A semantic segmentation method for a blast pile excavation area

Abstract

The application belongs to the technical field of open-pit mine intelligent shovel loading, blasting pile simulation, three-dimensional point cloud processing and semantic segmentation, and specifically discloses a semantic segmentation method for a blasting pile excavation area, which comprises the following steps: generating a blasting pile three-dimensional model based on preset blasting pile parameters and a physical engine, and extracting surface point coordinates and normal vectors; calculating excavability scores according to relative height, normalized radial distance, slope and orientation, and correcting the scores based on occlusion detection; generating excavation area labels according to the corrected scores and the proportion of excavable points; constructing physical prior features, encoding the features, and fusing the encoded features with embedded features of a point cloud semantic segmentation network to obtain enhanced point cloud features; and outputting semantic segmentation results of the blasting pile excavation area based on the enhanced point cloud features. The application can improve the reliability and engineering applicability of the segmentation results of the blasting pile excavation area.

Description

Technical Field

[0001] This application belongs to the technical fields of intelligent loading and unloading in open-pit mines, blasting simulation, 3D point cloud processing and semantic segmentation, and more specifically, relates to a semantic segmentation method for blasting excavation areas. Background Technology

[0002] In open-pit mining loading operations, blast piles are the direct targets of excavation equipment. The block size distribution, slope morphology, degree of boundary fragmentation, and local obstruction of the blast pile surface affect the bucket's entry position, excavation trajectory, and subsequent loading control. For unmanned or minimally manned loading systems, reliably identifying the blast pile excavation area is a fundamental step in selecting the loading location and making operational decisions.

[0003] The acquisition of real-world point clouds from blast piles is easily affected by factors such as observation perspective, dust, occlusion, safety distance, and work rhythm. It is typically difficult to repeatedly collect multiple sets of high-quality point cloud data under the same blast pile morphology at the mine site. Publicly available data is mostly geared towards perception of mine roads, vehicles, equipment, or the environment, and rarely directly covers the semantics of the excavation area on the blast pile surface. Even after acquiring blast pile point clouds, labeling the excavation area remains challenging because the excavation area is not a visual semantic category with relatively fixed boundaries like roads, vehicles, or buildings. Instead, it is an operational semantic related to the working conditions of the excavating equipment. Whether a surface point belongs to the excavation area usually depends on its height, distance from the edge of the pile, slope direction, slope gradient, and whether it is occluded by rocks above. Existing point cloud semantic segmentation methods typically rely on point coordinates, local neighborhoods, and contextual relationships to learn features and have been applied in road scenes, indoor scenes, and conventional object recognition. However, in blast pile scenes, with overlapping rocks, strong surface undulations, and complex boundary transitions, general segmentation networks that rely solely on geometric appearance are prone to misclassification in slope transition areas, exposed edge areas, and occluded areas. Existing blast pile simulation or measurement methods are mostly used for blast pile volume calculation, pile morphology analysis, block size distribution statistics, or blasting effect evaluation. They typically do not further integrate the simulated blast pile surface points with accessibility rules for loading operations to generate excavation area labels. Existing automatic labeling methods also largely rely on manual interaction, semi-supervised learning, or cross-modal mapping, making it difficult to directly provide calculation rules for excavable areas within the blast pile surface. These factors lead to high costs in constructing blast pile excavation area samples, unstable labeling rules, and difficulty in utilizing operational physical constraints in the segmentation model, thus affecting the reliability and engineering applicability of the segmentation results.

[0004] Therefore, improving the reliability and engineering applicability of the blasting excavation area segmentation results is an urgent problem to be solved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the purpose of this application is to provide a semantic segmentation method for blasting excavation areas, which can effectively improve the reliability and engineering applicability of blasting excavation area segmentation results.

[0006] To achieve the above objectives, in a first aspect, this application provides a semantic segmentation method for heap mining regions, comprising the following steps: S10, Generate a three-dimensional model of the blast pile based on preset blast pile parameters and a physics engine, and extract the coordinate information and normal vector information of surface points from the three-dimensional model of the blast pile; S20, calculate the excavability score of the surface point based on the relative height, normalized radial distance, slope information and orientation information, and correct the excavability score based on the occlusion detection results to obtain the corrected excavability score. S30, based on the corrected excavability score, the proportion of excavable points in the local area is statistically analyzed, and a blasting excavation area label is generated based on the proportion of excavable points. S40, construct physical prior features based on the coordinate information and the normal vector information, input the physical prior features into the physical perception feature encoder to obtain physical coding features, and then fuse the physical coding features with the point cloud embedding features of the point cloud semantic segmentation network to obtain enhanced point cloud features; during the training phase, use the burst mining region label as a supervision signal to train the point cloud semantic segmentation network and the physical perception feature encoder. S50, based on the enhanced point cloud features, output the semantic segmentation result of the burst mining region.

[0007] As a further preferred embodiment, in step S10, generating the three-dimensional model of the burst pile based on preset burst pile parameters and a physics engine includes: based on the bottom radius of the pile body... Anchorage Height correction factor and surface irregularity coefficient ,according to Determine the height of the blast pile ; Control parameters based on the profile shape at the height of the blast pile according to Determine different heights Effective radius at ( ), and use the noise function according to By perturbing the effective radius, an outer contour of the explosive pile with surface undulations is obtained, wherein, It is the azimuth angle. It is a two-dimensional noise function.

[0008] As a further preferred embodiment, the generation of the 3D model of the burst pile based on preset burst pile parameters and a physics engine further includes: based on normalized height... Normalized radial distance Baseline probability Power exponent parameters , and particle size separation strength ,according to Determine the probability of large-diameter stones appearing in the blast pile. ,in This is the horizontal distance from a surface point to the central axis of the stack. The stone model is divided into at least two particle size levels according to the particle size composition. The sampling ratio of each particle size level at different heights and radial positions is adjusted according to the probability of the occurrence of large-sized stones to obtain the distribution of blasted stones by stratified sampling and dynamic proportioning.

[0009] As a further preferred embodiment, in step S10, extracting the coordinate information and normal vector information of surface points from the blast pile 3D model includes: traversing the stone objects in the blast pile 3D model, filtering stone objects located below a preset underground height, and converting the vertex coordinates and normal vectors of the retained stone objects to the world coordinate system.

[0010] As a further preferred embodiment, in step S20, the exploitability score... Highly rated Marginal scoring Slope rating and orientation rating according to We get the weighted average, where , , , The weighting coefficient; the height score The edge score is used to represent the position of a surface point in the direction of the explosion height. The slope score is used to represent the position of a surface point relative to the maximum radius of the current height processing theory. The orientation score is used to indicate whether the slope corresponding to a surface point meets the working posture constraints of the excavating equipment. Used to represent the relationship between the horizontal projection of the surface point normal vector and the radial direction.

[0011] As a further preferred option, the height score according to Confirmed, among which The height of the surface point. The median of the effective height range, For the effective height range width, , , and These are the lower and upper limits of the effective height range, respectively; The edge score according to Confirmed, among which For normalized distance, The horizontal distance from a surface point to the central axis of the stack. At current height The theoretical maximum radius at that location, This is the edge threshold; The slope rating according to Confirmed, among which Let be the angle between the surface point normal vector and the vertical direction. This represents the maximum excavable slope. The orientation score according to Confirmed, among which The dot product of the horizontal projection of the surface point normal vector and the radial direction is given. The threshold is the outward threshold.

[0012] As a further preferred embodiment, in step S20, the occlusion detection result is obtained by ray detection: from surface points Along its normal vector A ray is emitted in a certain direction; if the ray is within a preset detection distance... If the surface point intersects with other colliding bodies, it is determined that the surface point is in an occluded state; if it does not intersect, it is determined that the surface point is not in an occluded state. Represented as: ,in, This indicates that the ray detection hit other colliders; The correction of the exploitability score based on the occlusion detection results includes: adjusting the occlusion status according to... Determine the revised exploitability score; for unoccluded points, the revised score remains unchanged. For occluded points, the corrected score is reduced to [missing value]. .

[0013] As a further preferred embodiment, in step S30, the step of calculating the proportion of excavable points in a local area based on the corrected excavability score and generating a blasting excavation area label based on the proportion of excavable points includes: calculating that the corrected excavability score in the local area is greater than or equal to a score threshold. The surface point ratio is used to obtain the excavable point ratio. ,like ≥ Then, the local area is marked as the mining area; where, The threshold value is used for the proportion of the excavated area. The labels for the blasting excavation area include at least one of point-level labels, stone object-level labels, and local area-level labels.

[0014] As a further preferred embodiment, in step S40, the physical prior features include relative height. Normalized radial distance and slope characteristics At least one of the following; wherein, relative height The radial distance is obtained by normalizing the height coordinates of the surface points; The slope characteristics are obtained by normalizing the distance from the surface point to the center of the pile in the horizontal plane. It is obtained from the vertical component of the surface point normal vector; The physical perception feature encoder includes a multilayer perceptron for mapping the physical prior features to the same feature dimension as the point cloud embedded features; the point cloud semantic segmentation network is a point cloud encoding and decoding network based on Point Transformer. The process of fusing the physical coding features with the point cloud embedding features of the point cloud semantic segmentation network includes: multiplying the physical coding features by a learnable scaling factor and then adding the result to the point cloud embedding features.

[0015] Secondly, this application provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the semantic segmentation method for the blasting excavation area described in any one of the above-mentioned methods.

[0016] This application achieves the following effects: By generating a 3D model of a blast pile based on preset blast pile parameters and a physics engine, the reliance on a large number of real labeled point clouds can be reduced, allowing samples to cover different shapes and occlusion relationships; by calculating and correcting the excavability score based on relative height, normalized radial distance, slope information, orientation information, and occlusion detection results, labels for the blast pile excavation area are generated, transforming the operating conditions of the excavation equipment into reproducible automatic labeling rules, making the labels reflect physical constraints such as height, edge, slope, orientation, and occlusion; by constructing physical prior features and fusing them into a point cloud semantic segmentation network, the network learns the appearance features of the point cloud while utilizing physical cues consistent with the labeling rules. The synergistic effect of these methods ensures that both the label generation rules and the network-learned features reflect the physical constraints of the excavation operation, thereby improving the reliability and engineering applicability of the blast pile excavation area segmentation results. Attached Figure Description

[0017] Figure 1This is a schematic diagram of the overall process of the automatic annotation and semantic segmentation method for the blasting excavation area provided in the embodiments of this application; Figure 2 This is a schematic diagram of the automatic generation process of the blasting excavation area label provided in the embodiments of this application; Figure 3 This is a schematic diagram of the point cloud semantic segmentation training and inference process based on physical prior features provided in the embodiments of this application; Figure 4 This is a detailed structural diagram of the physical perception feature encoding module and the point cloud semantic segmentation network provided in the embodiments of this application; Figure 5 This is a schematic diagram of the automatic labeling and semantic segmentation device for the blasting excavation area provided in the embodiments of this application; Figure 6 This is a schematic diagram comparing the performance of different semantic segmentation methods provided in the embodiments of this application; Figure 7 This is a schematic diagram of ablation experiment results for different combinations of physical prior features in the physical sensing feature encoding module provided in this application embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It should be noted that the blast pile excavation area provided in this application can be understood as the surface area of ​​the blast pile where the bucket of the excavating equipment can effectively contact and scoop up the material. This area typically needs to meet conditions such as surface exposure, suitable slope, accessible location, and no obstruction by overhead rocks. For the 3D mesh model of the blast pile... It can be used on surface vertices Perform binary classification labeling and output labels. , where 1 represents the excavated area and 0 represents the non-excavated area.

[0020] Please refer to Figure 1 This diagram illustrates the overall flow of an automatic annotation and semantic segmentation method for blast-dammed excavation areas provided in an embodiment of this application. Specifically, the method may include steps S1 to S9.

[0021] Step S1: Obtain the burst generation parameters. The burst generation parameters include the base radius of the pile. Anchorage Height correction factor Surface irregularity coefficient Parameters for controlling the profile shape Particle size separation strength And the proportions of components with different particle sizes.

[0022] Optionally, in one possible implementation of this embodiment, the parameters for generating the blast pile can be determined based on the bench height at the mine site, the material accumulation range after blasting, the angle of repose of coal and rock, the statistical results of particle size, or empirical parameters. Alternatively, random sampling can be performed within a preset range to obtain blast pile samples of different forms, such as gentle loose, gentle dense, steep loose, and steep dense.

[0023] Step S2: Generate a 3D model of the burst pile based on the burst pile generation parameters and the physics engine. In one implementation, the pile height... Determine according to formula (1): (1) in, The radius of the base of the pile is 1. Anchorage, For high correction factor, The surface irregularity coefficient is used to regulate the basic stable shape of the reactor structure. The height correction coefficient and the surface irregularity coefficient are used to adjust the height and surface roughness of the reactor structure.

[0024] At any height At this location, the effective radius of the pile body ( It can be determined according to formula (2): (2) in, These are parameters for controlling the profile shape. Adjustment is required. Afterwards, blast pile profiles with varying degrees of steepness can be obtained. The effective radius is used to constrain the initial distribution range of stone objects at the corresponding height.

[0025] To simulate the irregular boundary of the exploded surface, a noise function can be used to perturb the effective radius. The perturbed radius... It can be determined according to formula (3): (3) in, It is the azimuth angle. This is a two-dimensional noise function. In specific implementations, the noise function can be Perlin noise, or other noise functions that can generate continuous surface perturbations.

[0026] During the formation of a heap, stones of different sizes undergo natural sorting. To simulate this phenomenon, a normalized height can be used... and normalized radial distance Determine the probability of large-diameter stones. As shown in equation (4): (4) in, The horizontal distance from the point to the central axis of the pile body. As the baseline probability, and For the power exponent parameter, This refers to particle size separation intensity. A larger one... This makes it easier for large-diameter stones to be distributed in the outer area at the bottom of the pile; when When the value is smaller, the spatial separation of stones of different sizes decreases.

[0027] Optionally, in one possible implementation of this embodiment, the stone model can be divided into at least two particle size levels according to the particle size composition, and the sampling ratio of each particle size level at different heights and radial positions can be adjusted according to the probability of large-diameter stones to obtain the distribution of blasted stones with layered sampling and dynamic proportioning.

[0028] Optionally, in one possible implementation of this embodiment, the 3D model of the burst pile is generated by the Unity physics engine. It should be noted that Unity is only one possible implementation; other physics engines capable of supporting colliders, rigid body motion, and raycasting can also be used to implement the above steps. After the physics engine generates the burst pile, it can retain the mesh, colliders, world coordinate transformations, and normal vector information of each stone object.

[0029] Step S3: Extract the coordinate information and normal vector information of surface points from the 3D model of the blast pile. Specifically, iterate through each stone object in the 3D model of the blast pile, filter out stone objects located below a preset underground height, and transform the vertex coordinates and normal vectors of the remaining stone objects to the world coordinate system. The obtained data can be recorded as a coordinate array. and normal vector array .

[0030] Optionally, in one possible implementation of this embodiment, when traversing stone objects, if a stone object is empty, has no mesh vertices, or its bounding box's highest point is below a preset height threshold, then the stone object is skipped. This can reduce the impact of stones below ground and invalid meshes on subsequent label generation.

[0031] Step S4: Calculate the exploitability score of the surface points. For surface points... Overall score It can be determined according to formula (5): (5) in, , , , These are the weighting coefficients. , , , These are height score, edge score, slope score, and orientation score. The weighting coefficients can be determined based on excavation equipment parameters, site experience, and validation data.

[0032] High rating This indicates whether a surface point is within a suitable excavation height range. Let the surface point height be... The effective height range is [ , Then the high score can be determined according to formula (6): (6) in, , When a surface point is outside the effective height range, it is not treated as a priority excavation point.

[0033] Edge rating This is used to represent the position of a surface point relative to the edge of the stack. Let the horizontal distance from the surface point to the central axis of the stack be... The theoretical maximum radius at the current altitude is Normalized distance is The edge threshold is Then the edge score can be determined according to formula (7): (7) Among them, edge threshold This is used to distinguish between areas near the center of the stack and areas near the outer edge. The outer edge area, which is more easily accessible to loading equipment, receives a relatively higher edge score.

[0034] Slope rating This is used to indicate whether the slope corresponding to a surface point is suitable for bucket entry. Let the angle between the surface point's normal vector and the vertical direction be θ. The maximum excavable slope is The slope score can then be determined according to formula (8): (8) in, The selection criteria can be determined based on the working posture of the excavating equipment, the cutting capacity of the bucket, and on-site experience. Surface points with a slope less than or close to the maximum excavable slope are more suitable as candidate points for the excavation area.

[0035] Towards rating This is used to represent the degree of agreement between the surface point normal vector and the radial direction. Let the dot product of the horizontal projection of the surface point normal vector and the radial direction be... The outward threshold is Then the orientation score can be determined according to formula (9): (9) Among them, the outward threshold This score is used to determine whether the surface normal has a good exposed orientation. This score can reduce the occurrence of surface points facing the interior of the pile or located in a reverse position being mistakenly labeled as excavation areas.

[0036] Please refer to Figure 2 The diagram illustrates an automatic generation process for blasting excavation area labels provided in one embodiment of this application. Figure 2 The height score, edge score, slope score, and orientation score are all included in the excavability score calculation, and ray occlusion detection generates occlusion status. Both are used together to generate excavation area labels.

[0037] Step S5: Perform occlusion detection and correct the exploitability score. For surface points... From its position Along the normal vector A ray is emitted in a specific direction. If the ray is within the detection range... If a surface point intersects with other colliders, it is considered to be occluded. The occlusion state can be expressed by equation (10): (10) Optionally, in one possible implementation of this embodiment, the ray starting point can be offset by a preset distance along the normal vector direction, for example from... Rays begin to be emitted from the location, among which It is a positive number. This processing can reduce misjudgments caused by the intersection of the ray with the surface point's own collision body.

[0038] Based on the occlusion state, the modified exploitability score can be obtained according to Equation (11). : (11) in, This represents the occlusion suppression coefficient. For unoccluded points, the corrected score remains unchanged. For obstructed areas, the score is reduced. This treatment reduces the likelihood of recessed areas and areas covered by large rocks inside the pile being mistakenly identified as excavation areas.

[0039] Step S6: Generate blast excavation area labels based on the revised excavability score. In one implementation, statistically analyze local areas that meet the criteria... The surface point ratio is used to obtain the excavable point ratio. The proportion of excavable points can also be used as the exposure proportion of a local area. If If the condition is met, the local area is marked as a mining area; otherwise, it is marked as a non-mining area. As the scoring threshold, This is the proportional threshold.

[0040] Optionally, in one possible implementation of this embodiment, the local region can be a set of surface points corresponding to a single stone object, or a set of points divided according to spatial neighborhood, grid blocks, or sampling windows. The point-level label can be determined by the corrected excavability score of a single surface point, and the stone object-level label or the local region-level label can be determined by the proportion of excavable points within the corresponding range.

[0041] When exporting labels, surface point coordinates, normal vectors, excavation area labels, and occlusion labels can be exported as arrays respectively, resulting in... , , and .in, Indicates the tag for the mining area. Indicates an occluded state.

[0042] Step S7: Construct physical prior features. For the input point cloud, physical prior features can be constructed according to equation (12) based on the coordinate information and normal vector information. : (12) in, Relative height, For normalized radial distance, This refers to slope characteristics. In one implementation, It can be obtained by normalizing the height coordinates of the point. It can be obtained by normalizing the distance from a point in the horizontal plane to the center of the pile. It can be obtained from the vertical component of the normal vector.

[0043] The aforementioned physical prior features correspond at least partially to the scoring criteria in automatic labeling. In other words, the height, distance, and slope information used during label generation can continue to be used as input features during network training and inference, enabling the segmentation network to rely not only on the appearance of point clouds but also on physical cues in task judgment.

[0044] Step S8: Input the physical prior features into the physical perception feature encoding module. For example... Figure 4 As shown, the physical perception feature encoding module includes a multilayer perceptron, which maps low-dimensional physical prior features to the same dimension as the point cloud embedding features. If the point cloud embedding features are... Physical coding characteristics are This enhances the features of the point cloud. It can be determined according to formula (13): (13) in, The scaling factor is learnable. Using this fusion method, the network retains the contextual representation of the point cloud while acquiring height, distance, and slope information relevant to the mining operation.

[0045] Optionally, in one possible implementation of this embodiment, the physically-aware feature encoding module may include a combination of Linear layers, GELU activation layers, and Linear layers. The physically-aware feature encoding module does not alter the attention calculation process in the point cloud semantic segmentation network; instead, it injects physically encoded features into the point cloud embedding features to maintain the context modeling capability of the network backbone.

[0046] In one implementation, the point cloud semantic segmentation network employs a Transformer-based encoder-decoder structure, such as Point Transformer V3. It should be noted that Point Transformer V3 is only one optional backbone network; other point cloud segmentation networks capable of receiving point cloud embedded features and outputting point-by-point semantic results can also be used.

[0047] Step S9: Output semantic segmentation results based on enhanced point cloud features. During the training phase, the automatically labeled points are used... As a supervisory signal, the parameters of the point cloud semantic segmentation network and the physical perception feature encoding module are updated, so that the operational rules in the label generation process are incorporated into the network feature representation. In the inference stage, after inputting the burst point cloud to be processed, the network outputs the probability that each point belongs to the mined area and the non-mined area, and obtains the segmentation result accordingly.

[0048] In one implementation, the training loss can be a class-weighted cross-entropy loss. The class weights can be determined according to equation (14): (14) in, The total number of training samples. For the number of categories, For the first Number of sample points. This setting is used to alleviate the training bias caused by the imbalance in the number of points between the mined and unmined regions, so that local boundary regions and regions with significant changes in exposure ratio can be more fully constrained during training.

[0049] In the above implementation, blasting simulation, automatic labeling, and semantic segmentation work together. The simulation process provides blasting samples with controllable morphological differences, the automatic labeling process converts the loading operation conditions into reproducible labels, and the physical perception feature encoding module then uses these operation conditions as part of the network input. This forms a closed loop from sample generation and label generation to model training and inference.

[0050] Optionally, in one possible implementation of this embodiment, multiple 3D models of the burst pile can be generated and divided into training and testing sets. The multiple 3D models of the burst pile can cover different average slope angles, different surface irregularities, different average heights, and different proportions of occlusion points, so that the training samples simultaneously contain morphological differences such as gentle and steep slopes, loose and dense structures.

[0051] In one embodiment, the dataset contains 1,000 3D models of explosive piles, divided into training and testing sets at an 8:2 ratio, with each model sampling approximately 100,000 points. This dataset setup is only for validation purposes; in actual implementation, the number of models, points, and the division ratio can be adjusted based on computing resources, sampling density, and on-site mine conditions.

[0052] In one embodiment, such as Figure 6 As shown, after adopting the physical prior feature encoding module, the PTv3+PAFE model achieves an average intersection-over-union ratio (IoU), average class accuracy, and overall accuracy of 85.07%, 92.01%, and 93.72%, respectively, on the simulation dataset. Compared to PTv3Base, the PTv3+PAFE model improves the average IoU, average class accuracy, and overall accuracy by 2.30 percentage points, 2.80 percentage points, and 0.84 percentage points, respectively. The above data are used to illustrate the effectiveness of physical prior enhancement for semantic segmentation of blast-dammed excavation areas and are not intended to limit the scope of protection of this application.

[0053] In one embodiment, such as Figure 7 As shown, after ablation verification of relative height, normalized radial distance, and slope features, it is evident that introducing any one of these features individually can improve the average crossover ratio (CRO). When all three types of physical prior features are input, the average CRO reaches 85.07%. Even after removing any single physical prior feature, the result is still higher than the PTv3 Base baseline, indicating that different physical prior features have a complementary effect and can collectively enhance the point cloud semantic segmentation network's ability to discriminate blast-dump areas.

[0054] This application can also be applied to real-world pile point clouds. For pile point clouds acquired by real 3D vision devices, coordinate normalization, normal vector estimation, and physical prior feature construction can be performed first, and then the data can be input into a trained semantic segmentation network to obtain the segmentation results of the mined regions.

[0055] like Figure 3 As shown, optionally, in one possible implementation of this embodiment, the model training process and the field inference process are executed separately. The training server uses simulated pile burst samples and automatically labeled data to train a point cloud semantic segmentation network, and the field perception device calls the trained network to perform region segmentation on real pile burst point clouds.

[0056] Optionally, in one possible implementation of this embodiment, in order to adapt to real observation conditions, viewpoint constraints, occlusion truncation, incomplete sampling, or environmental disturbance simulation can be added during the simulation stage to make the training point cloud closer to the burst point cloud collected on site.

[0057] The segmentation results obtained by the above method can be used for bucket entry point selection, excavation trajectory planning, loading operation area prompting, and intelligent loading perception in open-pit mines. It should be noted that the above uses are only for illustrating the application scenarios of the segmentation results and do not limit the method of this application to specific models of excavating equipment.

[0058] like Figure 5 As shown, this application also provides an automatic annotation and semantic segmentation device for blast pile excavation areas. The device may include a blast pile simulation module, an automatic annotation module, a physical feature construction module, and a semantic segmentation module.

[0059] The bursting pile simulation module is used to execute steps S1 and S2, including obtaining bursting pile generation parameters and generating a three-dimensional bursting pile model based on preset bursting pile parameters and a physics engine.

[0060] The automatic labeling module is used to perform steps S3 to S6, including extracting the coordinate information and normal vector information of surface points from the 3D model of the blast pile, calculating the excavability score, performing occlusion detection, correcting the excavability score, and generating labels for the blast pile excavation area.

[0061] The physical feature construction module is used to execute step S7, which includes constructing physical prior features based on coordinate information and normal vector information. The semantic segmentation module is used to execute steps S8 and S9, which includes encoding the physical prior features and fusing them with point cloud embedded features, and outputting the semantic segmentation results of the burst mining region.

[0062] The modules described above can be implemented in software or through a combination of hardware and software. They can be deployed on the same electronic device or on different devices such as simulation servers, training servers, and on-site sensing devices. This embodiment does not impose any particular limitations on this.

[0063] According to embodiments of this application, an electronic device is also provided. The electronic device may include a processor, a memory, an input interface, an output interface, and a communication interface. The processor can execute one or more steps of the aforementioned automatic annotation and semantic segmentation method for blast-dammed excavation areas according to a computer program stored in the memory.

[0064] The processor can be a central processing unit, graphics processing unit, artificial intelligence computing chip, digital signal processor, controller, microcontroller, or other computing unit with processing capabilities. The memory can include random access memory, read-only memory, solid-state drive, disk, optical disk, or other media capable of storing programs and data.

[0065] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, which, when executed by a processor, causes the processor to perform the aforementioned automatic annotation and semantic segmentation method for blast-dig regions.

[0066] The program code used to implement the methods of this application may be written in one or more programming languages. The program code may be executed on a local electronic device, or partially on a local electronic device and partially on a remote server, or entirely on a remote server or distributed system.

[0067] It should be noted that, for the sake of ease of description, the foregoing method embodiments are presented as a series of steps. Those skilled in the art should understand that this application is not limited to the described order of steps. Where the technical solution of this application can be implemented, some steps may be executed in a different order, or they may be executed in parallel or simultaneously.

[0068] In the above embodiments, the descriptions of each embodiment have different focuses. Parts not described in detail in a particular embodiment can be found in the relevant descriptions of other embodiments.

[0069] Those skilled in the art can adjust the parameter range, physics engine, noise function, segmentation network, loss function, and deployment method according to equipment specifications, mine site conditions, and point cloud acquisition methods. These adjustments do not alter the technical approach of this application, which utilizes blast pile physics simulation, automatic marking of excavable areas, and physical prior augmentation segmentation to identify blast pile excavation areas.

[0070] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this application can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this application can be achieved, and this application does not impose any restrictions on this.

[0071] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A semantic segmentation method for a bursty mining region, characterized in that, Includes the following steps: S10, Generate a three-dimensional model of the blast pile based on preset blast pile parameters and a physics engine, and extract the coordinate information and normal vector information of surface points from the three-dimensional model of the blast pile; S20, calculate the excavability score of the surface point based on the relative height, normalized radial distance, slope information and orientation information, and correct the excavability score based on the occlusion detection results to obtain the corrected excavability score. S30, based on the corrected excavability score, the proportion of excavable points in the local area is statistically analyzed, and a blasting excavation area label is generated based on the proportion of excavable points. S40, construct physical prior features based on the coordinate information and the normal vector information, input the physical prior features into the physical perception feature encoder to obtain physical coding features, and then fuse the physical coding features with the point cloud embedding features of the point cloud semantic segmentation network to obtain enhanced point cloud features. During the training phase, the burst-dig region labels are used as supervision signals to train the point cloud semantic segmentation network and the physical perception feature encoder. S50, based on the enhanced point cloud features, output the semantic segmentation result of the burst mining region.

2. The semantic segmentation method for the excavated area according to claim 1, characterized in that, In step S10, generating a 3D model of the bursting pile based on preset bursting pile parameters and a physics engine includes: based on the bottom radius of the pile body... Anchorage Height correction factor and surface irregularity coefficient ,according to Determine the height of the blast pile ; Control parameters based on the profile shape at the height of the blast pile according to Determine different heights Effective radius at ( ), and use the noise function according to By perturbing the effective radius, an outer contour of the explosive pile with surface undulations is obtained, wherein, It is the azimuth angle. It is a two-dimensional noise function.

3. The semantic segmentation method for the excavated area according to claim 2, characterized in that, The process of generating a 3D model of a burst pile based on preset burst pile parameters and a physics engine also includes: based on normalized height... Normalized radial distance Baseline probability Power exponent parameters , and particle size separation strength ,according to Determine the probability of large-diameter stones appearing in the blast pile. ,in This is the horizontal distance from a surface point to the central axis of the stack. The stone model is divided into at least two particle size levels according to the particle size composition. The sampling ratio of each particle size level at different heights and radial positions is adjusted according to the probability of the occurrence of large-sized stones to obtain the distribution of blasted stones by stratified sampling and dynamic proportioning.

4. The semantic segmentation method for the excavated area according to claim 1, characterized in that, In step S10, extracting the coordinate information and normal vector information of surface points from the blast pile 3D model includes: traversing the stone objects in the blast pile 3D model, filtering stone objects located below a preset underground height, and converting the vertex coordinates and normal vectors of the retained stone objects to the world coordinate system.

5. The semantic segmentation method for the excavated area according to claim 1, characterized in that, In step S20, the exploitability score Highly rated Marginal scoring Slope rating and orientation rating according to We get the weighted average, where , , , The weighting coefficient; the height score The edge score is used to represent the position of a surface point in the direction of the explosion height. The slope score is used to represent the position of a surface point relative to the maximum radius of the current height processing theory. The orientation score is used to indicate whether the slope corresponding to a surface point meets the working posture constraints of the excavating equipment. Used to represent the relationship between the horizontal projection of the surface point normal vector and the radial direction.

6. The semantic segmentation method for the excavated area according to claim 5, characterized in that, The height rating according to Confirmed, among which The height of the surface point. The median of the effective height range, For the effective height range width, , , and These are the lower and upper limits of the effective height range, respectively; The edge score according to Confirmed, among which For normalized distance, The horizontal distance from a surface point to the central axis of the stack. At current height The theoretical maximum radius at that location, This is the edge threshold; The slope rating according to Confirmed, among which Let be the angle between the surface point normal vector and the vertical direction. This represents the maximum excavable slope. The orientation score according to Confirmed, among which The dot product of the horizontal projection of the surface point normal vector and the radial direction is given. The threshold is the outward threshold.

7. The semantic segmentation method for the excavated area according to claim 1, characterized in that, In step S20, the occlusion detection result is obtained through ray detection: from surface points Along its normal vector A ray is emitted in a certain direction; if the ray is within a preset detection distance... If the surface point intersects with other colliding bodies, it is determined that the surface point is in an occluded state; If they do not intersect, then the surface points are determined not to be in an occluded state; occlusion state Represented as: ,in, This indicates that the ray detection hit other colliders; The correction of the exploitability score based on the occlusion detection results includes: adjusting the occlusion status according to... Determine the revised exploitability score; for unoccluded points, the revised score remains unchanged. For occluded points, the corrected score is reduced to [missing value]. .

8. The semantic segmentation method for the excavated area according to claim 1, characterized in that, In step S30, the step of calculating the proportion of excavable points in a local area based on the corrected excavability score and generating a blast excavation area label based on the proportion of excavable points includes: calculating that the corrected excavability score in the local area is greater than or equal to a score threshold. The surface point ratio is used to obtain the excavable point ratio. ,like ≥ Then, the local area is marked as the mining area; where, The threshold value is used for the proportion of the excavated area. The labels for the blasting excavation area include at least one of point-level labels, stone object-level labels, and local area-level labels.

9. The semantic segmentation method for the excavated area according to claim 1, characterized in that, In step S40, the physical prior features include relative height. Normalized radial distance and slope characteristics At least one of the following; wherein, relative height The radial distance is obtained by normalizing the height coordinates of the surface points; The slope characteristics are obtained by normalizing the distance from the surface point to the center of the pile in the horizontal plane. It is obtained from the vertical component of the surface point normal vector; The physical perception feature encoder includes a multilayer perceptron for mapping the physical prior features to the same feature dimension as the point cloud embedded features; the point cloud semantic segmentation network is a point cloud encoding and decoding network based on Point Transformer. The process of fusing the physical coding features with the point cloud embedding features of the point cloud semantic segmentation network includes: multiplying the physical coding features by a learnable scaling factor and then adding the result to the point cloud embedding features.

10. An electronic device comprising a processor and a memory, the memory storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the semantic segmentation method for the blasting excavation area as described in any one of claims 1 to 9.

Images