Coal bunker coal pile volume calculation method based on point cloud semantic segmentation

Abstract

The present application belongs to the technical field of coal bunker coal pile volume calculation, and specifically discloses a coal bunker coal pile volume calculation method based on point cloud semantic segmentation, and the specific technical scheme is: step one, data acquisition and pretreatment: firstly, three-dimensional point cloud data in the coal bunker is acquired through laser radar, and original point cloud is cut and sampled and features are constructed; step two, network structure is constructed: based on PointNet++ multi-scale feature abstraction structure and by introducing a boundary perception Transformer module, fine semantic segmentation of the coal pile and background structure is realized; step three, point cloud post-processing and volume reconstruction: the coal pile point cloud obtained through segmentation is post-processed and three-dimensionally reconstructed, a closed grid model is generated and coal pile volume calculation is completed, the method can stably realize high-precision volume estimation under different coal pile scales, has good engineering application potential, and can meet the needs of real-time deployment and resource-limited deployment in industrial fields.

Description

Technical Field

[0001] This invention belongs to the field of coal bunker volume calculation technology, specifically relating to a coal bunker volume calculation method based on point cloud semantic segmentation. Background Technology

[0002] Coal, as a vital global energy source and industrial raw material, requires efficient and safe storage and management to ensure energy supply and industrial production. Coal bunkers, as core storage facilities, rely heavily on accurate perception of the volume, shape, and spatial distribution of coal piles for intelligent management. Domestic scholars have conducted extensive research on coal pile volume measurement. Dong Lihong et al. proposed a dynamic coal quantity measurement method based on binocular vision, acquiring three-dimensional information of the coal pile through binocular images and combining image segmentation and stereo matching algorithms to estimate the volume and mass of the coal pile, providing an effective solution for low-cost, non-contact, real-time coal quantity measurement in underground coal mines. Chen Ying et al., based on the principle of structured light, used projection fringes and image reconstruction methods to achieve three-dimensional reconstruction and volume calculation of large material piles, applicable to irregular objects such as coal piles and sand and gravel piles. Yang Yitie et al. proposed a three-dimensional reconstruction system for accumulated materials based on multi-line structured light, generating point cloud data through laser triangulation to calculate the volume of the accumulated materials.

[0003] Although the above methods have achieved coal pile volume measurement to some extent, they mostly rely on controlled lighting environment or the assumption of regular pile body. In actual coal bunker scenarios, due to factors such as dust interference, complex lighting and shading, vision-based or structured light-based methods are difficult to operate stably for a long time. At the same time, traditional methods have limited ability to characterize complex and irregular coal pile shapes, and their automation level and measurement accuracy are still insufficient, making it difficult to meet the application requirements of high-precision and automated sensing of coal pile volume in coal bunkers.

[0004] In recent years, with the development of LiDAR and 3D sensing technologies, 3D point clouds have become an important data form for environmental perception and scene understanding due to their rich spatial structure and geometric information. Point cloud semantic segmentation aims to assign semantic labels to each 3D point, enabling fine-grained analysis of complex 3D scenes. This allows for automatic and accurate 3D segmentation of coal piles and background structures (such as bin walls and equipment) within coal bunkers, providing a foundation for volume calculation and status monitoring. Early research primarily relied on handcrafted features (such as normals, curvature, and geometric descriptors) and traditional classifiers (such as SVM and random forests) for point-level recognition, but these methods exhibited poor robustness in complex scenes. In recent years, the introduction of deep learning methods has significantly improved the performance of point cloud segmentation. The PointNet series of networks pioneered a neural network structure for directly processing unordered point clouds. PointNet achieves invariance learning of set features by independently encoding each point and using max pooling, effectively avoiding information loss during point cloud serialization. However, this method neglects local neighborhood structural information and struggles to capture spatial contextual relationships. PointNet++ builds upon this foundation by proposing a multi-scale grouping mechanism and a hierarchical feature abstraction structure. This design enables the extraction of neighborhood geometric features at different scales, thus better adapting to non-uniformly sampled point clouds and significantly improving the ability to perceive complex structures. Subsequently, numerous studies have expanded upon this framework. DGCNN utilizes EdgeConv to dynamically construct local graph structures to enhance geometric relationship learning; KPConv introduces learnable kernel points to achieve continuous spatial convolution operations, achieving excellent performance on high-density point clouds. These methods also demonstrate superior performance on general datasets such as S3DIS and ScanNet.

[0005] However, point cloud segmentation in coal bunkers still faces many challenges: the internal environment of coal bunkers is complex, the point cloud density is uneven and the noise interference is strong, and the boundary between the coal pile and the background is blurred; the shape of the coal pile is irregular and the dynamic industrial scene has high requirements for model efficiency and robustness. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention proposes a method for calculating the volume of coal piles in coal bunkers based on point cloud semantic segmentation, which realizes the construction of a three-dimensional model of the coal pile and accurate volume calculation.

[0007] To achieve the above objectives, the technical solution adopted by this invention is: a method for calculating the volume of coal piles in coal bunkers based on point cloud semantic segmentation, the specific steps of which are as follows:

[0008] Step 1: Data Acquisition and Preprocessing: First, acquire 3D point cloud data of the coal bunker using lidar, and then perform block sampling and feature construction on the original point cloud;

[0009] Step 2: Constructing the network structure: Based on the PointNet++ multi-scale feature abstraction structure and introducing the boundary-aware Transformer module, fine semantic segmentation of the coal pile and background structure is achieved;

[0010] Step 3: Point cloud post-processing and volume reconstruction: Post-process and 3D reconstruction of the segmented coal pile point cloud are performed to generate a closed mesh model and complete the coal pile volume calculation.

[0011] In step one, the specific steps for data acquisition are as follows:

[0012] The three-dimensional spatial data inside the coal bunker was acquired using lidar scanning. LiDAR actively emits laser pulses and receives reflected signals, enabling it to accurately acquire the spatial coordinates of the target surface without relying on external lighting conditions. By scanning the interior of the coal bunker from multiple angles, high-density three-dimensional point cloud data covering the entire coal pile and bunker wall structure can be obtained, providing a reliable data foundation for subsequent semantic segmentation and volume calculation. The basic format of the point cloud is as follows: ,in, Represents the horizontal coordinate. The vertical height coordinates are represented by the label, which is a manually labeled category label used to distinguish coal piles from non-coal structures such as coal bunker walls. Due to the complex internal environment of coal bunkers, including issues such as coal dust obstruction, uneven surface reflection, and structural occlusion, 3D point clouds can directly reflect the true geometric shape and spatial distribution characteristics of the coal pile, are unaffected by changes in lighting, and have stronger environmental adaptability. Therefore, in the data processing and feature construction stages, feature representation and analysis are performed based on the geometric coordinate features of the point cloud, thereby improving the method's versatility and robustness in actual coal bunker scenarios.

[0013] Data Processing: To alleviate the GPU and memory pressure caused by large-scale coal bunker point clouds during model training, and to enhance the network's ability to model local geometric structures, a sliding window-based point cloud dicing strategy is adopted to preprocess the raw point cloud data. Specifically, only in the coal bunker... A fixed-size two-dimensional window is set in a horizontal plane and slides according to a given step size to generate a series of overlapping local point cloud patches. Let the window size be... The sliding step size is Then the local point cloud corresponding to the (m,n)th window can be represented as: (1);

[0014] in, , , Indicates the first line, number The set of points within the corresponding sliding window of the column. This represents the set of all points in the original point cloud data. Represents a set of point clouds The first in One point, Point exist Coordinates in direction, Point exist Coordinates in direction, Indicates the first A window in Starting coordinates in the direction, Indicates the first A window in Starting coordinates in the direction, Indicates the current point cloud scene in Minimum coordinate value in the direction, Indicates the current point cloud scene in The minimum coordinate value in the direction.

[0015] This strategy effectively mitigates the impact of uneven point cloud density on feature learning while ensuring complete coverage of the coal bunker scene, facilitating stable extraction of local geometric features by the model. During the block segmentation process, only local point clouds with more than a preset threshold of points within a block are retained to avoid interference from overly sparse regions during model training. For point cloud blocks that meet the conditions, a random sampling strategy is used to select a fixed number of points as network input; when the number of points within a block is insufficient, random resampling is used to fill in the gaps, thereby ensuring the consistency of the number of input points and meeting the requirements of deep networks for fixed-dimensional input. To improve the model's generalization ability in complex coal bunker scenes, various data augmentation operations are applied to local point cloud blocks during the training phase, including random rotation around the vertical axis, random scaling, and Gaussian noise perturbation, specifically as follows:

[0016] (2);

[0017] in, , , Represents the transformed th The coordinates of the points This represents the scaling factor, used to control the overall scale variation of the point cloud in space. Indicates circling Axis rotation angle The rotation matrix, Indicates dotted cloud The rotation angle of the shaft, This represents the perturbation term added to the point coordinates, typically used to simulate noise or random offsets.

[0018] In terms of feature construction, the original point cloud coordinates were not directly used as network input. Instead, a more discriminative feature representation was constructed by combining the absolute position and relative geometric relationship of the points. Specifically, for each local point cloud block, the centroid of the point cloud was first calculated, and the point coordinates were centered to highlight the local spatial structure features; at the same time, the point coordinates were linearly normalized to... The interval is used to preserve the relative spatial distribution information of points within the current window. Finally, the centered coordinates and normalized coordinates are concatenated to form a 6-dimensional geometric feature input network, thereby enhancing the model's ability to express local structural differences while ensuring geometric consistency.

[0019] During the testing phase, a sliding window segmentation strategy was also employed to cover the entire coal bunker scene, ensuring consistency in data distribution between training and testing. Unlike the training phase, the testing process required accurately mapping local prediction results back to the global point cloud coordinate system. Therefore, during the segmentation process, the global index of each local point in the original point cloud was recorded. When the same spatial region is covered by multiple sliding windows, the corresponding point will receive multiple prediction results. To address this, a majority voting strategy was used to fuse the predicted labels of global points. This involved statistically analyzing the prediction results of the point across all windows and selecting the category with the highest frequency as the final semantic label, effectively improving the smoothness and consistency of the full-scene segmentation results. Regarding point cloud block construction and feature construction, the testing and training phases remained consistent, both employing a fixed-point sampling strategy and a 6D geometric feature input format consisting of a concatenation of centralized and normalized coordinates. This consistent design helps reduce the training-test distribution offset, further improving the model's segmentation stability and robustness in real coal bunker scenes.

[0020] In step two, the overall network framework is based on PointNet++, with a boundary-aware Transformer module introduced at key locations. The network mainly consists of three parts: a multi-scale feature abstraction layer, a boundary-aware Transformer layer, and a feature propagation and point-by-point classification layer. First, the multi-scale feature abstraction module abstracts point cloud features within local neighborhoods at different scales through hierarchical sampling and grouping strategies. This allows for the simultaneous capture of detailed structural details and overall geometric morphology of the coal pile surface, effectively mitigating the impact of uneven point cloud density and scale variations in the coal bunker. Building upon this, the boundary-aware Transformer module is introduced to further model local features. This module uses a self-attention mechanism to characterize the relationship between points and their neighborhoods and combines prior boundary information to enhance the feature representation capability of the transition region between the coal pile and the bunker wall. This enables the network to pay more attention to key geometric changes in complex boundary regions, thereby improving the accuracy and robustness of semantic segmentation. Finally, the point cloud features extracted from high-level features are upsampled back to the original point cloud resolution layer by layer through the feature propagation module and fused with low-level geometric features. A point-by-point classifier is then used to perform semantic prediction for each point, achieving refined segmentation of the coal pile and the coal bunker wall. The overall network ensures segmentation accuracy while also adapting to complex coal bunker scenarios, providing a reliable semantic foundation for subsequent 3D reconstruction and volume calculation of coal piles.

[0021] Multi-scale feature extraction layer: Let the point cloud be... in, Indicates the first The 3D coordinates of each point. The input feature for each point is... ,in, Centralized coordinates are used to describe the spatial offset of a point relative to the centroid of the current point cloud block. To normalize the coordinates, the point cloud coordinates are linearly normalized and mapped to... The interval is used to preserve the relative positional distribution of points within a local block. Based on this, PointNet++ samples the points... field Within, with multi-scale radius Feature aggregation is performed to enhance the network's ability to model geometric structures at different scales. For each scale... The calculation form for local features is as follows:

[0022] (3);

[0023] in, Represents points within the neighborhood. This is a shared multilayer perceptron network. Relative features are introduced. and This process effectively eliminates the influence of absolute coordinates, allowing the network to focus more on the local geometric relationships between points. Max pooling, as a symmetric function, ensures the invariance of the point cloud's disorder during feature aggregation. Finally, local features extracted at different scales are concatenated and fused to form sampling points. Multi-scale feature representation:

[0024] (4);

[0025] This multi-scale feature abstraction and fusion strategy enables the network to simultaneously capture the detailed structure of the coal pile surface (small-scale neighborhood) and the overall morphology and spatial distribution features of the coal pile (large-scale neighborhood). It effectively addresses the common problems in coal bunker point clouds, such as uneven point density, significant scale changes, and complex structures, and provides high-quality feature representations for subsequent boundary-aware feature enhancement and semantic segmentation.

[0026] Boundary-Aware Point Transformer (BPT) Layer: In traditional Point Transformers, local attention weights are mainly determined by the query-key similarity between point pairs. This mechanism can effectively model local contextual relationships in point cloud semantic segmentation tasks. However, in coal bunker point cloud scenarios, the geometric transition region between the coal pile and the bunker wall often exhibits slow curvature changes, similar normals, and uneven point density. This makes it difficult for attention mechanisms that rely solely on feature similarity to accurately distinguish semantic boundaries, easily leading to missegmentation in the coal-wall interface region.

[0027] The Boundary Aware Point Transformer (BPT) module introduces boundary prior constraints to model the coal-wall transition region and incorporates these constraints into the attention weight calculation process, thereby enhancing the model's ability to model features of key boundary regions. The BPT module mainly consists of two parts: boundary prior prediction and boundary enhancement attention mechanism.

[0028] Boundary prior prediction: To characterize the probability that a point lies in the transition region between the coal pile and the coal bunker wall, a lightweight EdgeConv module is introduced to model local geometric changes and predict the boundary probability for each point. Specifically, for point... In its local neighborhood Inside, EdgeConv is used to aggregate the feature differences between points:

[0029] (5);

[0030] in, Indicates a structure composed of shared multilayer sensing mechanisms Mapping function, and Points With neighboring points The feature representation of . Based on this, through linear mapping and Sigmoid activation function prediction points Boundary probabilities:

[0031] (6);

[0032] in, For learnable weight matrix, Indicates the first The input feature vector of each sample, This represents the Sigmoid activation function. When a point is located in the transition zone between the coal pile and the coal bunker wall, the changes in its neighborhood characteristics are more significant, corresponding to... The value is relatively large; while in the area inside the coal pile or the inner wall of the silo, The value tends to be small. This boundary prior provides semantic boundary cues for the subsequent attention mechanism, guiding the model to focus on potential coal-wall boundary regions.

[0033] Boundary Enhanced Attention: Local Domain in Point Transformer Within, the original attention logits are:

[0034] (7);

[0035] in, For point The query indicates that, For point The key indicates, For feature dimensions.

[0036] BPT introduces boundary gating on this basis:

[0037] (8);

[0038] (9);

[0039] like If both are boundary points, then A larger value results in higher attention weights after softmax, thereby improving the discriminative power of boundary region features. Indicates the first When a central point performs feature aggregation on its neighboring points, the first... The normalized attention weights corresponding to each neighboring point This represents the natural exponential function, used to map correlation scores to positive values. Indicates the center point The first in its neighborhood Unnormalized correlation scores between points.

[0040] Finally, point The update features are:

[0041] (10);

[0042] in, For point The value vector of this mechanism enables the preservation of the Transformer's expressive power in normal regions while significantly enhancing attention to geometric details in the boundary transition zone, even in fuzzy boundary regions.

[0043] Feature propagation and point-by-point classification: through a multi-scale feature abstraction layer and a boundary-aware Point Transformer layer The spatial resolution of the point cloud features is reduced layer by layer after extraction and enhancement. In order to achieve point-by-point semantic prediction of the original point cloud, this invention adopts a feature propagation mechanism to upsample and fuse high-level abstract features layer by layer, and finally restore the original point cloud resolution.

[0044] (11);

[0045] This formula indicates that during the feature propagation process, the first... The first in the layer Features of each point From the higher-level semantic features Features of the current layer neighborhood The points are obtained through joint interpolation and fusion. Through interpolation, the network can progressively propagate deep abstract semantic information back to the original point cloud resolution while preserving shallow local geometric details, thereby improving the spatial accuracy of the point cloud semantic segmentation results. The points are then concatenated and fused with lower-level features, and the classification head uses a pointwise MLP.

[0046] (12);

[0047] This formula represents the classification prediction of the features of the i-th point. First, the point features are analyzed using the weight matrix W and the bias term b. A linear mapping is performed to obtain the response values ​​of the point in different semantic categories; then, the softmax function is used to normalize the response values ​​to obtain the probability distribution of point i belonging to each category. Finally, the semantic category of a point can be determined based on the maximum probability, and the semantic label (coal / coal bunker wall) of each point can be output.

[0048] In step three, prediction result fusion: During the testing phase, considering the large scale and uneven density of the coal bunker point cloud, this invention employs a sliding window segmentation method to segment the entire scene for prediction. The point cloud within each sliding window block is independently semantically classified by the network. However, due to overlapping areas between sliding windows, the prediction result of a single window only contains local contextual information, easily leading to segmentation discontinuities or label jitter at block boundaries. Therefore, this invention introduces a prediction result fusion strategy based on majority voting to improve the consistency and stability of the semantic segmentation results across the entire scene.

[0049] Sliding window block-level prediction: Within each local block divided by the sliding window, the network classifies the input point cloud point by point and outputs a semantic prediction label for each point.

[0050] (13);

[0051] in, Point In the The prediction results in each sliding window block are labeled 1 to represent the coal pile and 0 to represent the coal bunker wall. Since the sliding window slides with a fixed step size, the same global point is often included in multiple adjacent sliding window blocks, thus obtaining multiple independent prediction results.

[0052] Full-scene prediction result fusion: For each global point i, collect its predicted label set in all sliding window blocks containing that point:

[0053] (14);

[0054] in, Indicates the containing point The number of sliding window blocks.

[0055] This invention employs a majority voting strategy to determine the final semantic label of point i:

[0056] (15);

[0057] This formula represents a point cloud prediction result fusion method based on majority voting. Since the point cloud is segmented and predicted using a sliding window method during the testing phase, the same point may appear in multiple overlapping blocks, thus resulting in multiple prediction categories. For the indicator function, for the first Each point, statistically analyzed in In this prediction, they were classified into various categories. The number of occurrences is determined, and the category with the highest frequency is selected as the final prediction result. This method can reduce the random error caused by a single local prediction and improve the stability of global point cloud semantic segmentation results.

[0058] Enhanced consistency in boundary regions: The majority voting fusion mechanism has significant advantages in overlapping sliding window regions. Since boundary points are typically observed repeatedly by multiple sliding windows from different spatial locations and contexts, this strategy effectively suppresses the impact of single sliding window prediction errors, reduces the instability of boundary region labels, and significantly improves the segmentation continuity and overall consistency at the interface between the coal pile and the coal bunker wall. Ultimately, this fusion strategy yields a complete, seamless, and globally consistent semantic segmentation result for the coal bunker point cloud, providing reliable input data for subsequent coal pile surface reconstruction and volume calculation.

[0059] 3D Reconstruction and Geometric Calculation: Based on the semantic segmentation results of point clouds, this invention further performs 3D geometric reconstruction on point clouds identified as coal piles to construct a continuous and closed coal pile surface model, providing a geometric basis for subsequent volume calculations. First, point sets semantically labeled as coal piles are extracted, and coal bunker walls and other background points are removed to reduce interference from non-target areas in the reconstruction process. To obtain a smooth and topologically continuous coal pile surface, this invention employs the Poisson surface reconstruction method for 3D reconstruction of the coal pile point clouds. This method uses the normal information of the point cloud as a constraint, solves the Poisson equation to convert the discrete point cloud into a continuous implicit function, and generates a closed triangular mesh model based on this. Compared to reconstruction methods based on local stitching, Poisson reconstruction has better stability in handling point cloud sparsity, uneven sampling, and noise interference, effectively ensuring the overall smoothness and closure of the reconstructed surface, making it suitable for complex industrial scenarios such as coal bunkers. Before performing Poisson reconstruction, the normals of the coal pile point cloud are estimated. The normal directions are obtained by analyzing the local neighborhood geometry of each point. Principal component analysis based on k-nearest neighbors or fixed radius neighborhoods is used to calculate the local surface normals. Accurate normal estimation helps to enhance the model's ability to express the geometry of the coal pile surface and is an important prerequisite for ensuring the quality of Poisson reconstruction. After completing the surface reconstruction, a closed triangular mesh model composed of a large number of triangular patches is obtained.

[0060] This invention calculates the volume of the triangular mesh based on the divergence theorem, transforming the complex volume integral into the local volume decomposition and accumulation of mesh patches. For the volume integral derived from the vertices... The directed volume of the triangular facet formed can be expressed as:

[0061] (16);

[0062] The total volume estimate of the coal pile can be obtained by summing the volume contributions of all triangular faces in the closed grid.

[0063] Compared with existing technologies, the specific advantages of this invention are as follows: This invention proposes a semantic segmentation method based on PointNet++ and boundary-aware Transformer. The PointNet++ layer is used to achieve multi-scale local feature abstraction, ensuring full perception of coal pile details and geometric structures at different scales. The boundary-aware Transformer layer introduces boundary priors into the standard attention mechanism, enabling the network to automatically enhance its response in the coal-wall transition zone, effectively enhancing the modeling ability of key points and neighborhood features, thereby improving the segmentation accuracy and robustness of complex coal bunker scenes. In the post-processing stage, voxel downsampling, statistical filtering denoising, normal estimation, and Poisson reconstruction are combined to generate closed triangular meshes, and the coal pile volume is further calculated to achieve quantitative analysis of the segmentation results. Simultaneously, three-dimensional perception devices such as LiDAR are deployed in a real coal bunker environment to systematically collect data on the internal scene of the coal bunker, obtaining three-dimensional point cloud data with realistic geometric features and complex spatial morphology, providing a reliable data foundation for subsequent data preprocessing, model training, and performance verification. Attached Figure Description

[0064] Figure 1 A flowchart of point cloud segmentation and volume calculation technology for coal bunkers.

[0065] Figure 2 This is a diagram of a point cloud segmentation network framework based on PointNet++ and boundary-aware Transformer.

[0066] Figure 3 This is a diagram of the boundary sensing layer structure.

[0067] Figure 4 This is a 3D point cloud map of the coal bunker.

[0068] Figure 5 This is a comparison chart of training losses for different methods.

[0069] Figure 6 A graph comparing the accuracy of different testing methods.

[0070] Figure 7 Comparison of semantic segmentation results of coal bunker coal pile point clouds under different network models. Figure 7 (a) is a visualization result of the segmentation of the first type of coal bunker point cloud. Figure 7 (b) is a visualization result of the segmentation of the second type of coal bunker point cloud.

[0071] Figure 8 This is a comparison image of the 3D reconstruction results. Figure 8 (a) is a three-dimensional reconstruction result of the first type of coal pile. Figure 8 (b) is a three-dimensional reconstruction result of the second type of coal pile. Detailed Implementation

[0072] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0073] The method for calculating the volume of coal piles in coal bunkers based on point cloud semantic segmentation aims to achieve automatic and accurate differentiation between coal piles and coal bunker walls, and to provide a reliable basis for quantitative calculation of coal pile volume. The overall method mainly includes three parts: data acquisition and preprocessing, network structure design, and point cloud post-processing and volume reconstruction.

[0074] like Figure 1 As shown, firstly, three-dimensional point cloud data inside the coal bunker is acquired using LiDAR, and the original point cloud is segmented for sampling and feature construction. Secondly, based on the PointNet++ multi-scale feature abstraction structure and introducing the boundary-aware Transformer module, fine semantic segmentation of the coal pile and background structure is achieved. Finally, the segmented coal pile point cloud is post-processed and three-dimensionally reconstructed to generate a closed mesh model and complete the coal pile volume calculation, thus constructing a complete technical process from point cloud semantic segmentation to volume quantization analysis.

[0075] Step 1: Data Acquisition and Preprocessing: First, acquire 3D point cloud data of the coal bunker using lidar, and then perform block sampling and feature construction on the original point cloud;

[0076] The specific steps for data acquisition are as follows: Three-dimensional spatial data of the coal bunker's interior is acquired using lidar scanning. LiDAR actively emits laser pulses and receives reflected signals, enabling precise acquisition of the target surface's spatial coordinates without relying on external lighting conditions. By scanning the coal bunker's interior from multiple angles, high-density three-dimensional point cloud data covering the entire coal pile and bunker wall structure can be obtained, providing a reliable data foundation for subsequent semantic segmentation and volume calculation. The basic format of the point cloud is as follows: ,in, Represents the horizontal coordinate. The vertical height coordinates are represented by the label, which is a manually labeled category label used to distinguish coal piles from non-coal structures such as coal bunker walls. Due to the complex internal environment of coal bunkers, including issues such as coal dust obstruction, uneven surface reflection, and structural occlusion, 3D point clouds can directly reflect the true geometric shape and spatial distribution characteristics of the coal pile, are unaffected by changes in lighting, and have stronger environmental adaptability. Therefore, in the data processing and feature construction stages, feature representation and analysis are performed based on the geometric coordinate features of the point cloud, thereby improving the method's versatility and robustness in actual coal bunker scenarios.

[0077] Data Processing: To alleviate the GPU and memory pressure caused by large-scale coal bunker point clouds during model training, and to enhance the network's ability to model local geometric structures, a sliding window-based point cloud dicing strategy is adopted to preprocess the raw point cloud data. Specifically, only in the coal bunker... A fixed-size two-dimensional window is set in a horizontal plane and slides according to a given step size to generate a series of overlapping local point cloud patches. Let the window size be... The sliding step size is Then the first The local point cloud corresponding to each window can be represented as: (1);

[0078] in, , , Indicates the first line, number The set of points within the corresponding sliding window of the column. This represents the set of all points in the original point cloud data. Represents a set of point clouds The first in One point, Point exist Coordinates in direction, Point exist Coordinates in direction, Indicates the first A window in Starting coordinates in the direction, Indicates the first A window in Starting coordinates in the direction, Indicates the current point cloud scene in Minimum coordinate value in the direction, Indicates the current point cloud scene in The minimum coordinate value in the direction.

[0079] This strategy effectively mitigates the impact of uneven point cloud density on feature learning while ensuring complete coverage of the coal bunker scene, facilitating stable extraction of local geometric features by the model. During the block segmentation process, only local point clouds with more than a preset threshold of points within a block are retained to avoid interference from overly sparse regions during model training. For point cloud blocks that meet the conditions, a random sampling strategy is used to select a fixed number of points as network input; when the number of points within a block is insufficient, random resampling is used to fill in the gaps, thereby ensuring the consistency of the number of input points and meeting the requirements of deep networks for fixed-dimensional input. To improve the model's generalization ability in complex coal bunker scenes, various data augmentation operations are applied to local point cloud blocks during the training phase, including random rotation around the vertical axis, random scaling, and Gaussian noise perturbation, specifically as follows:

[0080] (2);

[0081] in, , , Represents the transformed th The coordinates of the points This represents the scaling factor, used to control the overall scale variation of the point cloud in space. Indicates circling Axis rotation angle The rotation matrix, Indicates dotted cloud The rotation angle of the shaft, This represents the perturbation term added to the point coordinates, typically used to simulate noise or random offsets.

[0082] In terms of feature construction, the original point cloud coordinates were not directly used as network input. Instead, a more discriminative feature representation was constructed by combining the absolute position and relative geometric relationship of the points. Specifically, for each local point cloud block, the centroid of the point cloud was first calculated, and the point coordinates were centered to highlight the local spatial structure features; at the same time, the point coordinates were linearly normalized to... The interval is used to preserve the relative spatial distribution information of points within the current window. Finally, the centered coordinates and normalized coordinates are concatenated to form a 6-dimensional geometric feature input network, thereby enhancing the model's ability to express local structural differences while ensuring geometric consistency.

[0083] During the testing phase, a sliding window segmentation strategy was also employed to cover the entire coal bunker scene, ensuring consistency in data distribution between training and testing. Unlike the training phase, the testing process required accurately mapping local prediction results back to the global point cloud coordinate system. Therefore, during the segmentation process, the global index of each local point in the original point cloud was recorded. When the same spatial region is covered by multiple sliding windows, the corresponding point will receive multiple prediction results. To address this, a majority voting strategy was used to fuse the predicted labels of global points. This involved statistically analyzing the prediction results of the point across all windows and selecting the category with the highest frequency as the final semantic label, effectively improving the smoothness and consistency of the full-scene segmentation results. Regarding point cloud block construction and feature construction, the testing and training phases remained consistent, both employing a fixed-point sampling strategy and a 6D geometric feature input format consisting of a concatenation of centralized and normalized coordinates. This consistent design helps reduce the training-test distribution offset, further improving the model's segmentation stability and robustness in real coal bunker scenes.

[0084] Step 2: Constructing the network structure: Based on the PointNet++ multi-scale feature abstraction structure and introducing the boundary-aware Transformer module, fine semantic segmentation of the coal pile and background structure is achieved.

[0085] The overall network framework is based on PointNet++, and boundary-aware Transformer modules are introduced at key locations. The overall structure is as follows: Figure 2 As shown, the network mainly consists of three parts: a multi-scale feature abstraction layer, a boundary-aware Transformer layer, and a feature propagation and point-by-point classification layer. First, the multi-scale feature abstraction module abstracts point cloud features within local neighborhoods at different scales through hierarchical sampling and grouping strategies. This allows for the simultaneous capture of detailed structural details and overall geometric morphology of the coal pile surface, effectively mitigating the impact of uneven point cloud density and scale variations in the coal bunker. Building upon this, a boundary-aware Transformer module is introduced to further model local features. This module uses a self-attention mechanism to characterize the relationship between points and their neighborhoods and combines prior boundary information to enhance the feature representation capability of the transition region between the coal pile and the bunker wall. This enables the network to pay more attention to key geometric changes in complex boundary regions, thereby improving the accuracy and robustness of semantic segmentation. Finally, the point cloud features extracted from the high-level features are upsampled back to the original point cloud resolution layer by layer through the feature propagation module and fused with low-level geometric features. A point-by-point classifier is then used to perform semantic prediction for each point, achieving refined segmentation of the coal pile and the coal bunker wall. The overall network ensures segmentation accuracy while also adapting to complex coal bunker scenarios, providing a reliable semantic foundation for subsequent 3D reconstruction and volume calculation of coal piles.

[0086] Multi-scale feature extraction layer: Let the point cloud be... in, Indicates the first The 3D coordinates of each point. The input feature for each point is... ,in, Centralized coordinates are used to describe the spatial offset of a point relative to the centroid of the current point cloud block. To normalize the coordinates, the point cloud coordinates are linearly normalized and mapped to... The interval is used to preserve the relative positional distribution of points within a local block. Based on this, PointNet++ samples the points... field Within, with multi-scale radius Feature aggregation is performed to enhance the network's ability to model geometric structures at different scales. For each scale... The calculation form for local features is as follows:

[0087] (3);

[0088] in, Represents points within the neighborhood. This is a shared multilayer perceptron network. Relative features are introduced. and This process effectively eliminates the influence of absolute coordinates, allowing the network to focus more on the local geometric relationships between points. Max pooling, as a symmetric function, ensures the invariance of the point cloud's disorder during feature aggregation. Finally, the local features extracted at different scales are concatenated and fused to form a multi-scale feature representation of sampling point p.

[0089] (4);

[0090] This multi-scale feature abstraction and fusion strategy enables the network to simultaneously capture the detailed structure of the coal pile surface (small-scale neighborhood) and the overall morphology and spatial distribution features of the coal pile (large-scale neighborhood). It effectively addresses the common problems in coal bunker point clouds, such as uneven point density, significant scale changes, and complex structures, and provides high-quality feature representations for subsequent boundary-aware feature enhancement and semantic segmentation.

[0091] Boundary-Aware Point Transformer (BPT) Layer: In traditional Point Transformers, local attention weights are mainly determined by the query-key similarity between point pairs. This mechanism can effectively model local contextual relationships in point cloud semantic segmentation tasks. However, in coal bunker point cloud scenarios, the geometric transition region between the coal pile and the bunker wall often exhibits slow curvature changes, similar normals, and uneven point density. This makes it difficult for attention mechanisms that rely solely on feature similarity to accurately distinguish semantic boundaries, easily leading to missegmentation in the coal-wall interface region.

[0092] The Boundary Aware Point Transformer (BPT) module introduces boundary prior constraints to model the coal-wall transition region and incorporates these constraints into the attention weight calculation process, thereby enhancing the model's ability to model features in key boundary regions. The BPT module mainly consists of two parts: boundary prior prediction and a boundary enhancement attention mechanism. Its overall structure is as follows: Figure 3 As shown.

[0093] Boundary Prior Prediction: To characterize the probability that a point lies in the transition region between the coal pile and the coal bunker wall, this invention introduces the lightweight EdgeConv module to model local geometric changes and predict the boundary probability of each point. Specifically, for point... In its local neighborhood Inside, EdgeConv is used to aggregate the feature differences between points:

[0094] (5);

[0095] in, Indicates a structure composed of shared multilayer sensing mechanisms Mapping function, and Points With neighboring points The feature representation of . Based on this, through linear mapping and Sigmoid activation function prediction points Boundary probabilities:

[0096] (6);

[0097] in, For learnable weight matrix, Indicates the first The input feature vector of each sample, This represents the Sigmoid activation function. When a point is located in the transition zone between the coal pile and the coal bunker wall, the changes in its neighborhood characteristics are more significant, corresponding to... The value is relatively large; while in the area inside the coal pile or the inner wall of the silo, The value tends to be small. This boundary prior provides semantic boundary cues for the subsequent attention mechanism, guiding the model to focus on potential coal-wall boundary regions.

[0098] Boundary Enhanced Attention: Local Domain in Point Transformer Within, the original attention logits are:

[0099] (7);

[0100] in, For point The query indicates that, For point The key indicates, For feature dimensions.

[0101] BPT introduces boundary gating on this basis:

[0102] (8);

[0103] (9);

[0104] Indicates the first When a central point performs feature aggregation on its neighboring points, the first... The normalized attention weights corresponding to each neighboring point This represents the natural exponential function, used to map correlation scores to positive values. Indicates the center point The first in its neighborhood The unnormalized correlation score between points; if If both are boundary points, then The larger the value, the higher the attention weight it receives after softmax, thus improving the discriminative power of boundary region features; ultimately, the point... The update features are:

[0105] (10);

[0106] in, For point The value vector of this mechanism enables the preservation of the Transformer's expressive power in normal regions while significantly enhancing attention to geometric details in the boundary transition zone, even in fuzzy boundary regions.

[0107] Feature propagation and point-by-point classification: After point cloud features are extracted and enhanced layer by layer through multi-scale feature abstraction layer and boundary-aware Point Transformer layer (BPT), their spatial resolution decreases layer by layer. In order to achieve point-by-point semantic prediction of the original point cloud, this invention adopts a feature propagation mechanism to upsample and fuse high-level abstract features layer by layer, and finally restore the original point cloud resolution.

[0108] (11);

[0109] This formula indicates that during the feature propagation process, the first... The first in the layer Features of each point From the higher-level semantic features Features of the current layer neighborhood The network achieves this through joint interpolation and fusion. Through interpolation, the network can progressively propagate deep abstract semantic information back to the original point cloud resolution while preserving shallow local geometric details, thereby improving the spatial accuracy of the point cloud semantic segmentation results. The network is then concatenated and fused with low-level features, and the classification head uses a pointwise MLP.

[0110] (12);

[0111] This expression represents the expression for the th... Classification and prediction are performed based on the features of each point, firstly through the weight matrix. and bias terms Point features A linear mapping is performed to obtain the response values ​​of the point in different semantic categories; then, the softmax function is used to normalize the response values ​​to obtain the probability distribution of point i belonging to each category. Finally, the semantic category of a point can be determined based on the maximum probability, and the semantic label of each point can be output.

[0112] Output the semantic label (coal / coal bunker wall) for each point.

[0113] Step 3: Point cloud post-processing and volume reconstruction: Post-process and 3D reconstruction of the segmented coal pile point cloud are performed to generate a closed mesh model and complete the coal pile volume calculation.

[0114] In step three, prediction result fusion: During the testing phase, considering the large scale and uneven density of the coal bunker point cloud, this invention employs a sliding window segmentation method to segment the entire scene for prediction. The point cloud within each sliding window block is independently semantically classified by the network. However, due to overlapping areas between sliding windows, the prediction result of a single window only contains local contextual information, easily leading to segmentation discontinuities or label jitter at block boundaries. Therefore, this invention introduces a prediction result fusion strategy based on majority voting to improve the consistency and stability of the semantic segmentation results across the entire scene.

[0115] Sliding window block-level prediction: Within each local block divided by the sliding window, the network classifies the input point cloud point by point and outputs a semantic prediction label for each point.

[0116] (13);

[0117] in, Point In the The prediction results in each sliding window block are labeled 1 to represent the coal pile and 0 to represent the coal bunker wall. Since the sliding window slides with a fixed step size, the same global point is often included in multiple adjacent sliding window blocks, thus obtaining multiple independent prediction results.

[0118] Full-scene prediction result fusion: for each global point Collect its set of predicted labels across all sliding window blocks containing that point:

[0119] (14);

[0120] in, Indicates the containing point The number of sliding window blocks.

[0121] This invention employs a majority voting strategy to determine the final semantic label of point i:

[0122] (15);

[0123] This formula represents a point cloud prediction result fusion method based on majority voting. Since the point cloud is segmented and predicted using a sliding window method during the testing phase, the same point may appear in multiple overlapping blocks, thus resulting in multiple prediction categories. For the indicator function, for the first Each point, statistically analyzed in The number of times each category c is identified in the prediction is used, and the category with the highest frequency is selected as the final prediction result. This method can reduce the random error caused by a single local prediction and improve the stability of global point cloud semantic segmentation results.

[0124] in, As an indicator function, if a certain category appears most frequently in the prediction results, it is used as the final semantic label for that point, thereby achieving effective integration of multi-window prediction information.

[0125] Enhanced consistency in boundary regions: The majority voting fusion mechanism has significant advantages in overlapping sliding window regions. Since boundary points are typically observed repeatedly by multiple sliding windows from different spatial locations and contexts, this strategy effectively suppresses the impact of single sliding window prediction errors, reduces the instability of boundary region labels, and significantly improves the segmentation continuity and overall consistency at the interface between the coal pile and the coal bunker wall. Ultimately, this fusion strategy yields a complete, seamless, and globally consistent semantic segmentation result for the coal bunker point cloud, providing reliable input data for subsequent coal pile surface reconstruction and volume calculation.

[0126] 3D Reconstruction and Geometric Calculation: Based on the semantic segmentation results of point clouds, this invention further performs 3D geometric reconstruction on point clouds identified as coal piles to construct a continuous and closed coal pile surface model, providing a geometric basis for subsequent volume calculations. First, point sets semantically labeled as coal piles are extracted, and coal bunker walls and other background points are removed to reduce interference from non-target areas in the reconstruction process. To obtain a smooth and topologically continuous coal pile surface, this invention employs the Poisson surface reconstruction method for 3D reconstruction of the coal pile point clouds. This method uses the normal information of the point cloud as a constraint, solves the Poisson equation to convert the discrete point cloud into a continuous implicit function, and generates a closed triangular mesh model based on this. Compared to reconstruction methods based on local stitching, Poisson reconstruction has better stability in handling point cloud sparsity, uneven sampling, and noise interference, effectively ensuring the overall smoothness and closure of the reconstructed surface, making it suitable for complex industrial scenarios such as coal bunkers. Before performing Poisson reconstruction, the normals of the coal pile point cloud are estimated. The normal directions are obtained by analyzing the local neighborhood geometry of each point. Principal component analysis based on k-nearest neighbors or fixed radius neighborhoods is used to calculate the local surface normals. Accurate normal estimation helps to enhance the model's ability to express the geometry of the coal pile surface and is an important prerequisite for ensuring the quality of Poisson reconstruction. After completing the surface reconstruction, a closed triangular mesh model composed of a large number of triangular patches is obtained.

[0127] This invention calculates the volume of the triangular mesh based on the divergence theorem, transforming the complex volume integral into the local volume decomposition and accumulation of mesh patches. For the volume calculated from the vertices... The directed volume of the triangular facet formed can be expressed as:

[0128] (16);

[0129] The total volume estimate of the coal pile can be obtained by summing the volume contributions of all triangular faces in the closed grid.

[0130] This invention deploys 3D sensing devices such as LiDAR in a real coal bunker environment to collect data on the internal scene of the coal bunker, constructing a coal bunker point cloud dataset for practical engineering applications. By rationally deploying sensor positions inside the coal bunker, a comprehensive scan of the coal pile surface and the coal bunker wall structure is achieved, acquiring 3D point cloud data with realistic geometric features and complex spatial morphology. The collected point cloud images are shown below. Figure 4 As shown. Data point cloud. The data is stored in a formatted manner, with tags including coal and coal bunker walls. The training set consists of 80% point cloud blocks, sliced ​​using a sliding window and randomly sampled 1024 points per block, with data augmentation (rotation, scaling, Gaussian noise) applied to each block. The validation set consists of 20% point cloud blocks, used for model parameter tuning and early stopping. The test set covers the entire coal bunker, sliced ​​using a sliding window, and uses majority voting fusion to ensure global consistency.

[0131] During the model training phase, this invention employs the Adam optimizer to update network parameters, with an initial learning rate set to 0.001. A weight decay coefficient is used to suppress overfitting. To further improve the model's convergence stability and performance in later training stages, a cosine annealing method is used for learning rate scheduling, allowing the learning rate to gradually decay during training. The training batch size is set to 8, balancing training stability and efficiency while ensuring sufficient GPU memory. The model is trained for 100 epochs, and validation set performance metrics are continuously monitored during training to evaluate the model's generalization ability.

[0132] To verify the effectiveness and advantages of the proposed method, this invention selected several classic point cloud segmentation networks for comparative experiments, including methods ranging from basic global feature learning to multi-scale local feature abstraction and graph convolution enhancement. Specifically: PointNet, as a pioneering method for point cloud processing, classifies and segments unordered point clouds by independently encoding each point and using max pooling to aggregate global features. Its advantages lie in its simple structure and high computational efficiency, but it ignores the geometric relationships of local neighborhoods, resulting in lower segmentation accuracy in complex scenes and boundary regions. PointNet++, based on PointNet, introduces a multi-scale grouping mechanism and hierarchical feature abstraction, enabling it to capture neighborhood geometric features at different scales and significantly improving its ability to perceive non-uniformly sampled point clouds and complex geometric structures. This method does not include a boundary awareness module, so segmentation blurring may still occur in the transition area between coal piles and coal bunker walls. DGCNN dynamically constructs local graph structures through EdgeConv, enabling it to learn the geometric relationships between points and enhance local feature representation. In high-density point cloud scenarios, its segmentation accuracy is superior to PointNet and ordinary PointNet++, but its sensitivity to boundary details remains limited. This invention, based on multi-scale feature abstraction, introduces a boundary-aware Transformer module. This module enhances the feature discrimination capability of the coal-wall transition region through a boundary prior attention mechanism, thereby achieving more accurate and robust segmentation in complex coal bunker scenarios. Compared with traditional methods, the method of this invention not only maintains the advantages of multi-scale feature extraction but also significantly improves the segmentation quality of boundary regions.

[0133] As shown in Table 1, the method of the present invention significantly outperforms other comparative methods in both overall accuracy (OA) and mean intersection-union ratio (mIoU). Specifically, the OA reaches [value missing]. mIoU reached Compared to the basic PointNet, it improves by [percentages]. and This result demonstrates that the boundary-aware Transformer module effectively enhances the model's ability to identify boundary regions and detailed structures in coal bunker point clouds, thereby significantly improving overall segmentation performance. Figure 5 The study demonstrates the loss variation of different methods during training: as the number of training epochs increases, the loss of all methods decreases, but the proposed method decreases at a faster rate, and after 20 epochs, it is significantly lower than other comparative models, eventually approaching zero. This result indicates that the proposed method has better convergence efficiency and a stronger ability to fit coal bunker point cloud data. Figure 6 The overall accuracy changes of each method on the test set: In the early stages of training, the OA growth trends of each method were similar. However, as the number of training epochs increased, the proposed ointNet++ multi-scale feature abstraction structure, by introducing a boundary-aware Transformer module and strengthening the feature representation of the coal-wall transition region in the attention mechanism, effectively improved the OA of the method in complex coal bunker scenarios, gradually leading the others. After 50 epochs, it consistently outperformed other models, ultimately approaching 1.0. Combined with... Figure 5 As can be seen from the loss curve, the proposed method converges quickly without overfitting, demonstrating stronger generalization ability and segmentation accuracy.

[0134]

[0135] Figure 7 The visualization results of the coal bunker point clouds (a) and (b) show that PointNet, based solely on global feature extraction, lacks sufficient ability to recognize local geometric structures and boundary regions, leading to significant missegmentation and blurring at the boundary between the coal pile and the bunker wall. PointNet++ introduces a multi-scale local feature abstraction mechanism, which can capture the spatial hierarchy information of the point cloud better, but local missegmentation still exists under complex geometric shapes. DGCNN enhances the modeling of inter-point relationships through dynamic graph convolution, resulting in high overall segmentation integrity, but discontinuities and clutter still exist in the edge transition regions. In contrast, this invention utilizes a boundary-aware Transformer module to effectively integrate boundary information during feature extraction, enabling the model to maintain a clear boundary in the transition zone between the coal pile and the bunker wall. The segmentation results are more refined, continuous, and close to the real annotation, fully verifying the superiority of this method in boundary recognition and global consistency.

[0136] Figure 8The invention includes 3D reconstruction results of two different coal piles (a) and (b). Based on the semantic segmentation and 3D reconstruction of the coal bunker point cloud, this invention further quantitatively analyzes and verifies the coal pile volume calculation results. Multiple groups of coal piles with different volume sizes were selected for experiments. The volume results calculated by the method of this invention were compared with known reference volumes, and the results are shown in Table 2. It can be seen that within different volume ranges, the method of this invention can accurately estimate the coal pile volume, with relative errors controlled within 5%, and the overall error level is relatively stable. Further analysis revealed that the volume calculation error slightly increases with the increase of the coal pile volume. This is mainly due to factors such as local occlusion inside the coal bunker, uneven point cloud density, and increased surface roughness of the coal pile, which have a certain impact on the accuracy of 3D reconstruction. However, overall, the error change trend is relatively gentle, and no obvious instability phenomenon has appeared.

[0137]

[0138] To address the challenges of complex coal bunker environments, irregular coal pile shapes, and blurred boundaries between coal piles and bunker walls, this invention proposes a method for calculating coal pile volume based on point cloud semantic segmentation. This method uses 3D LiDAR point clouds as its data foundation and employs deep learning to achieve automatic and accurate segmentation of the coal pile and bunker walls, providing reliable support for the quantitative calculation of coal pile volume. Methodologically, this invention... Based on a multi-scale feature abstraction structure, a boundary-aware Transformer module is introduced to enhance the feature representation of the coal-wall transition region in the attention mechanism, effectively improving the segmentation accuracy and robustness in complex coal bunker scenarios.

[0139] Experimental results show that the proposed method outperforms several comparative methods in terms of overall accuracy and average intersection-over-union ratio, verifying the effectiveness of the model structure. Based on this, the present invention combines point cloud post-processing with Poisson surface reconstruction to construct a 3D model of a coal pile and calculate its volume. Volume experiments demonstrate that this method can stably achieve high-accuracy volume estimation under different coal pile sizes, exhibiting good potential for engineering applications.

[0140] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the scope of the present invention.

Claims

1. A method for calculating the volume of a coal bunker pile based on point cloud semantic segmentation, characterized in that, The specific steps are as follows: Step 1: Data Acquisition and Preprocessing: First, acquire 3D point cloud data of the coal bunker using lidar, and then perform block sampling and feature construction on the original point cloud; Step 2: Constructing the network structure: Based on the PointNet++ multi-scale feature abstraction structure and introducing the boundary-aware Transformer module, fine semantic segmentation of the coal pile and background structure is achieved; The network structure consists of a multi-scale feature abstraction layer, a boundary-aware Transformer layer, and a feature propagation and point-by-point classification layer; The multi-scale feature abstraction module uses a hierarchical sampling and grouping strategy to abstract the features of point clouds in local neighborhoods at different scales, capturing the detailed structure and overall geometric shape of the coal pile surface. A boundary-aware Transformer module is introduced to further model local features. The relationship between a point and its neighborhood is characterized by a self-attention mechanism. The point cloud features extracted from high-level features are upsampled back to the original point cloud resolution layer by layer through the feature propagation module and fused with low-level geometric features. A point-by-point classifier is used to complete the semantic prediction of each point, thereby achieving fine segmentation of the coal pile and coal bunker wall. Multi-scale feature extraction layer: Let the point cloud be... Indicates the first The 3D coordinates of each point; the input feature of each point is... , Centralized coordinates are used to describe the spatial offset of a point relative to the centroid of the current point cloud block. To normalize the coordinates, the point cloud coordinates are linearly normalized and mapped to... Intervals are defined by the relative positional distribution of points within a local block; PointNet++ at sampling points field Within, with multi-scale radius For feature aggregation, the local features are calculated at each scale r as follows: (3)； in, Represents points within the neighborhood. For shared multilayer perceptron networks, relative features are introduced. and Max pooling, as a symmetric function, concatenates and fuses local features extracted at different scales to form sampling points. Multi-scale feature representation: (4)； The boundary-aware Point Transformer layer models the coal-wall transition region by introducing boundary prior constraints and incorporates them into the attention weight calculation process. The boundary-aware Point Transformer layer consists of two parts: boundary prior prediction and boundary-enhancing attention mechanism. Boundary prior prediction: A lightweight EdgeConv module is introduced to model local geometric changes and predict the boundary probability for each point. In its local neighborhood Inside, EdgeConv is used to aggregate the feature differences between points: (5)； in, Indicates a structure composed of shared multilayer sensing mechanisms Mapping function, and Points With neighboring points Feature representation, through linear mapping and Sigmoid activation function prediction points Boundary probabilities: (6); in, For learnable weight matrix, Indicates the first The input feature vector of each sample, This represents the Sigmoid activation function. ; Boundary Enhanced Attention: Local Domain in Point Transformer Within, the original attention logits are: (7)； in, For point The query indicates that, For point The key indicates, For feature dimensions; The boundary-aware Point Transformer layer introduces boundary gating on this basis: (8)； (9)； like If both are boundary points, then A larger value results in higher attention weights after softmax, thereby improving the discriminative power of boundary region features; Indicates the first When a central point performs feature aggregation on its neighboring points, the first... The normalized attention weights corresponding to each neighboring point This represents the natural exponential function, used to map correlation scores to positive values. Indicates the center point The first in its neighborhood Unnormalized correlation score between points; point The update features are: (10)； in, For point The value vector; Feature propagation and point-by-point classification: After the point cloud features are extracted and enhanced through multi-scale feature abstraction layer and boundary-aware Point Transformer layer, the feature propagation mechanism is used to upsample and fuse the high-level abstract features layer by layer, and finally restore the original point cloud resolution. (11)； This formula represents the feature of the i-th point in the l-th layer during feature propagation. From the higher-level semantic features Features of the current layer neighborhood Obtained by joint interpolation and fusion; The classification head uses pointwise MLP: (12)； This formula represents the classification prediction of the features of the i-th point. First, the point features are analyzed using the weight matrix W and the bias term b. A linear mapping is performed to obtain the response values ​​of the point in different semantic categories; then, the softmax function is used to normalize the response values ​​to obtain the probability distribution of point i belonging to each category. Finally, the semantic category of a point can be determined based on the maximum probability, and the semantic label of each point can be output. Step 3: Point cloud post-processing and volume reconstruction: Post-process and 3D reconstruction of the segmented coal pile point cloud are performed to generate a closed mesh model and complete the coal pile volume calculation.

2. The method for calculating the volume of a coal bunker based on point cloud semantic segmentation according to claim 1, characterized in that, In step one, the specific steps for data acquisition are as follows: by scanning the inside of the coal bunker from multiple angles, three-dimensional point cloud data covering the entire coal pile and bunker wall structure is obtained; The specific preprocessing steps are as follows: The original point cloud data is preprocessed using a sliding window-based point cloud dicing strategy, only in the coal bunker... A two-dimensional window of fixed size is set in a horizontal plane and slides according to a given step size to generate overlapping local point cloud patches; let the window size be... The sliding step size is Then the first The local point cloud representation corresponding to each window is as follows: ； in, , , This represents the set of points within the sliding window corresponding to the m-th row and n-th column. This represents the set of all points in the original point cloud data. Let i represent the i-th point in the point cloud set p. Point The coordinate value in the x-direction. Point The coordinate value in the y-direction. This represents the starting coordinates of the m-th window in the x-direction. This represents the starting coordinates of the nth window in the y-direction. This represents the minimum coordinate value of the current point cloud scene in the x-direction. This represents the minimum coordinate value of the current point cloud scene in the y-direction; During the block segmentation process, only local point clouds with more than a preset threshold of points within the block are retained. For point cloud blocks that meet the conditions, a random sampling strategy is used to select a fixed number of points as network input. When the number of points within the block is insufficient, it is supplemented by random repeated sampling. During the training phase, random rotation around the vertical axis, random scaling, and Gaussian noise perturbation are applied to local point cloud patches, specifically as follows: (2)； in, , , This represents the coordinates of the i-th point after the transformation. This represents the scaling factor, used to control the overall scale variation of the point cloud in space. Indicates the rotation angle around the z-axis The rotation matrix, This represents the rotation angle of the point cloud about the z-axis. This represents the perturbation term added to the point coordinates. The centroid of the point cloud is calculated for each local point cloud patch, and the point coordinates are centered, linearly normalized to a certain value. The interval is formed by concatenating the centered coordinates and the normalized coordinates to create a 6-dimensional geometric feature input network.

3. The method for calculating the volume of a coal bunker based on point cloud semantic segmentation according to claim 2, characterized in that, In step three, the entire scene is divided into blocks for prediction using a sliding window block method. The point cloud within each sliding window block is semantically classified independently by the network, and a prediction result fusion strategy based on majority voting is introduced. Sliding window block-level prediction: Within each local block divided by the sliding window, the network classifies the input point cloud point by point and outputs a semantic prediction label for each point. (13)； in, Point In the The prediction results in each sliding window block are labeled 1 to represent the coal pile and 0 to represent the coal bunker wall. Full-scene prediction result fusion: for each global point Collect its set of predicted labels across all sliding window blocks containing that point: (14)； in, This represents the number of sliding window blocks containing point i; The final semantic label of point i is determined using a majority voting strategy: (15)； As an indicator function, for the i-th point, it counts its position in... The number of times each category c is identified in the prediction is used, and the category with the highest frequency is selected as the final prediction result. ; Enhanced consistency of boundary regions: Semantic segmentation results of coal bunker point clouds are obtained through a fusion strategy; 3D Reconstruction and Geometric Calculation: Based on the semantic segmentation results of point clouds, 3D geometric reconstruction is performed on point clouds that are identified as coal piles. The point set with semantic tags as coal piles is extracted, the coal bunker walls and redundant background points are removed, and the coal pile point cloud is reconstructed in three dimensions using the Poisson surface reconstruction method. This method uses the normal information of the point cloud as a constraint, solves the Poisson equation, converts the discrete point cloud into a continuous implicit function, and generates a closed triangular mesh model on this basis. Before performing Poisson reconstruction, the normals of the coal pile point cloud are first estimated. The normal directions are obtained by analyzing the local neighborhood geometry of each point, and principal component analysis based on k-nearest neighbors or fixed-radius neighborhoods is used to calculate the local surface normals. After completing the surface reconstruction, a closed triangular mesh model is obtained. The volume of this triangular mesh is calculated based on the divergence theorem. For the part of the surface reconstructed from the vertex... The directed volume of the triangular facet formed is expressed as: (16)； The volume contribution of all triangular faces in the closed grid is summed to obtain the overall volume estimate of the coal pile.

Images