Data-driven strip mine excavation monitoring method and system based on 3D point cloud

Abstract

The invention relates to the technical field of mine engineering and geographic information, and discloses a data-driven strip mine excavation monitoring method and system based on 3D point cloud. A two-stage registration strategy is adopted for the multi-tense open-pit mine point cloud, EGS coarse registration and G-ICP fine registration are combined, and registration parameters are dynamically determined by analyzing quantile distribution of the nearest neighbor distance of the point cloud; performing change detection by using an M3C2 algorithm in combination with a voxel adaptation module, estimating registration uncertainty through the M3C2 distance standard deviation of a stable region, and generating a significant change mask; volume dynamic estimation is carried out based on DEM, and excavation and filling volume and net volume change are calculated through intersection-to-parallel ratio guided dynamic region segmentation and adaptive grid resolution interpolation. According to the method, the technical problems of low manual parameter adjustment efficiency, poor noise robustness and slow large-scale data processing of a traditional method are solved, and high-precision and automatic monitoring of the excavation volume of the strip mine is realized.

Description

Technical Field

[0001] This invention relates to the fields of mining engineering and geographic information technology, specifically to a data-driven open-pit mine excavation monitoring method and system based on 3D point clouds, applicable to scenarios such as dynamic monitoring of excavation volume, slope stability analysis, and resource reserve assessment in large-scale open-pit mines. Background Technology

[0002] As global mineral resource development moves towards intelligent and large-scale operations, efficient mining and safety monitoring in open-pit mines face severe challenges. Traditional open-pit mine monitoring mainly relies on single-point measurement methods such as total stations and GPS (Global Positioning System), which suffer from low data density, poor operational efficiency, and inability to capture subtle terrain changes. In recent years, the development of three-dimensional laser scanning (TLS) and UAV photogrammetry technologies has made 3D point cloud data the mainstream data source for open-pit mine monitoring; however, its application is still limited by the following technical bottlenecks:

[0003] (1) Parameter dependence problem: Traditional point cloud analysis methods, such as the Iterative ClosestPoint (ICP) algorithm, rely heavily on manually setting registration thresholds and clustering parameters, which makes it difficult to adapt to the complex and varied terrain conditions of open-pit mines. For example, the point cloud density of mines at different mining stages can vary by more than 10 times, and manual parameter adjustment is inefficient and can easily lead to volume estimation errors exceeding 5%.

[0004] (2) Insufficient noise robustness: Factors such as dust, light changes and equipment vibration in open-pit mines can introduce a large number of outliers into the point cloud. Existing methods, such as RANSAC (RANdom Sampling Consensus) based registration, often result in mismatches in high-noise scenarios, leading to a decrease in the accuracy of slope deformation monitoring.

[0005] (3) Bottleneck of large-scale data processing: A single scan of a large open-pit mine can generate billions of point cloud data. Traditional grid-based processing methods have a memory occupancy rate of over 90% and the volume calculation takes several hours, which cannot meet the real-time monitoring requirements.

[0006] (4) Difficulty in multi-source data fusion: Due to the different sampling methods, TLS and UAV point clouds have coordinate system deviations. Existing rigid registration methods are difficult to handle non-rigid deformation of mine slopes, such as local settlement caused by blasting.

[0007] (5) Limitations of existing typical technical solutions:

[0008] Currently, the technologies applied to mine change monitoring mainly include:

[0009] 1) Feature-based registration method: Point cloud alignment is performed by extracting geometric features such as planes and edges. However, the homogeneity of open-pit mine slopes results in a feature extraction rate of less than 30%, and the registration accuracy is only at the centimeter level.

[0010] 2) Deep learning methods: Although PointNet++ can automatically extract features, it requires tens of thousands of sets of labeled data, while mine safety restrictions make large-scale data collection costly.

[0011] 3) Traditional volume calculation method: Volume calculation based on Triangular Irregular Network (TIN) is prone to "hole effect" in sparse point cloud regions. Actual measurements show that when the point density is less than 1 point / m², the volume error exceeds 10%.

[0012] Constructing an open-pit mine monitoring system that requires no manual parameter tuning, adapts to noisy environments, and supports real-time processing of massive point clouds remains a critical technical challenge in this field. In particular, there are gaps in the following areas: a dynamic adaptive point cloud registration parameter generation mechanism; a change detection algorithm that integrates multi-scale geometric features; and an efficient volume estimation model for large-scale data. Summary of the Invention

[0013] This invention aims to provide a data-driven open-pit mine excavation monitoring method and system based on 3D point clouds. By automatically optimizing processing parameters through data statistical features, it solves the technical problems of low efficiency, poor noise robustness, and slow large-scale data processing in traditional methods, achieving high-precision (error <2%) and automated monitoring of open-pit mine excavation volume. The technical solution is as follows:

[0014] A data-driven open-pit mine excavation monitoring method based on 3D point clouds includes the following steps:

[0015] Step 1: A two-stage registration strategy is adopted for multi-temporal open-pit mine point clouds, combining EGS coarse registration and G-ICP fine registration. The registration parameters are dynamically determined by analyzing the quantile distribution of the nearest neighbor distance of the point cloud.

[0016] Step 2: Use the M3C2 algorithm combined with the voxel adaptation module to perform change detection, estimate the registration uncertainty by the standard deviation of the M3C2 distance in the stable region, and generate a mask of significant changes;

[0017] Step 3: Perform dynamic volume estimation based on the digital elevation model. Calculate the cut and fill volume and net volume changes by using dynamic region segmentation guided by the intersection-union ratio and adaptive grid resolution interpolation.

[0018] A data-driven open-pit mine excavation monitoring system based on 3D point clouds includes:

[0019] Adaptive point cloud registration module: It realizes coarse EGS registration and fine G-ICP registration, and dynamically determines the registration parameters by analyzing the quantile distribution of the nearest neighbor distance of the point cloud;

[0020] Change detection module: Integrates M3C2 algorithm and voxel adaptation module, estimates registration uncertainty by M3C2 distance standard deviation in stable regions, generates significant change mask, and outputs significant change regions;

[0021] Volume calculation module: used to calculate cut and fill volume based on adaptive digital elevation model grid;

[0022] Visualization module: Used to generate volume change cloud maps and 3D scene displays.

[0023] The beneficial effects of this invention are:

[0024] 1) This invention utilizes the statistical characteristics of point cloud data (such as point density distribution, nearest neighbor distance, etc.) to automatically derive the optimal processing parameters, avoiding manual intervention; it adopts an adaptive point cloud registration based on a coarse-to-fine strategy, which can dynamically determine the maximum corresponding distance, while constraining the dual distances between points and between points and the plane, thereby improving adaptability to non-planar terrain.

[0025] 2) This invention integrates the M3C2 algorithm with the voxel adaptation module for change detection, which can better adapt to the differences in point cloud density in different areas of open-pit mines, significantly improve the robustness to complex noise (such as measurement noise caused by dust and changes in illumination) and point cloud heterogeneity, and effectively filter out false changes caused by registration residuals and measurement noise.

[0026] 3) This invention uses dynamic mesh volume based on DEM, and even when the noise level is 100mm, the volume error can still be controlled within 2%.

[0027] 4) This invention solves the technical problems of low efficiency, poor noise robustness, and slow large-scale data processing of traditional methods by automatically optimizing processing parameters through data statistical features, thereby achieving high-precision (error <2%) and automated monitoring of open-pit mine excavation volume. Attached Figure Description

[0028] Figure 1 This is a diagram of the overall system architecture.

[0029] Figure 2(a) is a schematic diagram of registration uncertainty assessment - the relative position of the selected stable area (gray) in the open-pit mine point cloud and the overall point cloud (semi-transparent).

[0030] Figure 2(b) is a schematic diagram of registration uncertainty assessment - point cloud of stable region.

[0031] Figure 3(a) shows the volume change estimation process – DBSCAN clustering results.

[0032] Figure 3(b) shows the volume change estimation process – the difference between the two DEMs.

[0033] Figure 4(a) shows the original mine scan data.

[0034] Figure 4(b) shows the simulated mine scan data.

[0035] Figure 5(a) shows the impact of EGS voxel size on GPU and CPU memory usage.

[0036] Figure 5(b) shows the effect of EGS voxel size on runtime.

[0037] Figure 5(c) shows the effect of EGS voxel size on the relative errors of rotation and translation.

[0038] Figure 6 The graph shows the optimization results for the G-ICP quantile threshold.

[0039] Figure 7 This is a graph showing the effect of noise on M3C2 detection. Detailed Implementation

[0040] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0041] This invention proposes a comprehensive solution integrating "adaptive registration, robust change detection, and efficient volume estimation." Its core lies in automatically deriving optimal processing parameters using the statistical characteristics of point cloud data (such as point density distribution and nearest neighbor distance), avoiding manual intervention. Specifically, it includes:

[0042] 1) An adaptive point cloud registration module based on a coarse-to-fine strategy;

[0043] 2) Integrating the M3C2 (Multi-scale Model-to-Model Cloud Comparison) algorithm with the change detection module of the Voxel Adaptation Module (VAM);

[0044] 3) Dynamic mesh volume calculation module based on Digital Elevation Model (DEM).

[0045] like Figure 1As shown, the workflow of this invention is as follows: from multi-temporal point cloud input to net volume change output, including EGS (Exhaustive Grid Search) coarse registration, G-ICP (Generalized Iterative Closest Point) fine registration, M3C2 change detection, and DEM volume calculation.

[0046] 1. Adaptive point cloud registration:

[0047] (1) EGS coarse registration:

[0048] When the initial pose uncertainty of the point cloud is large, such as when there is no ground scan for GNSS (Global Navigation Satellite System) positioning, the EGS algorithm is used for global search. The specific process is as follows:

[0049] The source and target point clouds are voxels respectively, and the voxel size is dynamically adjusted according to the point cloud density. The calculation formula is as follows: ;

[0050] in, The nearest neighbor distance for the point cloud.

[0051] In the discretized rotation space The weighted cross-correlation coefficient is calculated in the core spatial model (used in the coarse registration stage of point cloud, representing all possible rotation states in three-dimensional space), with the search range set to ±45° (z-axis) and ±5° (x / y-axis), and the angle step size of 2°.

[0052] When the system has positioning equipment such as GNSS / IMU (Inertial Measurement Unit), this step can be skipped to improve efficiency.

[0053] (2) G-ICP precision registration:

[0054] The G-ICP algorithm is used to optimize the coarse registration results. Its innovation lies in dynamically determining the maximum correspondence distance.

[0055] Calculate the nearest neighbor distance distribution of the input point cloud, and select the 20th percentile as the threshold: ;

[0056] in, It represents the nearest neighbor distance for each point in the point cloud.

[0057] It simultaneously constrains both point-to-point and point-to-plane distances, enhancing adaptability to non-planar terrain.

[0058] 2. M3C2 change detection:

[0059] (1) Registration uncertainty assessment:

[0060] Select a stable area in the mine (such as an unexcavated rock surface) as a reference and extract the point cloud within the polygon mask.

[0061] Calculate the standard deviation of the M3C2 distance in the stable region as an estimate of the registration error: ;

[0062] The distance M3C2 is calculated along the local surface normal.

[0063] (2) Multiscale variation analysis:

[0064] The normal and projection dimensions of M3C2 are dynamically determined using the VAM module: ; ; ;

[0065] in, Point cloud density, The preset adjustment parameter is set to 0.01 to balance accuracy and efficiency. It is a proportional-integral control function; Normal dimension; The projection scale.

[0066] Calculate the detection threshold (Level of Detection, LoD) for each core point to distinguish between real changes and noise: ;

[0067] in, The preset adjustment parameters are taken as empirical values ​​(1.96 corresponds to the 95% confidence interval).

[0068] (3) Generation of significant change masks:

[0069] If the distance difference between two point clouds exceeds the local LoD, it is marked as a point of significant change: ;

[0070] This mask can effectively filter out more than 90% of spurious changes caused by noise.

[0071] 3. DEM volume calculation:

[0072] (1) Dynamic region segmentation:

[0073] The DBSCAN (Density-based clustering algorithm) algorithm is used to cluster the changing point cloud, with a cluster radius of [missing information]. The adaptive calculation is as follows: ;

[0074] in, The next nearest neighbor distance of the point of change. Preset adjustment parameters are used. A segmentation strategy guided by IOU (Intersection-over-Union ratio) focuses on regions with high variability, reducing computational load.

[0075] (2) DEM interpolation and mesh generation:

[0076] Determine a unified reference surface: Take the minimum z-coordinate of the two point cloud periods as the zero elevation point. ;

[0077] Dynamically calculate grid resolution: ;

[0078] in, The preset adjustment parameter is taken as an empirical value of 1.0. , These are the nearest neighbor distances for the two point clouds, respectively. This is the median calculation function.

[0079] Linear interpolation is used to generate the DEM to ensure the interpolation accuracy in sparse regions.

[0080] (3) Calculation of volume change:

[0081] Calculate the elevation difference between the two DEMs: ;

[0082] Calculate volume change by grid cell: ;

[0083] Accumulate excavation separately With fill volume: ; ;

[0084] Calculate the change in net volume: ;

[0085] This method can still control the volume error within 2% when the noise level is 100mm.

[0086] The overall system architecture is as follows Figure 1 As shown, the arrows indicate the direction of data flow between modules, and the dashed boxes represent adaptive parameter generation units, showing how each module dynamically adjusts processing parameters using point cloud statistical features.

[0087] The present invention provides a data-driven open-pit mine excavation monitoring system based on 3D point clouds, which specifically includes:

[0088] The adaptive point cloud registration module enables coarse EGS registration and fine G-ICP registration, and includes a dynamic parameter generation unit.

[0089] The change detection module integrates the M3C2 algorithm and the VAM module, and outputs areas of significant change.

[0090] The volume calculation module calculates the volume of cut and fill based on an adaptive DEM grid.

[0091] The visualization module generates volume change cloud maps and 3D scene displays.

[0092] The registration uncertainty assessment is shown in Figures 2(a) and (b). Figure 2(a) shows the relative position of the selected stable region (gray) in the open-pit mine point cloud to the overall point cloud (semi-transparent). Figure 2(b) shows the extracted stable region point cloud, used to calculate the registration error. The polygon mask was drawn using QGIS (Quantum GIS) software to ensure that the rock surface area unaffected by excavation was selected.

[0093] The volume change estimation process is shown in Figures 3(a) and (b). Figure 3(a) shows the DBSCAN clustering results, with red representing the unchanged area, blue representing the non-significant change segment, and rainbow colors representing IOU weights. Figure 3(b) is a difference map of the two DEMs, with warm colors representing fill and cold colors representing excavation, clearly reflecting the spatial distribution of mine excavation.

[0094] Figures 4(a) and (b) show the volumetric error comparison under different noise levels. The box plots in the figures show the relative error distribution of the method of the present invention under different real volumes (101.87m³-12733.24m³) within the noise range of 1mm-100mm. The results show that even under 100mm noise and "Hard" registration difficulty, the error is still less than 2%, verifying the robustness of the method.

[0095] The impact of EGS voxel size on performance is shown in Figures 5(a)-(c), which illustrate performance metrics as the EGS voxel size changes from 1m to 3m. Regarding memory usage, peak GPU memory usage decreased from 6.65 GiB to 0.26 GiB; runtime decreased from 16.4 minutes to 4.47 minutes; and registration accuracy improved from a median RRE of 1.3° at 2m voxels to 1.5° at 3m.

[0096] like Figure 6 As shown in the figure, the impact of the quantile threshold (10%-90%) of the maximum corresponding distance of G-ICP on performance is as follows: when the threshold is ≤30%, the volume error is stable at 0.15%-0.18%; when the threshold is 20%, a time saving of 14% and an average error of 0.16% are achieved, which is the optimal operating point.

[0097] like Figure 7 As shown, when 10mm-100mm Gaussian noise is added, the detection accuracy IoU decreases from 0.93 to 0.91, and the misclassification rate increases from 0.64% to 0.84%; the deformation area LoD increases from 5.8cm to 9.2cm, showing a linear correlation with the noise level.

Claims

1. A data-driven 3D point cloud based open pit excavation monitoring method, characterized in that, The method comprises the following steps: Step 1: A two-stage registration strategy is adopted for the multi-temporal open-pit point cloud, combined with EGS coarse registration and G-ICP fine registration, and the registration parameters are dynamically determined by analyzing the quantile distribution of the nearest neighbor distance of the point cloud; Step 2: The M3C2 algorithm is used in combination with the voxel adaptive module for change detection, the registration uncertainty is estimated by the M3C2 distance standard deviation of the stable area, and a significant change mask is generated; Step 3: Based on the digital elevation model, the volume is dynamically estimated, the dynamic region segmentation guided by the intersection-over-union ratio and the adaptive grid resolution interpolation are used to calculate the excavation and filling volume and the net volume change.

2. The 3D point cloud based data-driven open-pit excavation monitoring method according to claim 1, characterized in that, In step 1, when the initial pose uncertainty of the point cloud is greater than the set threshold, EGS coarse registration is adopted, and then G-ICP fine registration is used to optimize the coarse registration result; when the initial pose uncertainty of the point cloud is less than the set threshold, the EGS coarse registration is skipped, and the G-ICP fine registration is directly performed; The EGS coarse registration voxelizes the source point cloud and the target point cloud respectively, and the voxel size According to the point cloud density dynamic adjustment, the calculation formula is: ； wherein, is the point cloud nearest neighbor distance; is the median computation function; The weighted cross-correlation coefficient is calculated in the discretized rotation space; The G-ICP fine registration calculates the nearest neighbor distance distribution of the input point cloud, and selects the 20% quantile as the threshold: ； wherein, is the nearest neighbor distance for each point of the point cloud; is a quantile computation function.

3. The 3D point cloud based data-driven open-pit excavation monitoring method of claim 2, wherein, Step 2 is specifically: Step 2.1: Registration uncertainty evaluation; The stable area of the mine site is selected as the reference, and the point cloud in the polygon mask is extracted; M3C2 distance standard deviation of the stable region is calculated as a registration error estimate value, the calculation formula is: ； wherein M3C2 three-dimensional distance between volume elements, calculated along the local surface normal; is the standard deviation calculation function; Step 2.2: Multi-scale change analysis; The voxel adaptive module is used to dynamically determine the normal scale and projection scale of the M3C2 three-dimensional distance: ； ； ； wherein is a point cloud density, is a preset adjustment parameter, is a proportional-integral control function; is a normal scale; is a projection scale; The detection threshold LoD of each core point is calculated, which is used to distinguish real changes from noise: ； wherein, is a preset adjustment parameter; Step 2.3: Significant change mask generation; If the distance difference of two period point clouds exceeds the detection threshold LoD, it is marked as a significant change point is expressed as: ； wherein, is the distance difference of two phase point clouds.

4. The 3D point cloud based data-driven open-pit excavation monitoring method of claim 3, wherein, Step 3 is specifically: Step 3.1: Dynamic region segmentation; The change point cloud is clustered by using a density-based clustering algorithm, and a clustering radius The adaptive computation is as follows: ； wherein, is the next neighbor distance of the change point, is a preset adjustment parameter; is a mean value calculation function; Step 3.2: Digital elevation model interpolation and grid generation; Taking the minimum value of the z coordinate of two period point clouds As the zero point of altitude, a uniform reference surface is determined and is expressed as follows: ； wherein, and are the z-coordinates of the two phases of point clouds, respectively; Dynamic calculation of grid resolution ; ； wherein, , are the nearest neighbor distances of the two phases of point clouds respectively; is a preset adjustment parameter (taking an empirical value 1.0); is a median calculation function; Step 3.3: Volume change calculation; calculating an elevation difference between two digital elevation models : ； wherein, and are the elevations of the two digital elevation models, respectively; Calculating volume change by grid cell : ； : respectively accumulate the volume of excavation and the volume of fill : ； ； wherein, is the volume of a single grid cell; Calculating net volume change : 。 5. A 3D point cloud based data-driven open pit excavation monitoring system employing the monitoring method of any one of claims 1-4, characterized in that, It comprises: Adaptive point cloud registration module: EGS coarse registration and G-ICP fine registration are realized, and the registration parameters are dynamically determined by analyzing the quantile distribution of the nearest neighbor distance of the point cloud; Change detection module: The M3C2 algorithm and the voxel adaptive module are integrated, the registration uncertainty is estimated by the M3C2 distance standard deviation of the stable area, and a significant change mask is generated, and the significant change area is output; Volume calculation module: used for calculating the excavation and filling volume based on the adaptive digital elevation model grid; Visualization module: used for generating volume change cloud chart and three-dimensional scene display.

Images