stones, due to the lack of connection points, the geodesic distance increases sharply and exceeds the threshold, thus preventing clustering from crossing the gaps. This is fundamentally different from the standard method that only uses Euclidean distance or feature space distance: the standard method cannot perceive the topological structure of the point cloud surface and is prone to mistakenly merging stones on both sides of the gap; while the manifold constraint of this invention strictly restricts the segmentation boundary within the physical boundary of the stone, fundamentally preventing mistaken merging across gaps. Iterative optimization causes the segmentation boundary to converge to the geometric position of a local curvature maxima or abrupt change in normal vector—these positions precisely correspond to the contact boundary between adjacent stones.

[0147] In some embodiments of the present invention, the specific implementation of step S5 is as follows: like Figure 7 As shown, the particle size extraction module 240 performs directional bounding box fitting for each supervoxel region.

[0148] Sub-step S5-1: Principal direction estimation. The principal direction estimation submodule calculates the covariance matrix of the coordinates of all points within the supervoxel region, and solves for the eigenvalues ​​and eigenvectors. The three eigenvectors correspond to the directions of the major axis, median axis, and minor axis, respectively.

[0149] Sub-step S5-2: Size Extraction. The size extraction submodule rotates the point coordinates to a local coordinate system aligned with the principal direction, calculates the coordinate spans along the three axes, and obtains the major axis size L, the median axis size M, and the minor axis size S. It should be noted that since the sign (positive or negative) of the eigenvectors of the covariance matrix is ​​arbitrary, but the coordinate spans along the three axes (maximum value minus minimum value) are unaffected by the eigenvector sign, this rotation process has no impact on the subsequent particle size extraction results. For example, inverting a certain eigenvector will change the sign of the projected coordinates of all points on that axis, but the coordinate span (maximum value minus minimum value) remains unchanged; therefore, the major axis size L is unaffected.

[0150] Sub-step S5-3: Representative particle size calculation. The representative particle size calculation sub-module calculates the representative particle size value based on the definition of equivalent diameter of an equal-volume sphere: d_eq = (L × M × S)^(1 / 3).

[0151] This step-by-step approach can accurately capture the three-dimensional morphology of non-spherical stones through directional bounding box fitting; the equivalent diameter of a sphere of equal volume is a standard equivalent index for engineering screening tests, ensuring the comparability of the output results of this invention with those of traditional methods.

[0152] In some embodiments of the present invention, the specific implementation of step S6 is as follows: The gradation statistics module 250 groups representative particle size values ​​according to preset particle size intervals, accumulates the volume of all supervoxel regions falling into each interval (approximately using the volume of the oriented bounding box L×M×S), and plots a particle size distribution histogram and gradation curve (the horizontal axis represents the particle size interval, and the vertical axis represents the cumulative volume percentage). For example, the boundaries of the particle size intervals can be set to 100mm, 150mm, 200mm, 300mm, 400mm, 500mm, 600mm, 800mm, 1000mm, etc.

[0153] As an alternative, the gradation statistics module 250 can also use a multi-scale segmentation result voting fusion method to replace single-scale gradation statistics. Specifically: M different segmentation parameters are set (M≥3), and the seed point interval of each group of parameters is different. For example: fine scale (seed point interval = 1 / 3 of the minimum detection particle size, such as 33mm), medium scale (seed point interval = 1 / 2 of the minimum detection particle size, such as 50mm), and coarse scale (seed point interval = 1 / 1 of the minimum detection particle size, such as 100mm). Each group of parameters independently runs the complete ultravoxel segmentation and particle size extraction process to obtain M groups of particle size identification results.

[0154] For each segmentation result at each scale, each supervoxel region is treated as a stone candidate, and its spatial location (centroid coordinates), representative grain size, and occupied spatial range (oriented bounding box) are recorded. Then, a spatial correspondence is established between stone candidate sets at different scales: for stone candidate A at the fine scale and stone candidate B at the coarse scale, if the centroid of A is located inside the oriented bounding box of B, and the spatial overlap (cross-over-union ratio) of the two exceeds a preset threshold (e.g., 0.3), then A and B are determined to be different segmentation representations of the same stone at different scales.

[0155] For a set of candidate stones with cross-scale correspondences, a voting mechanism is used to determine the final identification result of the stone: (1) Existence voting -- If the location is identified as stone at at least two scales, the existence of stone at the location is confirmed; if it is identified as stone only at a single scale, it is marked as "pending confirmation" and enters the review process.

[0156] (2) Particle Size Voting – For confirmed stone materials, the representative particle size values ​​at each scale are weighted and averaged. The weighting principle is as follows: the medium scale (seed point interval = 1 / 2 of the minimum detected particle size) has the highest weight, and the weights of the fine and coarse scales decrease sequentially. The final representative particle size value d_final = Σ(w_i × d_i) / Σw_i, where d_i is the representative particle size value at the i-th scale, and w_i is the corresponding weight.

[0157] (3) Boundary voting - The final spatial boundary of the stone is obtained by weighted averaging of the oriented bounding boxes at each scale, with the weight allocation being the same as that of the particle size voting. In addition, for spatial regions that are identified as multiple stones at the fine scale and as a single stone at the coarse scale (i.e., over-segmented fragments on the surface of large stones), the voting fusion mechanism correctly identifies them as a single stone by combining existence voting and boundary voting, effectively eliminating the influence of over-segmentation.

[0158] This multi-scale voting fusion method effectively solves the recognition dilemma of single-scale methods in scenarios with large-span particle size distribution by independently performing segmentation at different scales and integrating the recognition results of each scale.

[0159] This step-by-step approach summarizes the particle size information of individual stones into statistical information reflecting the overall gradation of the entire storage area, providing a quantitative basis for project quality evaluation.

[0160] In some embodiments of the present invention, the specific implementation of step S7 is as follows: like Figure 8 As shown, the gap recognition module 260 performs the following sub-steps: Sub-step S7-1: Binary Voxel Mesh Construction. The binary voxel mesh construction submodule 261 divides the space occupied by the rockfill into a three-dimensional mesh, for example, the mesh resolution can be set to 10 mm. The spatial range of voxelization is defined as follows: In the horizontal direction, a preset margin is extended outward from the silo surface design boundary (e.g., 100 mm outward from the silo surface boundary) to ensure coverage of the entire rockfill area; in the vertical direction, the lower limit is the silo surface reference plane, and the upper limit is the maximum rockfill height. It is determined whether the center point of each voxel falls inside the oriented bounding box of any super-voxel region—if it falls in, a value of 1 is assigned (rock occupancy); if it does not fall in, a value of 0 is assigned (candidate gap). Voxels below the silo surface reference plane and exceeding the maximum rockfill height are excluded. The maximum rockfill height is obtained by statistically analyzing the difference between the maximum coordinate value of the point cloud on the rockfill surface in the vertical direction and the elevation of the silo surface reference plane.

[0161] Sub-step S7-2: 3D Connectivity Analysis. The 3D Connectivity Analysis submodule 262 performs a 26-neighborhood 3D connectivity analysis on all voxels with a value of 0, clustering spatially adjacent voxels with 0 values ​​into independent connected components. Small connected components with volumes smaller than a preset threshold are filtered out (considered noise artifacts). For example, when the voxel resolution is 10 mm, the volume of a single voxel is 1000 cubic millimeters. A volume threshold of 100 voxels (i.e., 100,000 cubic millimeters) can be set to filter out connected components smaller than this threshold.

[0162] Sub-step S7-3: 3D distance transformation. The 3D distance transformation submodule 263 calculates the shortest Euclidean distance from each 0-value voxel within each preserved gap connected region to the nearest 1-value voxel (stone boundary). A fast distance transformation algorithm with two scans is used, and its time complexity is linearly related to the total number of voxels.

[0163] Sub-step S7-4: Equivalent pore size and volume extraction. The equivalent pore size and volume extraction submodule 264 searches for local maxima in the distance transformation results of each gap's connected domain, using these as the equivalent pore size at that location (physically meaning the largest sphere diameter that can freely pass through that point). The minimum of all local maxima distance values ​​is taken as the effective equivalent pore size of the gap. This indicator directly determines whether self-compacting concrete can smoothly pass through the gap—the narrowest point is the key to determining fluidity. Gap volume = total number of voxels × volume of a single voxel.

[0164] This process discretizes the continuous space using voxelization, making it computationally comprehensible; it clusters scattered zero-value voxels into physically meaningful independent gaps through connected component analysis; and it assigns the geometric quantity of "distance to the boundary" to each voxel through distance transformation. The local maxima of the distance transformation correspond precisely to the "widest point" of the gap—that is, the equivalent aperture. This four-step process forms a complete chain for quantitative gap identification, achieving for the first time a quantitative description of the interparticle gaps within a rockfill mass.

[0165] In some embodiments of the present invention, the specific implementation of step S8 is as follows: like Figure 1 As shown, the report generation module 270 integrates the gradation statistics and gap identification results to generate a structured inspection report containing a particle size distribution histogram, a gradation curve (superimposed between the upper and lower limit envelopes of the design gradation curve), a calculated rockfill ratio, a gap volume distribution histogram, a gap equivalent pore size distribution map, and a gradation deviation early warning list. The report is then output to the construction quality control system through a standardized data interface (e.g., a RESTful API based on HTTP protocol and JSON data format).

[0166] This step presents the complex analysis results in an engineering-usable format, providing intuitive and quantitative data support for construction decisions.

[0167] As shown above, the entire processing flow is completed in real time at the edge. Data is initially processed locally at the acquisition terminal and then transmitted to the edge computing unit without being uploaded to the cloud, with end-to-end latency controlled in the minute range. This "end-edge" collaborative computing architecture ensures real-time performance while reducing dependence on network bandwidth, enabling the system to operate independently at the construction site.

[0168] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A real-time identification method for particle size distribution and gap distribution of rockfill based on three-dimensional laser scanning, executed by an edge computing unit, wherein the edge computing unit is communicatively connected to a data acquisition terminal, characterized in that, The method includes: (1) Acquire the three-dimensional point cloud data and two-dimensional image data of the rockfill surface synchronously collected by the data acquisition terminal; (2) Preprocess the three-dimensional point cloud data to obtain standardized point cloud data; (3) Spatial registration is performed between the two-dimensional image data and the normalized point cloud data to generate a fusion feature vector for each spatial point. The fusion feature vector includes the three-dimensional geometric features and two-dimensional texture features of the spatial point. (4) Based on the fusion feature vector of each spatial point, supervoxel clustering is performed with geodesic distance as manifold constraint condition to divide the point cloud on the surface of the rockfill into multiple supervoxel regions corresponding to each stone. (5) Extract three-dimensional size parameters for each of the supervoxel regions and determine the representative particle size value of the stone according to the preset particle size definition rules; (6) Statistically analyze the distribution of representative particle size values ​​in all supervoxel regions to generate particle size distribution information; (7) After marking the space occupied by each supervoxel region as occupied, perform three-dimensional morphological analysis on the remaining space to identify the distribution of interparticle gaps inside the rockfill and calculate the equivalent pore size and volume parameters of each gap region. (8) Generate and output a test report based on the particle size distribution information and the gap distribution information.

2. The method according to claim 1, characterized in that, The three-dimensional geometric features include the local normal vector direction and the local surface variation of the spatial point; the two-dimensional texture features are local texture descriptors extracted based on the corresponding pixel position of the spatial point in the two-dimensional image data.

3. The method according to claim 1, characterized in that, The geodesic distance in step (4) is obtained by calculating the shortest path on the graph structure constructed from the point cloud data. The weight of the edge used for geodesic distance calculation in the graph structure is the Euclidean distance between the two end nodes in three-dimensional space. The manifold constraint condition is: the geodesic distance between the candidate point and the seed point does not exceed a preset threshold. The physical meaning of the preset threshold is the expected minimum boundary interval between adjacent stones.

4. The method according to claim 1, characterized in that, The three-dimensional morphological analysis in step (7) includes: The space occupied by the riprap is voxelized into a binary 3D mesh; Perform connected component analysis on the binary 3D mesh to identify regions that are empty and spatially connected as gap regions; Perform a three-dimensional distance transformation on each gap region; The local maxima in the three-dimensional distance transformation result are used as the equivalent aperture of the gap region.

5. The method according to claim 1, characterized in that, The representative particle size value in step (5) is the equivalent diameter of the volumetric sphere of the stone corresponding to the supervoxel region, which is calculated by the geometric mean of the major axis, middle axis and minor axis dimensions of the oriented bounding box of the supervoxel region.

6. A real-time identification system for particle size distribution and gap distribution of rockfill based on three-dimensional laser scanning, characterized in that, include: A data acquisition terminal, which is equipped with a 3D laser scanner and at least one industrial camera; The data acquisition terminal is configured to move along a preset path above the rockfill concrete dam surface, the 3D laser scanner is configured to acquire 3D point cloud data of the rockfill body on the dam surface, and the industrial camera is configured to acquire 2D image data of the rockfill surface synchronously with the 3D laser scanner. An edge computing unit, which is communicatively connected to the data acquisition terminal; the edge computing unit is equipped with: A data preprocessing module is configured to perform denoising and spatial normalization processing on the three-dimensional point cloud data to obtain normalized point cloud data. A multi-source feature fusion module is configured to spatially register the two-dimensional image data with the normalized point cloud data, and generate a fused feature vector for each spatial point, wherein the fused feature vector includes the three-dimensional geometric features and two-dimensional texture features of the spatial point. The manifold constraint segmentation module is configured to perform supervoxel clustering based on the fused feature vectors of each spatial point, using geodesic distance as the manifold constraint condition, to segment the point cloud on the surface of the rockfill into multiple supervoxel regions corresponding to each stone. The particle size extraction module is configured to extract three-dimensional size parameters for each of the supervoxel regions and determine the representative particle size value of the stone according to a preset particle size definition rule. A gradation statistics module is configured to statistically analyze the distribution of representative particle size values ​​in all hypervoxel regions and generate particle size gradation information. The gap identification module is configured to mark the space occupied by each supervoxel region as occupied, perform three-dimensional morphological analysis on the remaining space, identify the distribution of particle gaps inside the rockfill, and calculate the equivalent pore size and volume parameters of each gap region. A report generation module is configured to generate and output a test report based on the particle size distribution information and the gap distribution information.

7. The system according to claim 6, characterized in that, The three-dimensional geometric features include the local normal vector direction and the local surface variation of the spatial point; the two-dimensional texture features are local texture descriptors extracted based on the corresponding pixel position of the spatial point in the two-dimensional image data.

8. The system according to claim 6, characterized in that, In the manifold constraint segmentation module, the geodesic distance is obtained by calculating the shortest path on the graph structure constructed from point cloud data. The weight of the edge used for geodesic distance calculation in the graph structure is the Euclidean distance between the two endpoints in three-dimensional space. The manifold constraint condition is: the geodesic distance between the candidate point and the seed point does not exceed a preset threshold. The physical meaning of the preset threshold is the expected minimum boundary interval between adjacent stones.

9. The system according to claim 6, characterized in that, In the gap recognition module, the three-dimensional morphological analysis includes: The space occupied by the riprap is voxelized into a binary 3D mesh; Perform connected component analysis on the binary 3D mesh to identify regions that are empty and spatially connected as gap regions; Perform a three-dimensional distance transformation on each gap region; The local maxima in the three-dimensional distance transformation result are used as the equivalent aperture of the gap region.

10. The system according to claim 6, characterized in that, The representative particle size value is the equivalent diameter of the volumetric sphere of the stone corresponding to the hypervoxel region, which is calculated by the geometric mean of the major axis, middle axis and minor axis dimensions of the oriented bounding box of the hypervoxel region.

Citation Information

Patent Citations

  • Transparent Modeling Methods for Rockfill Dams

    CN110363855B

  • Material pile lumpiness analysis method and device, electronic equipment and storage medium

    CN117237395A