Rockfill material grading detection method based on three-dimensional point cloud
By employing a 3D point cloud-based method for detecting the gradation of riprap, and combining deep learning and machine learning, this method solves the problems of time-consuming, labor-intensive, and low-precision segmentation of traditional detection methods, achieving efficient and accurate gradation detection.
Patent Information
- Application Number
- CN202610021980.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-08
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional methods for detecting the gradation of rockfill are time-consuming, labor-intensive, and inefficient, making it difficult to meet the high standards required for mechanized and intelligent construction of modern earth-rock dams. Furthermore, existing traditional point cloud segmentation methods have low segmentation accuracy and poor model generalization ability in complex scenarios.
A 3D point cloud-based method for detecting the gradation of riprap is adopted. This method involves collecting point cloud data, generating depth images, segmenting particles using an instance segmentation model, mapping back to the 3D point cloud, extracting the 3D morphological features of the particles, and then combining this with machine learning regression methods to plot the gradation curve.
It improves the accuracy and generalization ability of particle identification, overcomes the influence of light, and achieves efficient and accurate gradation detection. The detection results have a high degree of agreement with the actual gradation curve.
Smart Images

Figure CN122023281A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of rockfill gradation detection technology, specifically relating to a method for rockfill gradation detection based on three-dimensional point clouds. Background Technology
[0002] Reservoir dams are crucial infrastructure for optimizing water resource allocation and ensuring flood control safety. Rockfill dams, with their advantages of readily available materials, strong adaptability, good earthquake resistance, low carbon footprint, short construction period, and low cost, have become one of the fastest-growing and most competitive dam types. The stability and seepage resistance of rockfill dams primarily depend on the gradation composition of the dam body's filling materials; a reasonable gradation is crucial for ensuring the overall quality and safe operation of water conservancy projects. Traditional methods for testing dam material gradation mainly rely on sieve analysis, which suffers from problems such as being time-consuming, labor-intensive, inefficient, and having delayed feedback, making it difficult to meet the high standards required for mechanized and intelligent construction of modern earth-rock dams.
[0003] With the development of 3D laser scanning technology, point cloud-based gradation detection methods have gradually emerged. Compared to traditional image recognition methods, point cloud data can more completely preserve the 3D geometric information of particles, resulting in higher gradation conversion accuracy. However, existing traditional point cloud segmentation methods mostly rely on manually designed features, making it difficult to handle complex scenes with densely stacked particles and significant scale differences, leading to problems such as low segmentation accuracy and poor model generalization ability. In recent years, deep learning has made significant progress in the application of point cloud processing, with deep learning-based segmentation algorithms exhibiting stronger recognition capabilities and adaptability in complex environments.
[0004] Therefore, there is an urgent need to propose a method for detecting the gradation of riprap based on deep learning and 3D point cloud technology to solve the above problems. Summary of the Invention
[0005] To address the aforementioned shortcomings in existing technologies, this invention aims to provide a method for detecting the gradation of riprap based on three-dimensional point clouds, thereby solving the problem of low segmentation accuracy in traditional point cloud segmentation methods.
[0006] The technical solution adopted in this invention is as follows:
[0007] A method for detecting the gradation of riprap based on three-dimensional point clouds includes the following steps:
[0008] S1. Collect point cloud data of the riprap material;
[0009] S2. Project the point cloud of the riprap along the top-view direction to generate the corresponding depth image;
[0010] S3. Input the depth map into the instance segmentation model to obtain the segmentation mask for each stone particle;
[0011] S4. Map the segmentation mask back to the 3D point cloud and extract the 3D morphological features of each particle.
[0012] S5. Predict volume and particle size based on the three-dimensional features of particles, and plot the gradation curve.
[0013] Furthermore, step S2 specifically includes the following steps:
[0014] S21. Rotate the point cloud of the riprap to a standard horizontal coordinate system for depth projection.
[0015] First, the Random Sample Consensus (RANSAC) algorithm is used to perform plane fitting on the point cloud to extract the ground normal vector. For any point... P i Its distance to the fitting plane is: ;
[0016] When Di < ϵ (preset threshold), P i Points are considered interiors of the current plane. The RANSAC algorithm iteratively searches for the plane with the most interiors, ultimately determining the ground normal vector. .
[0017] Then, the point cloud is rotated to the horizontal coordinate system using the Rodrigues rotation formula, with the rotation matrix R2 as follows: ; ; ;
[0018] S22. Project the rotated point cloud along the Z-axis onto the XOY plane to construct a regular two-dimensional mesh.
[0019] Set the grid cell size to Δx×Δy, and iterate through all points, determining their index position in the grid based on their X and Y coordinates: ;
[0020] Within each grid cell, the maximum height projection principle is used to retain the point with the largest Z value in that region as the depth value, thereby generating the corresponding depth matrix: ;
[0021] The depth matrix D is normalized and converted into a grayscale image to generate a corresponding two-dimensional depth map. The grayscale value of each pixel reflects the surface height information of the region and serves as the input for the subsequent instance segmentation model.
[0022] Furthermore, the preferred instance segmentation model used in S3 is YOLOv8-seg, which is used to perform instance-level segmentation of stone particles in the depth map.
[0023] Furthermore, step S4 specifically includes the following steps:
[0024] S41. After obtaining the depth map segmentation mask, inversely project the depth map mask back into the point cloud mask. For any pixel position in the depth map... and depth value The corresponding point cloud coordinate calculation formula is as follows: ;
[0025] S42. Extract three-dimensional features from each segmented particle point cloud and calculate three-dimensional geometric features such as major axis, second major axis, minor axis, surface area, volume, sphericity, and roundness.
[0026] 1) Principal component analysis was used to calculate the principal size orientation of the particles.
[0027] The principal scale of rock particles is typically composed of three mutually perpendicular axial dimensions: the major axis (L), the second major axis (W), and the minor axis (T). Principal component analysis (PCA) is used to calculate the point cloud covariance matrix and solve for its eigenvalues and eigenvectors. This allows the particle's point cloud data to be projected onto the direction of maximum variance, thus aligning the projected coordinate system with the principal scale direction of the particle. The differences in the range of the projected point cloud coordinates along the three axes then correspond to the major, second major, and minor axes of the particle, respectively.
[0028] 2) For the volume V and surface area S of the rock particle point cloud, the Delaunay triangulation algorithm is used to construct a surface mesh model of the particles for calculation.
[0029] For a given triangular mesh, the position vectors of any vertex of a triangular face are respectively If the origin is taken as the fourth vertex of the tetrahedron, then the total volume of the entire mesh is... V The sum of the volumes of all tetrahedrons: ; ;
[0030] surface area S This can be obtained by summing the areas of all the triangular faces: ; ;
[0031] 3) Sphericity is used to measure the similarity of the shape of rock particles to an ideal sphere. It is defined as the ratio of the surface area of a particle to the surface area of a sphere with the same volume. ;
[0032] In the formula, V and S It is distributed as the volume and surface area of rock particles.
[0033] 4) Roundness is used to measure the smoothness of the particle's edges. Wadell et al. defined three-dimensional roundness as the average of the ratios of the radii of curvature of each edge on the particle profile to the radius of the largest inscribed sphere. ;
[0034] In the formula: r i For the first i The radius of a sphere fitted by a set of edges; R The radius of the largest inscribed sphere of the particle; N This represents the number of sharp edges of the particles.
[0035] 5) Radius of the largest inscribed sphere of the particle R 1 and the smallest circumsphere of the particle R 2. Using the Euclidean distance field calculation method, an internal point set is generated by uniformly sampling inside the particle. P in For each internal sampling point, find its minimum and maximum Euclidean distances to the outer surface of the particle. The maximum value of the minimum distances corresponding to all sampling points is the radius of the maximum inscribed sphere. R 1. The minimum of the maximum distances corresponding to all sampling points is the radius of the minimum circumscribed sphere. R 2: ; .
[0036] Furthermore, S5 specifically includes the following steps:
[0037] S51. Construct a volume prediction model based on the three-dimensional morphological features of particles. Use the three-dimensional geometric parameters of particles, such as major axis, second major axis, minor axis, surface area, and volume, as input features, and use machine learning regression methods to estimate the complete volume of particles.
[0038] Preferably, the volume prediction model is a regression model based on LightGBM.
[0039] S52. After completing the particle volume prediction, the secondary major axis of the particles is used as the basis for particle size division. Particle size intervals are set according to the actual needs of the project, and the proportion of particle volume in each particle size interval is calculated. The discrete gradation curve of the rockfill is plotted with particle size as the abscissa and volume distribution percentage as the ordinate.
[0040] Compared with the prior art, the present invention has the following advantages and beneficial effects: The present invention performs gradation conversion based on three-dimensional point cloud data, which effectively overcomes the shortcomings of image recognition methods such as being greatly affected by illumination and the inability of two-dimensional information to fully reflect the three-dimensional morphology of particles; by introducing a deep learning instance segmentation model, the present invention solves the problems of low segmentation accuracy and poor model generalization ability of traditional segmentation methods, thereby improving the accuracy and generalization ability of particle recognition. Attached Figure Description
[0041] Figure 1 This is a schematic diagram of point cloud data for rockfill material according to an embodiment of the present invention.
[0042] Figure 2 This is a point cloud projection depth map of the rockfill material according to an embodiment of the present invention.
[0043] Figure 3 This is a segmentation mask for the depth map of the rockfill in an embodiment of the present invention.
[0044] Figure 4 This is the point cloud segmentation result of the riprap material in an embodiment of the present invention.
[0045] Figure 5 This is the M1 gradation curve of the rockfill material in this embodiment of the invention.
[0046] Figure 6 This is the M2 gradation curve of the riprap in an embodiment of the present invention.
[0047] Figure 7 This is the M3 gradation curve of the riprap in an embodiment of the present invention.
[0048] Figure 8 This is the M4 gradation curve of the riprap in an embodiment of the present invention.
[0049] Figure 9 This is the M5 gradation curve of the riprap in an embodiment of the present invention.
[0050] Figure 10 This is the M6 gradation curve of the riprap in an embodiment of the present invention. Specific Implementation
[0051] The present invention will be further described below with reference to specific embodiments. These specific embodiments are further explanations of the principles of the present invention and are not intended to limit the present invention in any way. Any technology that is the same as or similar to the present invention does not exceed the scope of protection of the present invention.
[0052] like Figures 1-10 As shown, a method for detecting the gradation of riprap based on three-dimensional point clouds includes the following steps:
[0053] S1. Collect point cloud data of the rockfill.
[0054] Point cloud data was collected from the block stone using a handheld 3D laser scanner. A preview scan was performed before collection to familiarize the user with the appropriate scanning distance and movement speed. The point cloud quality was observed in real-time through the scanning software interface. If sparse areas were found, repeated scans were performed to improve accuracy. For areas with significant surface undulations, multiple coverage scans were conducted by adjusting the scanning angle and distance to ensure the completeness of the point cloud data for the stone pile. The collection results were saved upon completion. Figure 1 As shown.
[0055] S2. Project the point cloud of the riprap along the top-down direction to generate the corresponding depth image.
[0056] (1) Rotate the point cloud of the riprap to a standard horizontal coordinate system for depth projection. First, use the Random Sample Consensus (RANSAC) algorithm to perform plane fitting on the point cloud and extract the ground normal vector from it. For any point... P i Its distance to the fitting plane is: ;
[0057] When Di < ϵ (preset threshold), P i Points are considered interiors of the current plane. The RANSAC algorithm iteratively searches for the plane with the most interiors, ultimately determining the ground normal vector. Then, the point cloud is rotated to the horizontal coordinate system using the Rodrigues rotation formula, with the rotation matrix R2 as follows: ; ; ;
[0058] (2) Project the rotated point cloud along the Z-axis onto the XOY plane to construct a regular two-dimensional mesh. Set the mesh cell size to Δx×Δy, and traverse all points, determining their index positions in the mesh based on their X and Y coordinates: ;
[0059] Within each grid cell, the maximum height projection principle is used to retain the point with the largest Z value in that region as the depth value, thereby generating the corresponding depth matrix: ;
[0060] After normalization, the depth matrix D is converted into a grayscale image, generating the corresponding two-dimensional depth map, such as... Figure 2 As shown. The grayscale value of each pixel reflects the surface height information of that region, serving as input for subsequent instance segmentation models.
[0061] S3. Input the depth map into the instance segmentation model to obtain the segmentation mask for each stone particle.
[0062] The processed depth map is input into the instance segmentation model to obtain the segmentation mask for each stone particle, i.e., the corresponding contour information, thereby achieving target recognition and segmentation of the stone particles. Figure 3 As shown. In this embodiment, the instance segmentation model used is YOLOv8-seg, which introduces a more lightweight network structure and a stronger feature extraction backbone network, significantly reducing computational costs while maintaining model accuracy. Therefore, choosing YOLOv8-seg can achieve fast and stable instance segmentation of stone particles in complex backgrounds while ensuring high accuracy, meeting the dual requirements of real-time performance and accuracy in this invention.
[0063] S4. Map the segmentation mask back to the 3D point cloud and extract the 3D morphological features of each particle.
[0064] (1) After obtaining the depth map segmentation mask, inversely project the depth map mask back into the point cloud mask, such as... Figure 4 As shown. For any pixel position in the depth map and depth value The corresponding point cloud coordinate calculation formula is as follows: , ;
[0065] (2) Extract three-dimensional features from each segmented particle point cloud and calculate three-dimensional geometric features such as major axis, second major axis, minor axis, surface area, volume, sphericity, and roundness.
[0066] 1) Principal component analysis (PCA) is used to calculate the principal scale orientation of particles. The principal scale of rock particles is usually composed of three mutually perpendicular axial dimensions: the major axis (L), the second major axis (W), and the minor axis (T). PCA is used to calculate the point cloud covariance matrix and solve for its eigenvalues and eigenvectors. The point cloud data of the particles is then projected onto the direction with the largest variance, thereby aligning the projected coordinate system with the principal scale orientation of the particles. At this point, the difference in the range of the projected coordinates of the point cloud on the three axes corresponds to the major axis, second major axis, and minor axis of the particles, respectively.
[0067] 2) For the volume V and surface area S of the rock particle point cloud, the Delaunay triangulation algorithm is used to construct a surface mesh model of the particles for calculation. For a given triangular mesh, the position vectors of any vertex of a triangular facet are respectively... Taking the origin as the fourth vertex of the tetrahedron, the total volume of the entire mesh is... V The sum of the volumes of all tetrahedrons: ; ;
[0068] surface area SThis can be obtained by summing the areas of all the triangular faces: ; ;
[0069] 3) Sphericity is used to measure the similarity of the shape of rock particles to an ideal sphere. It is defined as the ratio of the surface area of a particle to the surface area of a sphere with the same volume. ;
[0070] In the formula, V and S It is distributed as the volume and surface area of rock particles.
[0071] 4) Roundness is used to measure the smoothness of the particle's edges. Wadell et al. defined three-dimensional roundness as the average of the ratios of the radii of curvature of each edge on the particle profile to the radius of the largest inscribed sphere. ;
[0072] In the formula: r i For the first i The radius of a sphere fitted by a set of edges; R The radius of the largest inscribed sphere of the particle; N This represents the number of sharp edges of the particles.
[0073] 5) Radius of the largest inscribed sphere of the particle R 1 and the smallest circumscribed ball of the particle R 2. Using the Euclidean distance field calculation method, an internal point set is generated by uniformly sampling inside the particle. P in For each internal sampling point, find its minimum and maximum Euclidean distances to the outer surface of the particle. The maximum value of the minimum distances corresponding to all sampling points is the radius of the maximum inscribed sphere. R 1. The minimum of the maximum distances corresponding to all sampling points is the radius of the minimum circumscribed sphere. R 2: ; ;
[0074] S5. Predict volume and particle size based on the three-dimensional features of particles, and plot the gradation curve.
[0075] (1) A volume prediction model based on the three-dimensional morphological features of particles is constructed. The input features include three-dimensional morphological parameters such as major axis, second major axis, minor axis, surface area, and volume. The particle volume is estimated using machine learning regression methods. This method can effectively compensate for the prediction of the complete volume of particles and improve the accuracy of volume estimation, thus addressing problems such as point cloud occlusion or incomplete scanning.
[0076] Compared to deep neural networks, traditional machine learning models typically exhibit higher training efficiency and stronger generalization ability in such tasks. Therefore, this embodiment uses LightGBM (Light Gradient Boosting Machine) to construct a volume prediction model. To further improve model performance, the Optuna hyperparameter optimization framework is introduced to automatically tune key parameters such as the learning rate, number of leaf nodes, and maximum depth, ensuring optimal volume prediction accuracy.
[0077] (2) After the volume prediction is completed, the secondary major axis of the particles is used as the particle size division standard. Particle size intervals are set according to the actual needs of the project, and the volume ratio of particles in each particle size interval is calculated. With particle size as the abscissa and volume distribution percentage as the ordinate, the discrete gradation curve of the rockfill can be plotted.
[0078] To verify the accuracy of this method, a sieve comparison test was conducted on sample dam materials selected at an actual construction site. Before the sieve operation, a handheld 3D laser scanner was used to scan the samples, collecting raw point cloud data. The method of this invention was then applied for rapid gradation detection. The detected gradation curve was compared with the sieve gradation curve. The gradation curves are shown below. Figures 5-10 As shown in the figure. Analysis revealed that the overall average absolute error between the particle mass fraction identification results for each particle size range and the actual gradation data was controlled within 5%. The gradation detection curve and the actual gradation curve showed a high degree of agreement, verifying the accuracy and applicability of this invention in practical engineering.
[0079] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.
[0080] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.
Claims
1. A method for detecting the gradation of riprap based on three-dimensional point clouds, characterized in that... Includes the following steps: S1. Collect point cloud data of the riprap material; S2. Project the point cloud of the riprap along the top-view direction to generate the corresponding depth image; S3. Input the depth map into the instance segmentation model to obtain the segmentation mask for each stone particle; S4. Map the segmentation mask back to the 3D point cloud and extract the 3D morphological features of each particle. S5. Predict volume and particle size based on the three-dimensional features of particles, and plot the gradation curve.
2. The method for detecting the gradation of riprap based on three-dimensional point clouds according to claim 1, characterized in that... S2 includes: S21. Rotate the point cloud of the riprap to a standard horizontal coordinate system for depth projection. First, the Random Sample Consensus (RANSAC) algorithm is used to perform plane fitting on the point cloud to extract the ground normal vector; for any point... P i Its distance to the fitting plane is: ; When Di < ϵ (preset threshold), P i Points are considered interiors of the current plane. The RANSAC algorithm iteratively searches for the plane with the most interiors, ultimately determining the ground normal vector. Then, the point cloud is rotated to the horizontal coordinate system using the Rodrigues rotation formula, with the rotation matrix R2 as follows: ; ; ; S22. Project the rotated point cloud along the Z-axis onto the XOY plane to construct a regular two-dimensional mesh; Set the grid cell size to Δx×Δy, and iterate through all points, determining their index position in the grid based on their X and Y coordinates: ; Within each grid cell, the maximum height projection principle is used to retain the point with the largest Z value in that region as the depth value, thereby generating the corresponding depth matrix: ; The depth matrix D is normalized and converted into a grayscale image to generate a corresponding two-dimensional depth map. The grayscale value of each pixel reflects the surface height information of the region and serves as the input for the subsequent instance segmentation model.
3. The method for detecting the gradation of riprap based on three-dimensional point clouds according to claim 1, characterized in that... The preferred instance segmentation model used in S3 is YOLOv8-seg.
4. The method for detecting the gradation of riprap based on three-dimensional point clouds according to claim 1, characterized in that... S4 includes: S41. After obtaining the depth map segmentation mask, inversely project the depth map mask back to the point cloud mask; for any pixel position in the depth map... and depth value The corresponding point cloud coordinate calculation formula is as follows: ; S42. Extract three-dimensional features from each segmented particle point cloud and calculate three-dimensional geometric features such as major axis, second major axis, minor axis, surface area, volume, sphericity, and roundness. 1) Principal component analysis was used to calculate the principal scale orientation of particles; The principal scale of rock particles is usually composed of three mutually perpendicular axial dimensions: the major axis L, the second major axis W, and the minor axis T. By using principal component analysis to calculate the covariance matrix of the point cloud and solve its eigenvalues and eigenvectors, the point cloud data of the particles is projected onto the direction with the largest variance, thereby aligning the projected coordinate system with the principal scale direction of the particles. At this time, the difference in the range of the projected coordinates of the point cloud on the three axes can correspond to the major axis, second major axis, and minor axis of the particles, respectively. 2) For the volume V and surface area S of the rock particle point cloud, the Delaunay triangulation algorithm is used to construct a surface mesh model of the particles for calculation; For a given triangular mesh, the position vectors of any vertex of a triangular face are respectively If the origin is taken as the fourth vertex of the tetrahedron, then the total volume of the entire mesh is... V The sum of the volumes of all tetrahedrons: ; ; surface area S This can be obtained by summing the areas of all the triangular faces: ; ; 3) Sphericity is used to measure the similarity of the shape of rock particles to an ideal sphere. It is defined as the ratio of the surface area of a particle to the surface area of a sphere with the same volume. ; In the formula, V and S Distribution is represented by the volume and surface area of rock particles; 4) Roundness is used to measure the smoothness of the edges and corners of particles. Wadell et al. defined three-dimensional roundness as the average of the ratios of the radii of curvature of each edge and corner on the particle profile to the radius of the largest inscribed sphere. ; In the formula: r i For the first i The radius of a sphere fitted by a set of edges; R The radius of the largest inscribed sphere of the particle; N The number of particle edges; 5) Radius of the largest inscribed sphere of the particle R 1 and the smallest circumscribed ball of the particle R 2. Using the Euclidean distance field calculation method, an internal point set is generated by uniformly sampling inside the particle. P in For each internal sampling point, find its minimum and maximum Euclidean distances to the outer surface of the particle. The maximum value of the minimum distances corresponding to all sampling points is the radius of the maximum inscribed sphere. R 1. The minimum of the maximum distances corresponding to all sampling points is the radius of the minimum circumscribed sphere. R 2: ; 。 5. The method for detecting the gradation of riprap based on three-dimensional point clouds according to claim 1, characterized in that... S5 includes: S51. Construct a volume prediction model based on the three-dimensional morphological features of particles. Use the three-dimensional geometric parameters of particles, such as major axis, second major axis, minor axis, surface area, and volume, as input features, and use machine learning regression methods to estimate the complete volume of particles. S52. After completing the particle volume prediction, the secondary major axis of the particles is used as the basis for particle size division. Particle size intervals are set according to the actual needs of the project, and the proportion of particle volume in each particle size interval is calculated. The discrete gradation curve of the rockfill is plotted with particle size as the abscissa and volume distribution percentage as the ordinate.
6. The method for detecting the gradation of riprap based on three-dimensional point clouds according to claim 5, characterized in that... The volume prediction model in S51 is a regression model of LightGBM.
