CN-122023961-A - Block stone point cloud data set manufacturing method based on discrete element stacking simulation

CN122023961ACN 122023961 ACN122023961 ACN 122023961ACN-122023961-A

Abstract

The invention discloses a method for manufacturing a block stone point cloud data set based on discrete element accumulation simulation, which relates to the fields of numerical modeling and computer vision, and comprises the steps of obtaining the geometric form of real block stone particles through a three-dimensional laser scanning technology and establishing a polymorphic particle template library; performing free accumulation simulation of the block stones based on a discrete element method to generate a three-dimensional accumulation model conforming to target grading, performing particle instance segmentation on the accumulation model through topology diagram construction and connected component analysis, endowing each particle with a unique instance label, performing uniform point cloud sampling on the surface of the three-dimensional model, removing shielding points by combining multi-view perspective projection and depth buffering principles, and generating a three-dimensional block stone point cloud data set with reasonable shielding and accurate labeling. The method has the advantages of high automation degree, true shielding treatment, accurate data labeling and the like, and can be applied to training and evaluation of the block point cloud instance segmentation model.

Inventors

Liao Yabin
YUAN KUN
DU TAO
HUANG GUOLIANG
YAO QIANG
WU YILONG
LI CHUNQUAN
LIU JINGBIAO
LI HONGTAO
JI PENG
SHI MENGNAN
ZENG JUN

Assignees

中国水利水电第七工程局有限公司
四川大学

Dates

Publication Date: 20260512
Application Date: 20260108

Claims (6)

1. A method for manufacturing a block stone point cloud data set based on discrete element accumulation simulation is characterized by comprising the following steps: s1, building a block stone particle three-dimensional model according to the shape of a block stone particle in actual engineering, and carrying out normalization processing on the particle three-dimensional model to build a standard particle form template library; S2, performing discrete element stacking simulation by using a real particle template to generate a three-dimensional stacking model; S3, constructing a grid topological graph according to the block stone pile three-dimensional model, automatically dividing each particle model based on a connected component algorithm, and endowing each particle with a unique instance label; S4, carrying out uniform sampling on the whole grids of the marked particle model to generate a complete point cloud, removing invisible points based on multi-view perspective projection and depth buffering principles, and combining visible point cloud data under a plurality of view angles to obtain a block stone point cloud data set with real structure, reasonable shielding and clear labels.
2. The method for manufacturing the block stone point cloud data set based on discrete element accumulation simulation according to claim 1 is characterized in that in the step S1, the establishment of a block stone particle three-dimensional model comprises the steps of collecting particle surface point cloud data by using a three-dimensional laser scanner, guiding the scanning data into modeling software for gridding treatment, including denoising, hole repairing and simplifying operation, to generate a closed three-dimensional curved surface model, and carrying out normalization treatment on the model to unify equivalent particle sizes to 1mm to construct a polymorphic particle template library, wherein particle forms include but are not limited to blocks, sheets and strips.
3. The method for producing a block stone point cloud data set based on discrete element stacking simulation according to claim 1, wherein the step S2 includes: S21, building a stone particle discrete element three-dimensional model based on a template library, setting physical parameters such as particle density, elastic modulus, friction coefficient and the like, and building a closed space bottom surface as a stacking area; S22, setting a particle size distribution range according to a target grading curve, generating particles with different particle sizes and shapes according to mass proportions by a plurality of particle factories, simulating the particles to fall and stack under the action of gravity to form a block stone pile, and calculating interaction force among the particles by using Hertz-Mindlin (no slip) and Standard Rolling Friction models in the simulation process; and S23, after the stacking simulation is completed, the three-dimensional model of the block stone pile is exported in an STL format, and meanwhile, a mass-particle size distribution curve of generated particles is output and stored in a TXT format.
4. The method for producing a block stone point cloud data set based on discrete element stacking simulation according to claim 1, wherein the step S3 comprises: S31, analyzing the derived STL file, extracting vertex coordinates and connection relations of triangular patches, and constructing a grid topological graph, wherein three vertexes of each patch are used as nodes in a graph structure, and the relation of sharing edges between the patches is used as undirected edges of the graph, so that a complete grid topological graph is formed; S32, carrying out connected component analysis on the topological graph by adopting a Breadth First Search (BFS) algorithm, identifying all closed particle examples, wherein each connected component corresponds to one particle and is assigned with a unique example label, and the algorithm comprises the following specific steps: 1) Randomly selecting an unvisited vertex as a starting point; 2) Accessing all vertices connected to the point; 3) Continuing recursively expanding the access range until the access range cannot be continued; 4) Grouping all visited vertices into a group to form a connected component; 5) The above process is repeated until all vertices are accessed.
5. The method for producing a block stone point cloud data set based on discrete element stacking simulation according to claim 1, wherein the step S4 comprises: S41, uniformly sampling the whole grid of the marked particle three-dimensional model by adopting an area weighted gravity center sampling method, distributing sampling points according to the area of triangular patches, generating an equal-density point cloud by utilizing gravity coordinates on each patch, setting the total sampling point as N, traversing each triangular patch, respectively calculating the area as A 1 ,A 2 , ···,A i , and setting the total area of all triangular patches as A total , wherein the sampling point N i of each triangular patch is as follows: ; Uniformly distributed sampling points are generated on the triangular surface patch by adopting a barycentric coordinate method, barycentric coordinate coefficients meeting the uniform distribution are randomly generated, and the coordinates of the sampling points are as follows: ; wherein: ; S42, setting a plurality of camera view angles, performing perspective projection on the point cloud data, reserving the point with the minimum depth at each pixel position by using a depth buffer mechanism, removing the invisible point which is shielded, merging the visible point clouds under the plurality of view angles, and removing the repeated points to form a three-dimensional block stone point cloud data set.
6. The method for producing a block stone point cloud data set based on discrete element stacking simulation according to claim 5, wherein the step S42 specifically includes: s421, for a group of marked stone point clouds, setting a matrix M which consists of N points and can be expressed as N rows and 3 columns and a matrix N of N rows and 1 column, wherein each row of the matrix M represents the three-dimensional coordinate of a point, and each row of the matrix N represents the particle instance label to which the point belongs; S422, setting the position and the direction of a camera, converting the coordinates of the point cloud from a world coordinate system to a camera coordinate system by using an external parameter matrix, wherein for any point in the point cloud, the coordinates under the camera coordinate system can be calculated by the following formula: ; Wherein R is a rotation matrix of 3 multiplied by 3 and represents the direction of the camera in the original coordinate system, T is a translation vector of 3 multiplied by 1 and represents the position of the camera in the original coordinate system; , And Three-dimensional coordinates of the point in a camera coordinate system; Projecting points from three-dimensional coordinates to a two-dimensional planar coordinate system according to perspective projection: ; where x and y represent the coordinates of points on the projection plane in mm, it is necessary to further convert them to the pixel coordinate system: ; The combination is represented in matrix form: ; Wherein u and v are coordinates of points on a projection pixel plane, f x and f y are focal lengths of cameras, and c x and c y are offset of principal points; s423, traversing the point coordinates of each row in the matrix M, and calculating point by point, if two points are projected to the same pixel position, reserving the point with the smallest depth, marking the rest invisible points and deleting the invisible points; S424, repeating the processing flow, obtaining visible point cloud data under a plurality of view angles, merging and de-duplicating the visible point cloud data, and finally forming a block stone point cloud data set with real scanning characteristics and clear instance labels.

Description

Block stone point cloud data set manufacturing method based on discrete element stacking simulation Technical Field The invention belongs to the field of numerical modeling and the field of computer vision, and particularly relates to a method for manufacturing a block stone point cloud data set based on discrete element stacking simulation. Background In modern hydraulic engineering construction, rock-fill is used as a main filling material of a rock-fill dam, and the rationality of grain composition of the rock-fill dam directly influences the compactness, impermeability and overall stability of the dam body. Reasonable grain composition can effectively reduce the risks of piping, seepage, sedimentation and the like, and has important significance for guaranteeing the construction quality and long-term operation safety of the hydraulic junction engineering. The traditional particle size distribution detection method mainly relies on manual screening, namely, representative samples are extracted on site, particles are subjected to particle size separation through a classification screening device, and the mass percentage of each particle size interval is counted. However, the method is complicated in operation process, time-consuming and labor-consuming, low in detection efficiency, and the detection result depends on experience and technical level of operators to a great extent, so that real-time and efficient particle grading monitoring is difficult to realize. In the current research of the novel grading detection method, the new technologies such as photography, camera shooting, laser scanning and the like are mostly adopted to identify and graded the stone particles. The particle grading analysis method based on the two-dimensional image is characterized in that the image of the particle surface is acquired through the image pickup equipment, and particle boundary identification is carried out by combining edge detection, a watershed algorithm or a deep learning model, so that the particle size distribution is estimated. The method has strong practicability and certain degree of automation, but is based on a two-dimensional projection image, is easy to be interfered by factors such as illumination change, shielding, image resolution and the like, can not comprehensively reflect the three-dimensional shape and volume characteristics of particles, and has systematic deviation and limited application range of calculation results. In contrast, the three-dimensional laser-based grading detection method can finish the space measurement of large-area irregularly-piled stones in a short time, and generates high-density point cloud data containing three-dimensional geometric information of particles. On the basis, the three-dimensional outline of the particle is extracted through a point cloud segmentation algorithm, and the volume and grading parameters of the particle are further calculated, so that the limitation of an image recognition method can be effectively overcome. The current particle segmentation method based on the three-dimensional point cloud mainly comprises a traditional segmentation algorithm and a segmentation algorithm based on deep learning. The conventional point cloud segmentation method generally depends on geometric attribute features, such as normal vectors, curvatures, euclidean distances and the like, and particle segmentation is realized through modes of edge detection, region growth, model fitting, clustering and the like. The method has certain advantages when processing the target object with regular geometry and clear boundary, but the segmentation precision is obviously reduced when facing the stone pile body with dense accumulation, complex particle morphology and obvious scale difference, and the grading calculation precision is required to be improved due to highly dependent parameter debugging and experience judgment. With the development of the deep learning technology, the point cloud segmentation method based on the deep learning can realize more accurate point cloud segmentation in a complex scene. However, training of point cloud segmented neural network models often requires a large number of data set point clouds. In reality, because particles are densely stacked, a large amount of shielding and adhesion exist among the particles, particle boundaries in point cloud data are fuzzy, the time and the labor are consumed for manually marking the particle examples point by point, the consistency is poor, the acquisition efficiency and the quality of training samples are seriously restricted, and the method becomes a big bottleneck for applying the deep learning method to block stone segmentation. Therefore, there is a need to propose a data set generation method for neural network training for block point cloud segmentation to solve the above-mentioned problems. Disclosure of Invention Aiming at the defects in the prior art, the invention aims to provide a method for manu