Method and device for automatically selecting shoveling points of loading machine for large material piles

Abstract

The invention discloses a large material pile-oriented loader shoveling point autonomous selection method and device, and belongs to the technical field of engineering machinery, and the method comprises the steps: obtaining three-dimensional point cloud data through a laser radar in response to a shoveling point selection request instruction; preprocessing the three-dimensional point cloud data to obtain preprocessed point cloud data; segmenting the preprocessed point cloud data through a segmentation network model to obtain a plurality of segmented point cloud data; based on the multiple pieces of segmented point cloud data, through a fullness rate prediction agent model, obtaining fullness rate data of the multiple candidate points; and on the basis of the fullness rate data of the multiple candidate points, an optimal spading point is obtained through a marine predator optimization algorithm, a control instruction set is generated in response to the optimal spading point, and the control instruction set is used for controlling the target loader to execute spading operation corresponding to the optimal spading point. The invention aims to overcome the defects of low efficiency, poor stability and the like of manual selection of a shoveling point in a large material pile scene.

Description

Technical Field

[0001] This invention discloses a method and device for loader digging points to be autonomously selected for large material piles, belonging to the field of construction machinery technology. Background Technology

[0002] As a core piece of engineering machinery for bulk material handling in mining, ports, and large warehouses, the selection of digging points directly determines the operating efficiency, bucket full rate, and fuel consumption of loaders. Currently, this operation mainly relies on the driver's visual perception and work experience. However, in large material pile operation scenarios, the material pile has a complex shape and large scale, and the driver's field of vision cannot fully cover the material pile. Manual selection is difficult to guarantee the efficiency and stability of the operation, and is prone to problems such as fluctuations in bucket full rate, repeated digging, and increased energy consumption. Although there are existing methods for material pile shape recognition and digging point generation based on camera and LiDAR fusion perception, these methods are limited by the camera's field of vision and are difficult to obtain the complete surface features of large material piles. Moreover, some methods only rely on point cloud geometric features to select digging points, lacking prediction and evaluation of digging efficiency, and cannot meet the digging operation requirements in large material pile scenarios. Summary of the Invention

[0003] To address the shortcomings of low efficiency and poor stability in manually selecting digging points in large material pile scenarios, as well as the deficiencies of existing fusion perception methods in obtaining complete features of the material pile and lacking a mechanism for predicting and evaluating digging efficiency, this invention proposes a method and device for the autonomous selection of digging points by loaders in large material pile environments. Based on acquiring complete three-dimensional point cloud data of the material pile, the method achieves autonomous evaluation and intelligent selection of digging points through point cloud processing, segmentation, efficiency prediction, and algorithm optimization, thereby improving the operating performance of loaders in large material pile environments.

[0004] The technical solution of the present invention is as follows:

[0005] According to a first aspect of the present invention, a method for autonomously selecting a loader's digging point for a large stockpile is provided, comprising: in response to a digging point selection request command, acquiring three-dimensional point cloud data via lidar; preprocessing the three-dimensional point cloud data to obtain preprocessed point cloud data; segmenting the preprocessed point cloud data using a segmentation network model to obtain multiple segmented point cloud data; based on the multiple segmented point cloud data, obtaining full bucket rate data for multiple candidate points using a full bucket rate prediction proxy model; based on the full bucket rate data of the multiple candidate points, obtaining an optimal digging point using a marine predator optimization algorithm; and generating a control instruction set in response to the optimal digging point, the control instruction set being used to control the target loader to perform the digging operation corresponding to the optimal digging point.

[0006] Furthermore, the 3D point cloud data is preprocessed to obtain preprocessed point cloud data, including: using a tightly coupled iterative Kalman filter algorithm to perform multi-frame point cloud superposition processing on the 3D point cloud data to obtain superimposed point cloud data; using statistical filtering and voxel filtering to process the superimposed point cloud data to obtain filtered point cloud data; and dividing the filtered point cloud data into material pile and background point clouds based on spatial distribution and surface morphology features to obtain preprocessed point cloud data.

[0007] Furthermore, based on multiple segmented point cloud data, a full bucket rate prediction surrogate model is used to obtain full bucket rate data for multiple candidate points. This includes: dividing the preprocessed point cloud data into two dimensions on the horizontal projection plane with a preset grid size; extracting the material pile boundary points using the Alpha-shape algorithm to obtain an initial boundary point set; performing B-spline curve interpolation on the initial boundary point set to obtain a boundary curve; uniformly sampling the boundary curve to obtain a complete boundary point set; sampling the complete boundary point set to obtain initial sampling points; extracting the neighborhood point cloud corresponding to each initial sampling point; obtaining quadratic surface parameters for multiple sampling points through quadratic surface fitting; and obtaining full bucket rate data for multiple candidate points based on the quadratic surface parameters of multiple sampling points using the full bucket rate prediction surrogate model.

[0008] Further, the preprocessed point cloud data is divided into two dimensions on the horizontal projection plane with a preset grid size. The Alpha-Shape algorithm is used to extract the boundary points of the material pile to obtain an initial boundary point set. This includes: dividing the preprocessed point cloud data into two dimensions on the horizontal projection plane with a preset grid size to obtain multiple grid-divided point cloud data; judging the multiple grid-divided point cloud data based on a set height threshold to obtain a boundary grid index set; projecting the material pile point cloud onto the x–y plane, and using the Alpha-Shape algorithm to calculate the boundary contour of the boundary grid index set to obtain the initial boundary point set.

[0009] Furthermore, based on the full-bucket rate data of multiple candidate points, the optimal digging point is obtained through the marine predator optimization algorithm, including: obtaining multiple expected values ​​of full-bucket rate data based on the full-bucket rate data of multiple candidate points; obtaining the comprehensive fitness of multiple digging points through a comprehensive fitness function based on the multiple expected values ​​of full-bucket rate data, the path cost term from the digging point to the fixed unloading point, and the distance penalty term between the digging point and the previous digging point; obtaining the fitness values ​​of multiple candidate points through the marine predator optimization algorithm based on the comprehensive fitness of multiple digging points; and searching the fitness values ​​of multiple candidate points to obtain the optimal digging point.

[0010] According to a second aspect of the present invention, a loader autonomously selects digging points for large stockpiles, comprising: an acquisition module, configured to acquire three-dimensional point cloud data via lidar in response to a digging point selection request command; a processing module, configured to preprocess the three-dimensional point cloud data to obtain preprocessed point cloud data; a segmentation module, configured to segment the preprocessed point cloud data using a segmentation network model to obtain multiple segmented point cloud data; a prediction module, configured to obtain full bucket rate data of multiple candidate points based on the multiple segmented point cloud data using a full bucket rate prediction proxy model; and an optimization module, configured to obtain an optimal digging point based on the full bucket rate data of the multiple candidate points using a marine predator optimization algorithm, and generate a control instruction set in response to the optimal digging point, the control instruction set being used to control the target loader to perform digging operations corresponding to the optimal digging point.

[0011] According to a third aspect of the present invention, a terminal is provided, comprising:

[0012] One or more processors;

[0013] Memory for storing the one or more processor-executable instructions;

[0014] Wherein, the one or more processors are configured as follows:

[0015] Perform the method described in the first aspect of the embodiments of the present invention.

[0016] According to a fourth aspect of the present invention, a non-transitory computer-readable storage medium is provided, wherein when instructions in the storage medium are executed by a processor of a terminal, the terminal is enabled to perform the method described in the first aspect of the present invention.

[0017] According to a fifth aspect of the present invention, an application product is provided that, when the application product is running on a terminal, causes the terminal to execute the method described in the first aspect of the present invention.

[0018] The beneficial effects of this invention are as follows:

[0019] This invention provides a method and apparatus for autonomously selecting shovel points for loaders operating on large stockpiles. It acquires 3D point cloud data of the large stockpile using lidar, preprocesses it using a segmentation network model to obtain block point cloud data, then outputs the full-bucket rate data of candidate points based on a full-bucket rate prediction proxy model. Finally, it uses a marine predator optimization algorithm to select the optimal shovel point and generate a control command set. This method overcomes the limitations of manual selection and the shortcomings of existing methods in terms of field of view and evaluation mechanism, enabling autonomous and intelligent selection of shovel points for large stockpiles. This effectively improves the full-bucket rate and efficiency of loader operations, reduces repetitive shoveling, lowers fuel consumption, and ensures the efficiency and stability of loading operations in large stockpile scenarios.

[0020] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an exemplary embodiment of a method for loader shovel point autonomous selection for large stockpiles.

[0022] Figure 2 This is a schematic diagram illustrating the extraction of initial boundary points of a loader's autonomous selection of digging points for large stockpiles, according to an exemplary embodiment.

[0023] Figure 3 This is a schematic diagram of B-spline curve interpolation of the stockpile boundary points, illustrating an autonomous selection method for loader digging points in large stockpiles according to an exemplary embodiment.

[0024] Figure 4 This is a schematic diagram illustrating the determination of the digging direction and the fitting of a quadratic surface to the neighborhood point cloud of a loader for autonomous selection of digging points in a large stockpile, according to an exemplary embodiment.

[0025] Figure 5 This is a schematic block diagram illustrating the structure of a loader digging point autonomous selection device for large stockpiles, according to an exemplary embodiment.

[0026] Figure 6 This is a schematic block diagram of a terminal structure according to an exemplary embodiment. Detailed Implementation

[0027] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] In the description of this invention, it should be noted that the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0029] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0030] This invention provides a method for loader digging points to be autonomously selected for large stockpiles. The method is implemented by a terminal, which can be a desktop computer or a laptop computer, etc., and the terminal includes at least a CPU.

[0031] Example 1

[0032] Figure 1 This is a flowchart illustrating an exemplary method for autonomously selecting the digging point of a loader for large stockpiles. The method is used by a terminal and includes the following steps:

[0033] Step S10: In response to the excavation point selection request command, acquire three-dimensional point cloud data through lidar;

[0034] In step S10, the loader hardware mainly includes: lidar, industrial control computer, inertial navigation measurement unit, and vehicle controller. The lidar on the loader continuously scans the material pile at the work site from multiple angles and in multiple orientations, collecting comprehensive three-dimensional point cloud data, and then sends the three-dimensional point cloud data to the terminal, i.e., to the industrial control computer.

[0035] Step S20: Preprocess the 3D point cloud data to obtain preprocessed point cloud data;

[0036] In step S20, the FAST-LIO2 algorithm based on tightly coupled iterative Kalman filtering is used to perform multi-frame point cloud superposition processing on the three-dimensional point cloud data to obtain superimposed point cloud data, which can significantly improve the point cloud density and geometric detail restoration of the material pile surface, making the overall shape of the material pile more complete and the boundary clearer, thereby providing a high-precision data foundation for subsequent feature extraction and model training.

[0037] Subsequently, preprocessing operations are performed on the fused point cloud data, including statistical filtering and voxel filtering to remove outliers, downsample, and remove noise points from the superimposed point cloud data, resulting in filtered point cloud data. This reduces data redundancy and improves processing efficiency while maintaining the main geometric structure features.

[0038] The filtered point cloud data is finely labeled, and divided into two categories, "material piles" and "background," based on spatial distribution and surface morphology features. The constructed dataset covers different material pile sizes, shapes, and background environments, exhibiting good diversity and representativeness, providing reliable data support for subsequent model training and performance verification, thus serving as the preprocessed point cloud data.

[0039] Step S30: The preprocessed point cloud data is segmented using a segmentation network model to obtain multiple segmented point cloud data.

[0040] In step S30, before introducing how to segment the preprocessed point cloud data using a segmentation network model, the following describes in detail how to train the segmentation network model: the preprocessed point cloud dataset is randomly divided into a training set, a validation set, and a test set in a ratio of 70%:15%:15% to ensure the sufficiency and generalization ability of the model training.

[0041] Building upon this foundation, a PointNet++ network structure is constructed. This network achieves efficient modeling of the spatial structure and geometric features of the point cloud of the material pile through hierarchical feature extraction and local feature aggregation mechanisms. The network input consists of preprocessed point cloud coordinates and their geometric attributes, such as normal vectors, and the output is a point-by-point semantic category label (material pile / background).

[0042] Supervised learning was employed during model training. The training set served as the primary data source, while the validation set was used to dynamically monitor model convergence and prevent overfitting. The Adam optimizer was selected for optimization, with the cross-entropy loss function as the main objective, used to measure the difference between the predicted class and the true label. During training, the accuracy, intersection-over-union ratio (IoU), and loss value trends were monitored simultaneously to evaluate model performance.

[0043] Once training reaches stable convergence, the model's performance is evaluated using a validation set across different training epochs. The optimal model weights are selected based on the dual criteria of minimum validation loss and highest segmentation accuracy. Finally, the selected model is used to evaluate its performance and perform generalization validation on a test set, ensuring that its segmentation accuracy and robustness on unseen data meet the expected requirements. This validated model is then output as the final segmentation model, resulting in the segmentation network model.

[0044] Subsequently, the preprocessed point cloud data is segmented using a segmentation network model to obtain multiple segmented point cloud data.

[0045] Step S40: Based on multiple segmented point cloud data, the full-bucket rate data of multiple candidate points is obtained through the full-bucket rate prediction proxy model;

[0046] In step S40, firstly, the preprocessed point cloud data is divided into two dimensions on the horizontal projection plane using a preset grid size. Then, the Alpha-shape algorithm is used to extract the material pile boundary points, resulting in an initial boundary point set, such as... Figure 2 As shown, the specific content is as follows:

[0047] (1) Grid division and height statistics.

[0048] On the horizontal projection plane (x–y plane) of the point cloud of the material pile, a two-dimensional division is performed with a preset grid size r.

[0049] Let the point cloud set be P = {(x k ,y k ,z k If |k=1,2,…,N}, then the point (x k ,y k The raster index (i,j) corresponding to ) is calculated as follows:

[0050]

[0051] Among them, (x min ,y min ) represents the minimum coordinates of the point cloud projection.

[0052] The height value of each grid cell is defined as the highest height value of the point cloud within that grid cell:

[0053]

[0054] Where G(i,j) represents the set of points belonging to grid (i,j).

[0055] For empty graticules (those without point distribution), their height is interpolated and padded using the average height of the effective graticules in the neighborhood:

[0056]

[0057] in Represents the set of neighboring grid cells. This represents the number of valid grid cells within the neighborhood.

[0058] (2) Height quantile threshold segmentation.

[0059] Sort all valid values ​​of H(i,j) in ascending order and select P. q The height value at that location is used as the threshold T H :

[0060] T H =Quantile(H,P) q ), (5)

[0061] Where Pq P represents the quantile percentage parameter. q ∈(0,1), for example, when P q When T = 0.10, H The corresponding height value is ranked in the 10th percentile among all grid cells.

[0062] Grids below a threshold are considered to be at the boundary of the material pile, while grids above the threshold are considered to be in the main area of ​​the material pile. The occupancy marker matrix is ​​defined accordingly:

[0063]

[0064] Where O(i,j) = 1 indicates that the grid belongs to the boundary region of the material pile, and O(i,j) = 0 indicates that the grid belongs to the main region of the material pile.

[0065] Therefore, the set of boundary raster indices can be obtained:

[0066] B={(i,j)|O(i,j)=1} (7)

[0067] (3) Alpha-Shape boundary extraction.

[0068] Project the point cloud of the material pile onto the x–y plane, and use the Alpha-Shape algorithm to calculate the boundary contour S of the point set. α Its core constraint is: If there exists a circle with radius α that passes through x p ,x q And it contains no other points, then (x p ,x q )∈S α .

[0069] (4) Initial boundary point set generation.

[0070] Based on the boundary raster index set B determined in step (2), extract the elements located within these rasters that satisfy (x k ,y k )∈S α The point cloud data. The final initial boundary point set is obtained:

[0071] P init ={(x k ,y k ,z k )∈P|(i k j k )∈B,(x k ,y k ))∈S α} (8)

[0072] This point set delineates the edges of the stockpile's shape, providing high-precision input for subsequent boundary refinement and excavation point planning.

[0073] Secondly, B-spline curve interpolation is performed on the initial boundary point set to obtain the boundary curve, such as... Figure 3 As shown, the specific content is as follows:

[0074] To obtain a continuous and smooth representation of the boundary point set and ensure the uniform distribution of points in space, the initial boundary point set is first...

[0075] P init ={Q0,Q1,…,Q m} (9)

[0076] Perform cubic B-spline curve interpolation. Assume the curve order is k = 4. For each data point, a parameter value u needs to be assigned first. i (i.e., the position on the curve), so as to map the discrete points to the parameter interval [0,1]. To better reflect the actual spatial distribution between points, the chord length method can be used for parameterization:

[0077] u0=0 (10)

[0078]

[0079] This method ensures that the curve parameter u i The distance between points is proportional to the actual distance between them, i = 1, 2, ..., m, thereby reducing excessive bending or stretching of the curve in certain areas.

[0080] Subsequently, a quasi-uniform node vector U = {u0, u1, ..., u} is constructed. n+k}:

[0081]

[0082] Where the internal node t j Calculated using the average parameter method:

[0083]

[0084] This node vector construction method ensures that the two ends of the curve pass precisely through the boundary points, j = k, k+1, ..., n, while the internal node distribution is reasonable, making the curve smooth and continuous as a whole.

[0085] Based on the node vectors and parameter values, define the cubic B-spline basis function N. j，k (u)(Cox–de Boor recurrence formula):

[0086]

[0087] Then establish a system of interpolation equations:

[0088]

[0089] Where P j For control points, k > 1, i = 0, 1, ..., m, the complete control point set P = {P0, P1, ..., Pm} is obtained by solving this linear system. n Therefore, a continuous B-spline curve can be constructed to represent it:

[0090]

[0091] Where, u∈[u0, u m The curve [is defined] for each parameter u. i It precisely passes through the initial boundary point and maintains a smooth transition between control points.

[0092] Next, uniform sampling is performed on the boundary curve to obtain a complete set of boundary points, as detailed below:

[0093] To further obtain a complete set of boundary points that are uniformly distributed in space, the curve C(u) is uniformly sampled over the parameter interval:

[0094] u sample ={u0,u0+Δu,u0+2Δu,...,u m} I (14)

[0095]

[0096] Where N represents the number of sampling points, calculate the corresponding curve coordinates:

[0097] P boundary ={C(u s )|u s ∈u sample} (16)

[0098] Through uniform sampling, the final P is obtained boundary It preserves the geometric features of the original boundary and is evenly distributed in space, providing reliable input data for subsequent processing.

[0099] Then, the complete set of boundary points is sampled to obtain the initial sampling points, as detailed below:

[0100] Determine the set of boundary points P of the stockpile. boundary Then, the initial digging point s was selected using the direct random indexing method. j This method first stores the boundary point set in a randomly accessible linear structure, denoted by a capacity of N. Then, using a uniformly distributed random number generator, it extracts an index i within the integer interval [0, N-1], and directly accesses P. boundary[i] and define it as the initial sampling point s j .

[0101] Next, the neighborhood point cloud corresponding to each initial sampling point is extracted, and quadratic surface fitting is used to obtain the quadratic surface parameters of multiple sampling points, such as... Figure 4 As shown, the specific content is as follows:

[0102] For the excavation point s j , based on its coordinates on the horizontal plane Let P be the center of a circle, and let r = W / 2 be the radius of the neighborhood. boundary Extracting the neighborhood point set:

[0103]

[0104] Where W is the width of the loader bucket.

[0105] Next, for the neighborhood point set B j Coordinates on the plane {(x i y i Linear fitting is performed. Compared with the traditional least squares method, principal component analysis (PCA) can obtain more stable and noise-resistant fitting results.

[0106] First, calculate the centroid of the neighboring points:

[0107]

[0108] Then construct a decentralized data matrix:

[0109]

[0110] Calculate the covariance matrix:

[0111]

[0112] Find the largest eigenvector of S:

[0113] v j =(v x ,v y ) (twenty two)

[0114] As the direction vector of the fitted line, the line passes through the point

[0115] This direction corresponds to two possible normal vectors:

[0116]

[0117] To determine the correct orientation of the excavation, first calculate the coordinates of the center of mass of the material pile on the horizontal plane:

[0118]

[0119] Where N is the total number of points in the point cloud on the surface of the stockpile.

[0120] Define a vector pointing from the candidate point to the centroid of the stockpile:

[0121]

[0122] Subsequently, by judging the relationship between the normal vector and w j The dot product sign determines the digging direction. To ensure the digging action is directed towards the center of the stockpile, the normal vector with a positive dot product is chosen as the final digging direction vector:

[0123]

[0124] Finally, the digging direction vector d j Defined as a unit normal vector pointing towards the center of mass of the stockpile, this ensures that the digging action conforms to physical reality, enabling the loader to dig efficiently. This vector represents the direction of digging on the horizontal plane, but is not yet expressed as an angle. To facilitate subsequent neighborhood point cloud extraction and loader attitude calculation, the direction vector needs to be converted into a digging angle:

[0125] θ j =arctan2(d y d x (27)

[0126] This angle reflects the degree of deflection of the digging direction relative to the global positive x-axis: when θ j When θ = 0, the digging direction is consistent with the x-axis; when θ = 0, the digging direction is consistent with the x-axis. j When θ > 0, the direction deflects counterclockwise; when θ < 0, the direction deflects counterclockwise. j When the value is greater than 0, the direction deflects clockwise.

[0127] To facilitate subsequent extraction of the neighboring point cloud of the excavation point and align the excavation direction with the x-axis, the entire stockpile point cloud P is processed. pile Apply a rotational transformation about the z-axis:

[0128]

[0129] P rot =R z (-θ j )P pile (29)

[0130] After rotation, the digging direction coincides with the positive x-axis direction. At this time, observing the material pile from the positive x-axis direction will provide a material pile perspective that is consistent with the digging direction.

[0131] The point cloud P after rotation rotIn the middle, the y-coordinate of the rotated candidate digging point is used. Centered on the bucket, extract the neighborhood point cloud with a width equal to the bucket width W:

[0132]

[0133] In the cropped point cloud, search for the minimum coordinate value along the x-axis:

[0134]

[0135] The corresponding point is the point on the material pile that the loader's blade first contacts when digging at that digging point. Accordingly, the candidate digging point is updated to the actual digging point in the rotated coordinate system:

[0136]

[0137] Subsequently, in the interval x∈[x min ,x min Extract the point cloud of the neighborhood of the excavation point within the +L] section:

[0138] P shovel_area ={(x, y, z)∈P slice |x min ≤x≤x min +L} (33)

[0139] Where L is the bucket length. Finally, rotate the updated digging point back to the original coordinate system:

[0140]

[0141] By following the steps above, the digging point s can be obtained. j The corresponding standardized neighborhood point cloud and actual digging points, where the digging direction has been transformed into a clear angular form.

[0142] To further characterize the geometric features of the material pile surface at the digging point, the point cloud P in the neighborhood of the digging point is analyzed. shovel_area Perform quadratic surface fitting. Assume there are N sample points in the point cloud, with coordinates (x, y, y) respectively. i y i , z i ), i = 1, 2, ..., N. The local surface can be approximated by the general quadratic surface equation:

[0143] z(x, y) = ax 2 +bxy+cy 2 +dx+ey+f (35)

[0144] Where a, b, c, d, e, and f are the surface parameters to be determined.

[0145] To obtain the above surface parameters, the quadratic surface equation is discretized and written in matrix form:

[0146] z=Xα (36)

[0147] in:

[0148]

[0149]

[0150] By estimating the surface parameters using the least squares method, we can obtain:

[0151]

[0152] The optimal parameter solution for the local surface can reflect the true geometric shape of the material pile surface in the neighborhood of the digging point.

[0153] Finally, based on the quadratic surface parameters of multiple sampling points, the full-bowl rate prediction proxy model is used to obtain the full-bowl rate data of multiple candidate points, as detailed below:

[0154] Before calculating using the full-bucket-rate prediction agent model, we will first introduce the training of the full-bucket-rate prediction agent model:

[0155] Perform actual digging experiments and record the bucket fullness F. j The full-capacity rate is defined as:

[0156]

[0157] Among them, V load,j For V represents the volume of material in the bucket after excavation at the excavation point. bucket This refers to the rated volume of the bucket.

[0158] The parameters [a, b, c, d, e, f] obtained from the quadratic surface fitting are used as input features, with the measured full-bucket ratio F as the input feature. j As output, construct the initial training samples:

[0159] Sample j =([a,b,c,d,e,f],F j (41)

[0160] After excavation is completed, the point cloud of the material pile is re-segmented, and boundary point extraction and interpolation are performed. Excavation points are randomly selected, and the quadratic surface parameters and the bucket fullness rate are calculated. This process is repeated M times, and the results are summarized to form the initial dataset.

[0161]

[0162] Where, x j This represents the corresponding input feature vector. This dataset provides fundamental data support for the subsequent development of a full-bucket rate prediction model.

[0163] The training process for the full-capacity prediction model is as follows:

[0164] Based on the obtained initial sample data in F represents the parameter vector for fitting the quadratic surface of the stockpile. i ∈[0,1] represents the corresponding full-bowl rate, and a full-bowl rate prediction agent model f(x) is constructed using the Bayesian optimization method.

[0165] Gaussian process regression is chosen as the core of the surrogate model. Assume the fullness rate y satisfies:

[0166]

[0167] Where ∈ represents Gaussian noise. The surrogate model establishes a mapping relationship between the input parameter x and the fullness y through a Gaussian process:

[0168]

[0169] Where μ(x) is the mean function, usually taken as μ(x) = 0; k(x,x′) is the kernel function, here a radial basis function is used:

[0170]

[0171] in The characteristic length scale matrix, Let V be the signal variance.

[0172] Bayesian optimization selects the next optimal candidate input point x using a sampling function. * This is done to maximize potential gains in the parameter space. The acquisition function here uses the expected improvement as the acquisition function:

[0173] x t+1 =arg maz x∈χ EI(x) (46)

[0174]

[0175] Where y best This is the highest full-bowl rate observed to date. Let EI be the feasible region of the quadratic surface parameters. The analytical expression for EI is...

[0176] EI(x)=(μ(x)-y best )Φ(Z)+σ(x)φ(Z) (48)

[0177]

[0178] Where μ(x) and σ(x) are the mean and standard deviation of the surrogate model predictions, respectively, and Φ(·) and φ(·) are the cumulative distribution function and probability density function of the standard normal distribution, respectively.

[0179] The selected optimal quadratic surface parameters x t+1 The data is mapped to the corresponding digging point location, where the loader performs actual digging to obtain the true bucket full rate y. t+1 Update the training dataset:

[0180]

[0181] and utilize The Gaussian process surrogate model is retrained to gradually improve the model's prediction accuracy.

[0182] Through multiple iterations, the surrogate model achieves a balance between exploration and utilization in the parameter space. The optimization process terminates when the desired improvement value satisfies the convergence condition EI < ∈ [0, 0]. Ultimately, a high-precision full-bucket rate prediction model f is obtained. final (x) can obtain the full-bowl rate data of multiple candidate points based on the quadratic surface parameters of multiple sampling points through the full-bowl rate prediction proxy model.

[0183] Step S50: Based on the full bucket rate data of multiple candidate points, the optimal digging point is obtained through the marine predator optimization algorithm. In response to the optimal digging point, a control instruction set is generated. The control instruction set is used to control the target loader to perform the digging operation corresponding to the optimal digging point.

[0184] In step S50, firstly, based on the full-bucket rate data of multiple candidate points, multiple expected values ​​of full-bucket rate data are obtained. The specific steps are as follows:

[0185] To quantitatively evaluate the operational performance of different candidate digging points, a comprehensive fitness function is constructed to measure the overall performance of digging points in terms of loading efficiency, path cost, and operational continuity. The comprehensive fitness function consists of three parts: the expected value of the full bucket rate predicted based on a Bayesian surrogate model, the path cost term based on a hybrid A* algorithm, and the distance penalty term between the digging point and the previous digging point.

[0186] Let the set of candidate digging points be:

[0187]

[0188] Each digging point P i Corresponding quadratic surface parameter vector x i Based on the obtained proxy model f final (x) can predict the expected value of the full bucket rate at the digging point.

[0189] (1) Expected value of full capacity

[0190] The predicted mean μ(x) output by the surrogate model i This indicates the expected fill rate for that position:

[0191] E i =μ(x i (52)

[0192] To eliminate the influence of dimensions, it is normalized:

[0193]

[0194] Secondly, based on the expected values ​​of multiple full bucket rates, the path cost from the digging point to the fixed unloading point, and the distance penalty term between the digging point and the previous digging point, the comprehensive fitness of multiple digging points is obtained through a comprehensive fitness function, as detailed below:

[0195] Digging point P i to fixed unloading point P d The driving path is planned using the HybridA* algorithm. This algorithm considers the kinematic characteristics of the loader, such as its minimum turning radius and steering constraints, while also taking into account the feasibility and smoothness of the path. The path cost consists of three parts:

[0196] 1. Path length cost:

[0197]

[0198] Where P i,k N represents the k-th node on the path. i This represents the number of nodes in the path.

[0199] 2. Cumulative steering angle cost:

[0200] Define every three consecutive path nodes (P) i,k-1 ,P i,k ,P i,k+1 The steering angle is

[0201]

[0202] The cumulative steering angle cost is defined as:

[0203]

[0204] 3. Cost of cumulative reversal times:

[0205] If the angle between adjacent path segments is |Δθ i,k |Greater than the threshold θ revThis is considered one reversal operation, and the cumulative number of reversals is:

[0206]

[0207] in This is an indicator function.

[0208] Combining the three indicators, the total path cost is defined as:

[0209] C i =α L C L，i +α θ C θ，i +α r C r，i (58)

[0210] Where α L ,α θ ,α r This is the path cost weighting coefficient, reflecting the relative importance of path length, turning smoothness, and reversal frequency in path complexity assessment.

[0211] The path cost, after normalization, is as follows:

[0212]

[0213] Since path cost is a negative indicator, the smaller the value, the better the operation path.

[0214] To ensure the continuity and efficiency of loader digging operations, the next digging point P should be encouraged to be... i Compared to the previous excavation point P last Get as close as possible. Define the Euclidean distance between the two as:

[0215]

[0216] Distance D i The smaller the value, the closer the two excavation locations are, and the better the continuity of the operation. To ensure comparability between different candidate points, the distance values ​​are normalized:

[0217]

[0218] Normalized distance term It is used as a negative indicator in the overall fitness calculation to constrain the spatial distribution of continuous operations and prevent the digging positions of two adjacent digging operations from being too dispersed.

[0219] Taking into account the expected full bucket rate, path cost, and operational continuity, the comprehensive fitness function for the digging point is defined as follows:

[0220]

[0221] Where w1, w2, w3 ≥ 0, and satisfy:

[0222] w1 + w2 + w3 = 1 (63)

[0223] In the formula, The normalized result representing the expected value of the full-capacity rate is a positive indicator; This is the normalized value of the total path cost, and it is a negative indicator. The normalized value of the distance between digging points is used to characterize the continuity of operations and is a negative indicator. The weighting coefficients reflect the relative importance of different evaluation indicators in the comprehensive assessment: w1 reflects the degree of focus on loading efficiency; w2 controls the impact of path complexity on the comprehensive performance assessment; w3 is used to balance the spatial distribution of the work area and prevent a decrease in work efficiency caused by excessively large spacing between consecutive digging points.

[0224] Then, based on the comprehensive fitness of multiple digging points, the fitness values ​​of multiple candidate points are obtained through the marine predator optimization algorithm. The specific steps are as follows:

[0225] Each candidate digging point s j Its horizontal plane coordinates (x) have been obtained through B-spline curve and neighborhood boundary point analysis. j ,y j and bucket direction angle θ j To achieve efficient digging by the loader, it is necessary to search for the optimal digging point in the continuous x-axis space:

[0226] S best =(x * ,y * ,θ * (64)

[0227] The B-spline curve is represented as:

[0228]

[0229] Therefore, for a given x-coordinate, the y-coordinate can be uniquely determined:

[0230] y = S(x) (66)

[0231] The digging angle θ can be calculated using the neighborhood boundary points and the PCA fitting function from step 5):

[0232] θ=f θ (x, y) (67)

[0233] Through the above processing, the optimal digging point search problem can be transformed into a one-dimensional continuous optimization problem:

[0234]

[0235] The overall fitness function is defined as follows:

[0236]

[0237] The Marine Predators Algorithm (MPA) optimizes the search by simulating the global exploration and local exploitation behaviors of marine predators during hunting. Its core mechanisms include predator-prey interaction, a three-stage search strategy, and fish aggregation effects (FADs). The specific steps for finding the optimal digging point using MPA are as follows:

[0238] Initialization phase:

[0239] In the search interval [x min x max N individuals are randomly generated within the [group]. Each individual corresponds to a potential digging point, representing a set of candidate solutions. Key algorithm parameters are defined: Max_iter is the maximum allowed number of iterations, and Iter is the current iteration round number. Simultaneously, the globally optimal individual (predator) is initialized, with the individual with the best fitness in the initial population set as the current optimal solution, Elite.

[0240] Three-stage search strategy:

[0241] The algorithm balances global exploration (expanding the search range) and local development (fine-tuning) capabilities through a three-stage iteration. Different individual update rules are adopted in different stages, and the core relies on two random walk modes: Brownian motion and Lévy flight.

[0242] Brownian motion is a continuous random motion in which the particle displacement follows a normal distribution. It manifests as short-distance, high-frequency random jumps, making it suitable for precise searches within local regions. Its mathematical description is:

[0243] X t+1 =X t +σ·∈ (70)

[0244] Where X t Let σ be the current position, σ be the step size, and ∈ ~N(0,1) be a standard normal random variable.

[0245] Lévy flight is a non-Gaussian random walk with a step size following a Lévy distribution. It is characterized by mostly short-distance movements, occasionally involving extremely long jumps, making it suitable for global exploration. Its mathematical description is:

[0246] Phase 1: Exploration Phase

[0247] This phase primarily involves global exploration, where all prey move randomly using Brownian motion, traversing the search space to discover potential optimal regions. The individual update formula is:

[0248]

[0249]

[0250] in, Let i be the position of the i-th prey in round t. It is a standard normally distributed random vector (which follows Brownian motion properties). is a random number in the interval [0,1], and ⊙ represents element-wise multiplication.

[0251] Phase 2: Mixed Phase

[0252] This phase balances exploration and development, with the population divided into two parts based on location and updated using different strategies:

[0253] For the first group of individuals: continue global exploration using Lévy flight, utilizing its long jump capability to expand the search area. The update formula is:

[0254]

[0255] The latter part of the individual: It moves closer to the current optimal solution Elite through Brownian motion, focusing on local development, and the update formula is:

[0256]

[0257] in, Let P be a random vector following a Lévy distribution, P be a random number in the interval [0, 1], and CF be the predator decay coefficient, where CF = (1 - Iter / Max_iter). 2·(Iter / Max_iter) .

[0258] Phase 3: Convergence Phase

[0259] This stage focuses on localized, refined development. The entire population approaches the Elite through Lévy flight. At this point, the probability of long jumps during Lévy flight decreases, enabling rapid convergence to the optimal solution.

[0260] The individual update formula is:

[0261]

[0262] (3) FADs effect (random perturbation mechanism)

[0263] To avoid the algorithm getting trapped in local optima (premature convergence), the Fish Aggregating Devices (FADs) effect is introduced, and the updated individual positions are randomly perturbed:

[0264]

[0265] in, Let r be a binary random matrix, and r be a random number in the interval [0,1]. The population index is randomly shuffled to simulate random interactions among prey.

[0266] (4) Iterative update and optimal solution output

[0267] After each iteration, the fitness value F(x) of all individuals is calculated.

[0268] Finally, the fitness values ​​of multiple candidate points are searched to obtain the optimal digging point.

[0269] Calculate the fitness value F(x) of all individuals, and update the individual with the better fitness to the global optimum (Elite). When the iteration reaches Max_iter, stop the search and output the optimal digging point parameters: optimal x-coordinate: Thus, the optimal ordinate y is obtained. * =S(x * ) and optimal bucket angle θ * =f θ (x * ,y * The final optimal digging point is represented as:

[0270] S best =(x * y * θ * (80)

[0271] Among them, S best This is the final optimal digging point.

[0272] Using steps S10-S50 of this embodiment, three-dimensional point cloud data of a large material pile is acquired by lidar. After preprocessing and segmentation by a network model, block point cloud data is obtained. Then, based on the full bucket rate prediction proxy model, the full bucket rate data of candidate points is output. Finally, the optimal digging point is selected by the marine predator optimization algorithm and a control command set is generated. This can overcome the limitations of manual selection and the shortcomings of existing methods in terms of vision and evaluation mechanism, realize the autonomous and intelligent selection of digging points for large material piles, effectively improve the full bucket rate and efficiency of loader operation, reduce repeated digging behavior, reduce fuel consumption, and ensure the efficiency and stability of loading operations in large material pile scenarios.

[0273] Furthermore, the 3D point cloud data is preprocessed to obtain preprocessed point cloud data, including: using a tightly coupled iterative Kalman filter algorithm to perform multi-frame point cloud superposition processing on the 3D point cloud data to obtain superimposed point cloud data; using statistical filtering and voxel filtering to process the superimposed point cloud data to obtain filtered point cloud data; and dividing the filtered point cloud data into material pile and background point clouds based on spatial distribution and surface morphology features to obtain preprocessed point cloud data.

[0274] Furthermore, based on multiple segmented point cloud data, a full bucket rate prediction surrogate model is used to obtain full bucket rate data for multiple candidate points. This includes: dividing the preprocessed point cloud data into two dimensions on the horizontal projection plane with a preset grid size; extracting the material pile boundary points using the Alpha-shape algorithm to obtain an initial boundary point set; performing B-spline curve interpolation on the initial boundary point set to obtain a boundary curve; uniformly sampling the boundary curve to obtain a complete boundary point set; sampling the complete boundary point set to obtain initial sampling points; extracting the neighborhood point cloud corresponding to each initial sampling point; obtaining quadratic surface parameters for multiple sampling points through quadratic surface fitting; and obtaining full bucket rate data for multiple candidate points based on the quadratic surface parameters of multiple sampling points using the full bucket rate prediction surrogate model.

[0275] Further, the preprocessed point cloud data is divided into two dimensions on the horizontal projection plane with a preset grid size. The Alpha-Shape algorithm is used to extract the boundary points of the material pile to obtain an initial boundary point set. This includes: dividing the preprocessed point cloud data into two dimensions on the horizontal projection plane with a preset grid size to obtain multiple grid-divided point cloud data; judging the multiple grid-divided point cloud data based on a set height threshold to obtain a boundary grid index set; projecting the material pile point cloud onto the x–y plane, and using the Alpha-Shape algorithm to calculate the boundary contour of the boundary grid index set to obtain the initial boundary point set.

[0276] Furthermore, based on the full-bucket rate data of multiple candidate points, the optimal digging point is obtained through the marine predator optimization algorithm, including: obtaining multiple expected values ​​of full-bucket rate data based on the full-bucket rate data of multiple candidate points; obtaining the comprehensive fitness of multiple digging points through a comprehensive fitness function based on the multiple expected values ​​of full-bucket rate data, the path cost term from the digging point to the fixed unloading point, and the distance penalty term between the digging point and the previous digging point; obtaining the fitness values ​​of multiple candidate points through the marine predator optimization algorithm based on the comprehensive fitness of multiple digging points; and searching the fitness values ​​of multiple candidate points to obtain the optimal digging point.

[0277] Example 2

[0278] Figure 5This is a schematic block diagram illustrating the structure of a loader digging point autonomous selection device for large stockpiles, according to an exemplary embodiment. The device includes:

[0279] The acquisition module 210 is used to acquire three-dimensional point cloud data through lidar in response to the excavation point selection request command;

[0280] Processing module 220 is used to preprocess the 3D point cloud data to obtain preprocessed point cloud data;

[0281] The segmentation module 230 is used to segment the preprocessed point cloud data through a segmentation network model to obtain multiple segmented point cloud data.

[0282] The prediction module 240 is used to obtain the full-bucket rate data of multiple candidate points based on multiple segmented point cloud data through a full-bucket rate prediction proxy model.

[0283] The optimization module 250 is used to obtain the optimal digging point based on the full bucket rate data of multiple candidate points through the marine predator optimization algorithm. In response to the optimal digging point, a control instruction set is generated. The control instruction set is used to control the target loader to perform the digging operation corresponding to the optimal digging point.

[0284] Example 3

[0285] Figure 6 This is a structural block diagram of a terminal provided in an embodiment of this application. The terminal can be the terminal described in the above embodiments. The terminal can be a portable mobile terminal, such as a smartphone or tablet computer. The terminal may also be referred to as user equipment, portable terminal, or other names.

[0286] Typically, a terminal includes a processor and memory.

[0287] The processor may include one or more processing cores, such as a quad-core processor or an octa-core processor. The processor may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). The processor 301 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, the processor may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content required to be displayed on the screen. In some embodiments, the processor may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0288] The memory may include one or more computer-readable storage media, which may be tangible and non-transitory. The memory may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory are used to store at least one instruction, which is executed by a processor to implement a loader digging point autonomous selection method for large stockpiles provided in this application.

[0289] Those skilled in the art will understand that Figure 6 The structure shown does not constitute a limitation on the terminal and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0290] Example 4

[0291] In an exemplary embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, implements a method for autonomous selection of loader digging points for large stockpiles as provided in all embodiments of the present application.

[0292] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0293] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including—but not limited to—electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, capable of transmitting, propagating, or transmitting programs for use by or in connection with an instruction execution system, apparatus, or device.

[0294] The program code contained on a computer-readable medium may be transmitted using any suitable medium, including—but not limited to—wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0295] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, as well as conventional procedural programming languages ​​such as "C" or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0296] Example 5

[0297] In an exemplary embodiment, an application product is also provided, including one or more instructions that can be executed by the processor of the aforementioned device to complete the aforementioned method for autonomous selection of loader digging points for large stockpiles.

[0298] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A method for autonomously selecting loader digging points for large stockpiles, characterized in that, include: In response to the request command to select the excavation point, three-dimensional point cloud data is acquired through LiDAR; The three-dimensional point cloud data is preprocessed to obtain preprocessed point cloud data; The preprocessed point cloud data is segmented using a segmentation network model to obtain multiple segmented point cloud data. Based on multiple segmented point cloud data, the full-bucket rate data of multiple candidate points are obtained through a full-bucket rate prediction proxy model; Based on the full bucket rate data of multiple candidate points, the optimal digging point is obtained through the marine predator optimization algorithm. In response to the optimal digging point, a control instruction set is generated, which is used to control the target loader to perform the digging operation corresponding to the optimal digging point.

2. The method for autonomous selection of loader digging points for large stockpiles according to claim 1, characterized in that, The three-dimensional point cloud data is preprocessed to obtain the preprocessed point cloud data, including: The three-dimensional point cloud data is processed by multi-frame point cloud overlay using a tightly coupled iterative Kalman filter algorithm to obtain the overlay point cloud data. The superimposed point cloud data is processed using statistical filtering and voxel filtering to obtain filtered point cloud data. Based on spatial distribution and surface morphology features, the filtered point cloud data is divided into material pile and background point cloud to obtain the preprocessed point cloud data.

3. The method for autonomous selection of loader digging points for large stockpiles according to claim 1, characterized in that, Based on multiple segmented point cloud data, the full-bowl rate prediction proxy model is used to obtain full-bowl rate data for multiple candidate points, including: The preprocessed point cloud data is divided into two dimensions on the horizontal projection plane with a preset grid size, and the material pile boundary points are extracted using the Alpha-shape algorithm to obtain the initial boundary point set. B-spline curve interpolation is performed on the initial boundary point set to obtain the boundary curve; The boundary curve is uniformly sampled to obtain a complete set of boundary points; The complete set of boundary points is sampled to obtain initial sampling points; For each initial sampling point, the corresponding neighborhood point cloud is extracted, and the quadratic surface parameters of multiple sampling points are obtained through quadratic surface fitting. Based on the quadratic surface parameters of multiple sampling points, the full-bowl rate data of multiple candidate points are obtained through the full-bowl rate prediction proxy model.

4. The method for autonomous selection of loader digging points for large stockpiles according to claim 3, characterized in that, The preprocessed point cloud data is divided into two dimensions on its horizontal projection plane using the preset grid size. The Alpha-shape algorithm is used to extract the material pile boundary points, resulting in the initial boundary point set, which includes: The preprocessed point cloud data is divided into two dimensions on the horizontal projection plane using the preset grid size to obtain point cloud data with multiple grid divisions. Based on a set height threshold, the point cloud data after multiple grid divisions are judged to obtain the boundary grid index set; The point cloud of the material pile is projected onto the x–y plane, and the boundary contour of the boundary raster index set is calculated using the Alpha-Shape algorithm to obtain the initial boundary point set.

5. The method for autonomous selection of loader digging points for large stockpiles according to claim 3, characterized in that, Based on the full-bucket rate data of multiple candidate points, the optimal digging point is obtained through the marine predator optimization algorithm, including: Based on the full-bottle rate data of multiple candidate points, multiple expected values ​​of full-bottle rate data are obtained; Based on the expected values ​​of multiple full bucket rate data, the path cost term from the digging point to the fixed unloading point, and the distance penalty term between the digging point and the previous digging point, the comprehensive fitness of multiple digging points is obtained through a comprehensive fitness function. Based on the comprehensive fitness of multiple digging points, the fitness values ​​of multiple candidate points are obtained through the marine predator optimization algorithm; The fitness values ​​of multiple candidate points are searched to obtain the optimal digging point.

6. A loader digging point autonomous selection device for large stockpiles, characterized in that, include: The acquisition module is used to acquire three-dimensional point cloud data through LiDAR in response to the excavation point selection request command; The processing module is used to preprocess the three-dimensional point cloud data to obtain preprocessed point cloud data; The segmentation module is used to segment the preprocessed point cloud data using a segmentation network model to obtain multiple segmented point cloud data. The prediction module is used to obtain the full-bucket rate data of multiple candidate points based on multiple segmented point cloud data through a full-bucket rate prediction proxy model; The optimization module is used to obtain the optimal digging point based on the full bucket rate data of multiple candidate points through the marine predator optimization algorithm, and to generate a control instruction set in response to the optimal digging point. The control instruction set is used to control the target loader to perform the digging operation corresponding to the optimal digging point.

7. A terminal, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to perform the steps of the loader digging point autonomous selection method for large stockpiles as described in any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When executed by the processor, the program instructions implement the steps of the loader digging point autonomous selection method for large stockpiles as described in any one of claims 1 to 5.

Images