A method for segmenting three-dimensional laser point cloud of a blast pile based on multi-feature fusion

Abstract

The application relates to the technical field of blast pile segmentation, and discloses a blast pile three-dimensional laser point cloud segmentation method based on multi-feature fusion, which comprises the following steps: carrying out denoising and downsampling pretreatment on collected blast pile three-dimensional laser point cloud data to obtain pretreated point clouds; extracting multi-dimensional features of each point from the pretreated point clouds, wherein the multi-dimensional features comprise curvature features, point cloud density features and spatial context features; highlighting local details through the curvature features and the point cloud density features; maintaining global consistency through the spatial context features; and carrying out semantic pre-classification on the pretreated point clouds based on a CANUPO classifier. The application effectively solves the over-merging or under-segmentation problems of traditional methods in typical scenes such as block adhesion, serious stacking and boundary ambiguity by fusing three types of complementary features, namely local geometry, spatial distribution and global structure, and introducing a semantic-guided comprehensive similarity function in the clustering stage.

Description

Technical Field

[0001] This invention relates to the field of burst pile segmentation, specifically to a burst pile 3D laser point cloud segmentation method based on multi-feature fusion. Background Technology

[0002] Mineral resources are a crucial foundation of the national economy, and their extraction efficiency directly impacts national industrial production and energy security. Currently, open-pit mining has become a primary mode of mineral resource extraction, and blasting operations, as the core technology of open-pit mining, play a vital role in improving resource extraction efficiency and reducing production costs through quality assessment. The size of the blasted rock mass refers to the geometric dimensions of the fragments within the broken ore and rock deposits formed after blasting. Fragments that are too small will cause losses and dilution, while fragments that are too large will lead to transportation difficulties and may even require secondary blasting. Therefore, the distribution of blasted rock mass size can be used to evaluate blasting quality in order to optimize blasting plans. Three-dimensional laser scanning technology can accurately, quickly, and non-contactly acquire high-density three-dimensional spatial information of target surfaces, offering significant advantages in data acquisition accuracy and efficiency, and has become an important technological support in the field of smart mines.

[0003] Three-dimensional laser point cloud data of blast piles contains rich geometric and spatial feature information, such as curvature and point cloud density. These features provide important support for improving the accuracy of 3D laser point cloud segmentation of blast piles. Due to the complex surface morphology of blast piles, the boundaries between ore and rock usually have good geometric continuity, while the adhered areas exhibit obvious spatial feature differences. For example, in the adhered areas of ore and rock point clouds, the point cloud density is often significantly higher than that in non-adhesive areas.

[0004] Therefore, the effective fusion of multiple features of blast pile 3D laser point cloud can effectively characterize the spatial distribution of ore and rock and their adjacency relationship, which plays an important role in improving the segmentation accuracy of blast pile 3D laser point cloud. However, the existing blast pile 3D laser point cloud segmentation methods still do not make full use of spatial and geometric features such as curvature and point cloud density, resulting in limited segmentation accuracy, complex process and lack of interpretability of results. Summary of the Invention

[0005] This invention provides a multi-feature fusion-based three-dimensional laser point cloud segmentation method for explosive piles, which solves the problems mentioned in the background art that the existing three-dimensional laser point cloud segmentation methods for explosive piles still do not fully utilize spatial and geometric features such as curvature and point cloud density, resulting in limited segmentation accuracy, complex process and lack of interpretability of results.

[0006] This invention provides the following technical solution: a method for segmenting three-dimensional laser point clouds of burst-type piles based on multi-feature fusion, comprising the following steps: The collected 3D laser point cloud data of the explosive pile was denoised and downsampled to obtain a preprocessed point cloud. Multidimensional features of each point are extracted from the preprocessed point cloud, including curvature features, point cloud density features, and spatial context features. Local details are highlighted by the curvature features and point cloud density features; Global consistency is maintained through the aforementioned spatial context features; The preprocessed point cloud is semantically pre-classified based on the CANUPO classifier, dividing the point cloud into at least two categories, and outputting the semantic label and confidence score of each point. Based on the semantic tags, the weights of curvature, density, and spatial context features in the conditional Euclidean clustering process are adaptively set to complete the segmentation of each ore block in the blast pile.

[0007] As an optional scheme of the multi-feature fusion-based three-dimensional laser point cloud segmentation method for burst-type piles described in this invention, wherein: the curvature feature extraction method includes: For each point in the preprocessed point cloud, its k-nearest neighbor set is searched using a KD-Tree, and a local quadratic surface is fitted using the least squares method, as shown in the formula: in, and The horizontal coordinate; In the height direction; , , , and The coefficients to be determined are obtained by the least squares method; Then, the first and second partial derivatives at this point are calculated using the following formula: in, For along Rate of change of direction; For along Rate of change of direction; Solving for the principal curvatures based on the coefficients E, F, and G of the first fundamental form and the related terms of the second fundamental form. The formula is: Finally, the mean curvature is obtained. The formula is: .

[0008] As an optional scheme of the burst-pile 3D laser point cloud segmentation method based on multi-feature fusion described in this invention, wherein: the first basic form coefficient The calculation is as follows: First basic form coefficient The calculation is as follows: First basic form coefficient The calculation is as follows: .

[0009] As an optional scheme of the multi-feature fusion-based three-dimensional laser point cloud segmentation method for burst-type piles described in this invention, wherein: the method for extracting point cloud density features includes: Centered on a point, with a set radius Construct a spherical neighborhood; Counting the number of points in the neighborhood using a KD-Tree The point cloud density is calculated using the following formula: in, Let r be the volume of the sphere with radius r.

[0010] As an optional scheme of the multi-feature fusion-based three-dimensional laser point cloud segmentation method for bursty piles described in this invention, the method for extracting spatial context features includes: The Local Polar Coordinate Representation Module (LPR) is used to convert the Cartesian coordinates of local neighborhood points into polar coordinates to achieve rotation-invariant local geometric representation. The Dual Distance Attention Pooling Module (DDAP) is based on geometric distance. Distance from features Through attention weights By integrating the two, a comprehensive response is obtained, achieving adaptive aggregation of multi-scale information. The formula is as follows: Global Context Feature Module (GCF) calculates the local neighborhood volume. Total volume of global point cloud The ratio, used to measure the spatial proportion of a point or region in the global structure, is expressed by the formula: .

[0011] As an optional scheme of the multi-feature fusion-based three-dimensional laser point cloud segmentation method for bursty piles described in this invention, the semantic pre-classification method includes: For each point in the preprocessed point cloud, a multi-scale local geometric feature vector is constructed; Establish a spherical neighborhood around each point; Principal component analysis was performed on the neighborhood point set at each scale to obtain three principal eigenvalues ​​λ1≥λ2≥λ3; Calculate the variance ratio of each eigenvalue and construct a 2*N dimensional eigenvector to characterize the geometric morphological properties of the point at different scales; The multi-scale feature vector is input into the CANUPO classifier, and it is projected onto the maximum separability plane through linear discriminant analysis. The semantic category label of the point is output according to the confidence formula, that is, the large block category or the small block category.

[0012] As an optional scheme of the burst pile three-dimensional laser point cloud segmentation method based on multi-feature fusion described in this invention, the CANUPO classifier uses linear discriminant analysis for projection and determines the optimal projection direction by calculating the ratio of intra-class distance to inter-class distance between samples. The confidence level of the semantic category label is calculated using a posterior probability formula and used to dynamically adjust the feature weights.

[0013] As an optional scheme of the burst-pile 3D laser point cloud segmentation method based on multi-feature fusion described in this invention, it further includes an eigenvalue variance ratio, which is used to quantify the degree of distribution concentration of local point clouds in different dimensions, and the formula is: in, It represents the variance proportion along the first principal component direction, reflecting the extensibility of the point cloud along the longest axis; This represents the proportion of variance along the direction of the second principal component. This represents the variance proportion along the direction of the third principal component, reflecting the degree of dispersion of the point cloud in the normal direction.

[0014] As an optional scheme of the three-dimensional laser point cloud segmentation method for blast piles based on multi-feature fusion described in this invention, the method is as follows: based on conditional Euclidean clustering, while satisfying the geometric distance threshold, it forces the priority clustering of semantic points of the same type, avoiding excessive merging of dissimilar points due to geometric proximity, thereby achieving high-precision segmentation of each mineral and rock block in the blast pile.

[0015] As an optional scheme of the burst-pile 3D laser point cloud segmentation method based on multi-feature fusion described in this invention, the conditional Euclidean clustering uses a fusion function to determine whether two points belong to the same cluster, the formula being: in, for The curvature characteristics; Point cloud density features; For context feature vectors; , and The weighting coefficients are adaptively set based on the semantic pre-classification results; The preset similarity threshold; (.) is the discriminant function; Furthermore, two points are only grouped into the same cluster when the Euclidean distance between them is less than the geometric distance threshold, the semantic labels are consistent or the semantic confidence is higher than the tolerance, and the fusion function value satisfies the above inequality. This forces the introduction of semantic consistency constraints during the clustering process, thus avoiding excessive merging of heterogeneous mineral blocks due to geometric proximity.

[0016] The present invention has the following beneficial effects: 1. This burst-pile 3D laser point cloud segmentation method based on multi-feature fusion effectively solves the problems of over-merging or under-segmentation that traditional methods are prone to in typical scenarios such as block adhesion, severe stacking, and blurred boundaries by fusing three complementary features: local geometry, spatial distribution, and global structure, and introducing a semantically guided comprehensive similarity function in the clustering stage.

[0017] 2. This multi-feature fusion-based 3D laser point cloud segmentation method utilizes the CANUPO classifier to perform semantic pre-classification of the point cloud into large or small blocks, and dynamically adjusts the weights of each feature in clustering accordingly. This allows the algorithm to adaptively respond to the differences in geometric characteristics at different particle size scales. This mechanism overcomes the limitations of fixed thresholds or uniform feature weights in multi-scale scenarios, significantly enhancing the method's generalization ability and stability in environments with coexisting gravel areas, medium-sized blocks, and large boulders. Attached Figure Description

[0018] Figure 1 This is the overall technical roadmap of the present invention.

[0019] Figure 2 This is a flowchart of the semantic pre-classification process of the present invention.

[0020] Figure 3 This is a flowchart of the clustering and segmentation process based on semantic geometric-spatial feature fusion according to the present invention.

[0021] Figure 4 This is a schematic diagram of the first explosion pile segmentation result of the present invention.

[0022] Figure 5 This is a schematic diagram of the second explosion pile segmentation result of the present invention.

[0023] Figure 6 This is a schematic diagram of the third explosion pile segmentation result of the present invention.

[0024] Figure 7 This is a schematic diagram of the simulated burst pile segmentation results of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figures 1-7 A method for segmenting 3D laser point clouds in a burst-heap based on multi-feature fusion includes the following steps: The collected 3D laser point cloud data of the explosive pile was denoised and downsampled to obtain a preprocessed point cloud. Multidimensional features of each point are extracted from the preprocessed point cloud, including curvature features, point cloud density features, and spatial context features. Local details are highlighted by the curvature features and point cloud density features; Global consistency is maintained through the aforementioned spatial context features; The preprocessed point cloud is semantically pre-classified based on the CANUPO classifier, dividing the point cloud into at least two categories, and outputting the semantic label and confidence score of each point. Based on the semantic labels, the weights of curvature, density, and spatial context features in the clustering process are adaptively set, and a comprehensive similarity function for conditional Euclidean clustering is constructed. Based on the comprehensive similarity function, conditional Euclidean clustering is performed on the point cloud to complete the segmentation of each ore and rock block in the blast pile.

[0027] Specifically, in point cloud segmentation tasks, curvature, point cloud density, and contextual features are three key spatial features. Curvature characterizes the local geometric properties of the mineral surface, point cloud density reflects the spatial distribution and structural differences of the region, while contextual features model complex spatial structures by establishing local and global topological relationships. These three features are complementary at the representational level: curvature and point cloud density highlight local details, while contextual features emphasize global consistency. By fusing multi-feature information, more comprehensive, robust, and high-precision segmentation of 3D laser point clouds from explosive piles can be achieved. Among them, the extraction of curvature features: Curvature characteristics are used for preliminary identification and differentiation of the size of blasted ore blocks. Smaller ore blocks have rough surfaces and sharp edges, resulting in higher curvature values; while larger ore blocks have smoother surfaces and relatively lower curvature values. For any point in the point cloud... First, find its KD-Tree Nearest neighbor set Then, the least squares method is used to fit the local quadratic surface (Equation 1), at point At this point, its first and second partial derivatives are calculated using Formula 2. Principal curvature. The mean curvature can be obtained using Formula 3. The mean curvature can be calculated using Formula 4. As a local unevenness index, it can effectively distinguish the surface structure of different mineral and rock blocks, providing a geometric basis for subsequent segmentation; Formula 1: ; in, and The horizontal coordinate; In the height direction; , , , and The coefficients to be determined are obtained by the least squares method; Formula 2: ; in, For along Rate of change of direction; For along Rate of change of direction; Formula 3: ; Formula 4: .

[0028] Subsequently, the point cloud density features are extracted: Point cloud density features, as a key indicator for characterizing the spatial distribution and structural differences of blast piles, play a significant role in regional discrimination and boundary enhancement in further accurate identification of ore point clouds. This feature effectively reveals the size and stacking state information of ore blocks: large ore areas, due to their relatively flat surfaces and simple structures, exhibit sparse and uniform point cloud distribution, displaying low-density characteristics; while rubble accumulation areas, due to their complex surfaces and numerous gaps, result in dense point cloud reflections and disordered distribution, exhibiting high-density characteristics. Especially in the critical regions of ore stacking and blurred boundaries, abrupt changes in density features can effectively indicate the contact relationships between different blocks, providing crucial spatial statistical basis for subsequent segmentation. To quantify this feature, a KD-Tree spatial index is established to count the number of points within the radius neighborhood of each point. The point cloud density is calculated using Formula 5. Formula 5: ; in, For The volume of a sphere with radius .

[0029] Subsequently, contextual features are extracted: To address the issues of local feature confusion and scale sensitivity caused by the complex morphology, granularity differences, and blurred boundaries in burst-type 3D laser point clouds, this invention introduces a Spatial Contextual Features (SCF) model to achieve multi-level spatial semantic relationship modeling. This model consists of three parts: a Local Polar Representation (LPR) module, a Dual-Distance Attentive Pooling (DDAP) module, and a Global Contextual Features (GCF) module.

[0030] The LPR module achieves rotation-invariant local geometric representation by converting Cartesian coordinates to polar coordinates and enhances the directional robustness of features through centroid orientation correction. The DDAP module addresses geometric distance... Distance from features An attention fusion mechanism (Equation 6) is introduced to achieve adaptive aggregation of multi-scale information. The GCF module encodes global spatial relationships through volume ratio features (Equation 7), thereby improving the model's ability to distinguish structures at different scales.

[0031] Formula 6: in, For the first The comprehensive characteristic response of each point This refers to the attention weighting coefficient. Formula 7: ; in, It is a volume ratio characteristic. For local neighborhood volume, This represents the total volume of the global point cloud.

[0032] Subsequently, semantic pre-classification of mineral and rock point clouds for precise segmentation is performed, including: To address the issue of varying curvature sensitivity caused by uneven ore size distribution, this invention innovatively embeds the CANUPO classifier into a multi-feature fusion segmentation framework. (See reference...) Figure 2The core of this method lies in using the CANUPO classifier to perform semantic pre-segmentation of the original point cloud, dividing it into two categories: large blocks and small blocks. This provides crucial semantic prior information for subsequent conditional clustering segmentation based on features such as curvature and density. By introducing this pre-classification step, the accuracy degradation caused by using uniform curvature weights due to differences in block size is effectively avoided.

[0033] Specifically, the embedded CANUPO classifier achieves semantic partitioning through multi-scale local dimensionality analysis. It defines a series of spherical neighborhoods (radius) around each point. (Representing the analysis scale), and performing principal component analysis (PCA) on neighborhood points at each scale to obtain eigenvalues ​​λ1≥λ2≥λ3. By calculating the variance proportions of each eigenvalue, the geometric characteristics of the local point cloud in different dimensions such as 1D linear, 2D planar, or 3D volumetric can be accurately described. In N... s Repeat this process at each scale, constructing a 2×N scale for each point. s The feature vectors are multi-dimensional, thus fully recording the multi-scale dimensional evolution features. Based on this high-dimensional feature space, CANUPO uses a linear classifier to project it onto the maximally separable plane, while providing a classification confidence formula. This ultimately achieves robust classification of large and small blocks. This semantic pre-segmentation process deeply integrates CANUPO's classification capabilities with the segmentation mainline of this invention, providing structured semantic support for the subsequent adaptive weighted fusion of features such as curvature and density, significantly improving the segmentation accuracy and adaptability of complex bursty point clouds in multi-scale scenarios.

[0034] It should be noted that the eigenvalue variance ratio is used to quantify the degree of concentration of local point cloud distribution across different dimensions, and the formula is: in, It represents the variance proportion along the first principal component direction, reflecting the extensibility of the point cloud along the longest axis; This represents the proportion of variance along the direction of the second principal component. The variance percentage in the direction of the third principal component reflects the degree of dispersion of the point cloud in the normal direction. Subsequently, multi-feature fusion and clustering segmentation of the 3D laser point cloud of the burst pile were performed, see [reference]. Figure 3Unlike traditional methods that rely solely on a single geometric attribute, this mechanism comprehensively considers three complementary information components during clustering: curvature, point cloud density, and contextual features. This achieves unified modeling of local morphology, spatial distribution, and semantic relationships. Curvature characterizes the geometric undulations and boundary changes of the ore surface, point cloud density reflects the spatial aggregation characteristics of ore blocks, and contextual features reveal the semantic dependencies between point clouds. Based on this, the invention obtains ore category information through semantic pre-classification and introduces a category-adaptive weighting mechanism in the feature fusion stage, enabling ore at different scales to exhibit differentiated responses during feature fusion. This design significantly enhances the discriminativeness and robustness of feature fusion, allowing the algorithm to simultaneously consider boundary sensitivity and regional consistency in complex blasting scenarios.

[0035] Based on the above mechanism, this invention further proposes a conditional Euclidean clustering method using multi-feature weighted fusion. This method is an improved algorithm that introduces physical meaning and semantic constraints on the basis of traditional Euclidean clustering. Compared with traditional methods that rely solely on Euclidean distance, CEC can comprehensively utilize geometric, physical, and semantic features for clustering, solving the limitations caused by distance-driven Euclidean clustering in complex scenes. For example, between mineral rocks and the surface, or between adjacent rock blocks with significantly different properties, although their geometric distance is close, their physical or semantic attributes are significantly different. If clustering is based solely on geometric distance, over-merging is likely to occur, affecting the segmentation accuracy. CEC effectively avoids this problem through feature condition constraints, significantly improving the segmentation accuracy of complex burst point clouds.

[0036] It should be noted that conditional Euclidean clustering uses a fusion function to determine whether two points belong to the same cluster. The formula is as follows: in, for The curvature characteristics; Point cloud density features; For context feature vectors; , and The weighting coefficients are adaptively set based on the semantic pre-classification results; The preset similarity threshold; (.) is the discriminant function; Furthermore, two points are only grouped into the same cluster when the Euclidean distance between them is less than the geometric distance threshold, the semantic labels are consistent or the semantic confidence is higher than the tolerance, and the fusion function value satisfies the above inequality. This forces the introduction of semantic consistency constraints during the clustering process, thus avoiding excessive merging of heterogeneous mineral blocks due to geometric proximity. In summary, this invention has been applied to an open-pit copper mine, using 3D laser point cloud data of blast piles to study its effectiveness. A RIEGL VZ1000 3D laser scanner was used to acquire 20 sets of blast pile 3D laser point cloud data. Considering the special nature of the mining environment, large-scale acquisition of real blast pile 3D laser point cloud data presents many limitations. To compensate for the insufficient data volume and further verify the universality of the method, a simulated ore-rock accumulation was constructed in a controlled environment. This simulated accumulation consisted of 32 large-sized ore blocks, 50 small-sized ore blocks, and 338 pieces of gravel. Point cloud data from 23 perspectives were acquired using a FARO S150 3D laser scanner. To further improve the realism of the simulated blast pile, point cloud data from three perspectives were selected for subsequent experiments.

[0037] To comprehensively evaluate the segmentation performance of this method on 3D laser point clouds of different types of explosive piles, three metrics—precision (P), recall (R), and average particle size accuracy (D)—were selected for quantitative evaluation. Experimental results are presented from... Figures 4-6 It can be intuitively seen that this method can achieve relatively ideal segmentation results for both real 3D laser point clouds of blasted piles No. 3, 11, and 17, and simulated blasted pile 3D laser point clouds. Especially on real blasted pile data, the model can distinguish the boundaries of adjacent blocks well, significantly reducing the phenomena of "over-merging" and "under-segmentation", and demonstrating strong structural sensitivity and boundary preservation ability.

[0038] As shown in the quantitative results in the table above, the No. 3 blast pile exhibits the highest segmentation accuracy, with all three indicators exceeding 0.85, including a precision rate of 0.9127. This indicates that the proposed method still possesses strong recognition and classification capabilities even under complex surface morphologies. In contrast, the segmentation performance of the simulated blast pile data is slightly lower, with an average particle size accuracy of only 0.7127. This difference mainly stems from the inconsistency in the distribution of data features: the real blast pile has a wider distribution of ore and rock particles and a more complex geometric structure, providing more comprehensive contextual information for the model; while the simulated blast pile, due to construction limitations, has a relatively simple block shape and particle size distribution, resulting in a reduced sensitivity of the model to changes in feature scale, thus affecting the segmentation accuracy. In summary, the experimental results verify the effectiveness and robustness of the context-feature-based segmentation method proposed in this invention in real blast pile 3D laser point clouds, especially demonstrating outstanding performance in scenarios with complex structures and significant scale variations.

[0039] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0040] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for segmenting three-dimensional laser point clouds of a burst-type pile based on multi-feature fusion, characterized in that, Includes the following steps: The collected 3D laser point cloud data of the explosive pile was denoised and downsampled to obtain a preprocessed point cloud. Multidimensional features of each point are extracted from the preprocessed point cloud, including curvature features, point cloud density features, and spatial context features. Local details are highlighted by the curvature features and point cloud density features; Global consistency is maintained through the aforementioned spatial context features; The preprocessed point cloud is semantically pre-classified based on the CANUPO classifier, dividing the point cloud into at least two categories, and outputting the semantic label and confidence score of each point. Based on the semantic tags, the weights of curvature, density, and spatial context features in the conditional Euclidean clustering process are adaptively set to complete the segmentation of each ore block in the blast pile.

2. The method for segmenting three-dimensional laser point clouds of a burst-type pile based on multi-feature fusion according to claim 1, characterized in that, The method for extracting curvature features includes: For each point in the preprocessed point cloud, its k-nearest neighbor set is searched using a KD-Tree, and a local quadratic surface is fitted using the least squares method, as shown in the formula: Where x and y are the horizontal coordinates; z represents the height direction; , , , and The coefficients to be determined are obtained by the least squares method; Then, the first and second partial derivatives at this point are calculated using the following formula: in, For along Rate of change of direction; For along Rate of change of direction; Based on the first fundamental form coefficients , and The second fundamental form related terms, solving for the principal curvature. The formula is: Finally, the mean curvature is obtained. The formula is: 。 3. The method for segmenting three-dimensional laser point clouds of a burst-type pile based on multi-feature fusion according to claim 2, characterized in that: The first fundamental form coefficient E is calculated as follows: The first basic form coefficient F is calculated as follows: The first basic form coefficient G is calculated as follows: 。 4. The method for segmenting three-dimensional laser point clouds of a burst-type pile based on multi-feature fusion according to claim 1, characterized in that: The method for extracting point cloud density features includes: Centered on a point, with a set radius Construct a spherical neighborhood; Counting the number of points in the neighborhood using a KD-Tree method The point cloud density is calculated using the following formula: in, Let r be the volume of the sphere with radius r.

5. The method for segmenting three-dimensional laser point clouds of a burst-type pile based on multi-feature fusion according to claim 1, characterized in that: The method for extracting spatial context features includes: The Local Polar Coordinate Representation Module (LPR) is used to convert the Cartesian coordinates of local neighborhood points into polar coordinates to achieve rotation-invariant local geometric representation. The Dual Distance Attention Pooling Module (DDAP) is based on geometric distance. Distance from features Through attention weights By integrating the two, a comprehensive response is obtained, achieving adaptive aggregation of multi-scale information. The formula is as follows: Global Context Feature Module (GCF) calculates the local neighborhood volume. Total volume of global point cloud The ratio, used to measure the spatial proportion of a point or region in the global structure, is expressed by the formula: 。 6. The method for segmenting three-dimensional laser point clouds of a burst-type pile based on multi-feature fusion according to any one of claims 1-4, characterized in that, The semantic pre-classification method includes: For each point in the preprocessed point cloud, a multi-scale local geometric feature vector is constructed; Establish a spherical neighborhood around each point; Principal component analysis was performed on the neighborhood point set at each scale to obtain three principal eigenvalues ​​λ1≥λ2≥λ3; Calculate the variance ratio of each eigenvalue and construct a 2*N dimensional eigenvector to characterize the geometric morphological properties of the point at different scales; The multi-scale feature vector is input into the CANUPO classifier, and it is projected onto the maximum separability plane through linear discriminant analysis. The semantic category label of the point is output according to the confidence formula, that is, the large block category or the small block category.

7. The method for segmenting three-dimensional laser point clouds of a burst-type pile based on multi-feature fusion according to claim 6, characterized in that: The CANUPO classifier uses linear discriminant analysis for projection and determines the optimal projection direction by calculating the ratio of intra-class distance to inter-class distance. The confidence level of the semantic category label is calculated using a posterior probability formula and used to dynamically adjust the feature weights.

8. The method for segmenting three-dimensional laser point clouds of a burst-type pile based on multi-feature fusion according to claim 7, characterized in that, It also includes the eigenvalue variance ratio, which is used to quantify the degree of concentration of local point cloud distribution in different dimensions, and the formula is: in, It represents the variance proportion along the first principal component direction, reflecting the extensibility of the point cloud along the longest axis; This represents the proportion of variance along the direction of the second principal component. This represents the variance proportion along the direction of the third principal component, reflecting the degree of dispersion of the point cloud in the normal direction.

9. The method for segmenting three-dimensional laser point clouds of a burst-type pile based on multi-feature fusion according to claim 8, characterized in that: Based on conditional Euclidean clustering, while satisfying the geometric distance threshold, it forces the priority clustering of semantic points of the same type, avoiding excessive merging of dissimilar points due to geometric proximity, thereby achieving high-precision segmentation of each mineral and rock block in the blast pile.

10. The method for segmenting three-dimensional laser point clouds of a burst-type pile based on multi-feature fusion according to claim 9, characterized in that, The conditional Euclidean clustering uses a fusion function to determine whether two points belong to the same cluster. The formula is as follows: in, for The curvature characteristics; Point cloud density features; For context feature vectors; , and The weighting coefficients are adaptively set based on the semantic pre-classification results; The preset similarity threshold; (.) is the discriminant function; Furthermore, two points are only grouped into the same cluster when the Euclidean distance between them is less than the geometric distance threshold, the semantic labels are consistent or the semantic confidence is higher than the tolerance, and the fusion function value satisfies the above inequality. This forces the introduction of semantic consistency constraints during the clustering process, thus avoiding excessive merging of heterogeneous mineral blocks due to geometric proximity.

Images