Local updating system and method for high-precision map of strip mine

Abstract

The invention provides a local updating system and method for a high-precision map of a strip mine, and the system comprises six modules: a multi-source data collection module, an intelligent change detection module, an incremental three-dimensional reconstruction module, a space-time database management module, a visual verification module and a mining and stripping amount measurement module. PointNet + + and U-Net networks are fused to realize change region extraction with the precision less than or equal to 1m, only local reconstruction is carried out on the change region, centimeter-level seamless fusion is realized by adopting graph optimization and Laplacian grid deformation, and updating is completed in a short time. The model supports version second-level rollback, WebGL30FPS rendering and TIN method automatic amount calculation, the error is less than 1.2%, the bandwidth is saved by 90%, the GPU occupation is reduced by 75%, the requirement of an unmanned mine card for the current trend less than or equal to 30 minutes under the pace of daily propulsion of a strip mine is met, mining while updating is realized, and a real-time data base is provided for intelligent scheduling and digital twin mines.

Description

Technical Field

[0001] This invention relates to the fields of mine surveying and mapping and 3D real-scene modeling technology, specifically a high-precision map local update system and method for open-pit mines. Background Technology

[0002] Open-pit mines employ a continuous operation mode of "daily advancement of the mining and stripping face and dynamic relocation of loading and unloading areas," which leads to sudden changes in the end of transport roads, slopes, curvatures, and temporary avoidance zones within hours, sometimes at the centimeter or even meter level. Simultaneously, unforeseen factors such as blasting, route rerouting, and rain erosion further exacerbate the rapid "drifting" of road alignment. Unmanned mining trucks rely on high-precision maps for centimeter-level positioning and determination of drivable areas. Insufficient map real-time accuracy will directly trigger safety stops or entry into abandoned road sections, causing production interruptions and safety hazards.

[0003] Existing technologies generally adopt a full-scale, all-element high-precision map update mode, which involves periodically (usually weekly or monthly) re-collecting, stitching point clouds, labeling features, converting formats, and distributing the entire road network in the mining area. This mode exhibits drawbacks in open-pit mine scenarios, including long update cycles, high computational resource consumption, high data transmission bandwidth, low production matching, and severe redundant updates. It cannot match the "daily-progressive" production rhythm of mines and requires additional planning for shutdown or avoidance periods, impacting production capacity.

[0004] Therefore, the existing full update mode has become the main constraint on the large-scale and normalized operation of unmanned mining trucks in open-pit mines. There is an urgent need for a high-precision map local update method and system that is incremental, fast, and has low bandwidth, which is oriented towards dynamic mining scenarios, in order to solve the problem of the mismatch between map timeliness and the continuous production rhythm of the mine. Summary of the Invention

[0005] The main objective of this invention is to address the problems existing in the prior art and provide a high-precision map local update system and method for open-pit mines.

[0006] The technical solution of the present invention is as follows: a high-precision map local update system for open-pit mines, the system comprising: a multi-source heterogeneous data acquisition module, used to acquire multi-source real-scene three-dimensional data of the open-pit mine excavation face at a high frequency not lower than the mining production process cycle.

[0007] The intelligent change detection and region extraction module is communicatively connected to the multi-source heterogeneous data acquisition module. It is used to automatically identify and extract areas where the working face is continuously excavated and the terrain is rapidly changing, within the point cloud-image-semantic multimodal data fusion framework, with a positioning uncertainty threshold of ≤1m.

[0008] The incremental 3D reconstruction and update module is communicatively connected to the intelligent change detection and region extraction module. It is used to perform local real-scene 3D model reconstruction only on the changed region, and to use graph optimization and seamless stitching algorithms to fuse the newly generated local 3D model with the historical global 3D model to generate an updated global real-scene 3D model.

[0009] The spatiotemporal database and version management module is used to perform hierarchical storage, version management, and incremental rollback of the global real-scene 3D model.

[0010] The visualization, interaction and verification platform is used to perform 3D visualization, human-computer interaction verification, and error statistics on the updated global real-scene 3D model.

[0011] The automatic calculation module for mining and stripping work volume is used to automatically calculate the current mining and stripping volume based on the updated global real-scene 3D model, and generate progress comparison reports and statistical charts.

[0012] Furthermore, the multi-source heterogeneous data acquisition module includes: a vehicle-mounted / roadside mobile acquisition subsystem, which consists of a lidar, a camera, and a GNSS / IMU integrated navigation unit, and is installed on an unmanned mining truck, patrol vehicle, or fixed pole position, for acquiring raw point cloud and image data of the working face and surrounding environment at a frequency of 10Hz-20Hz.

[0013] The UAV aerial survey subsystem is used to perform oblique photogrammetry on the macroscopic area of ​​the mining face and spoil heap to obtain centimeter-level topographic change data.

[0014] The production operation system interface is used to obtain the operating location, planned path, and operation semantic information of excavators and bulldozers in real time from the mine scheduling system through the API interface, and to serve as a prior constraint for change detection.

[0015] Furthermore, the intelligent change detection and region extraction module includes a multimodal registration unit, which is used to complete the sub-pixel matching of current data and historical models based on the initial GNSS / IMU pose value and the NDT-ICP joint algorithm.

[0016] The Variation Semantic Segmentation Unit is used to output a probability map at the pixel / voxel level in the joint feature space of point cloud and image by employing a cross-modal deep learning network that integrates PointNet++ and U-Net.

[0017] The region aggregation and contour vectorization unit is used to perform morphological filtering, clustering, and Ramer-Douglas-Peucker vectorization on the probability map to generate closed variable polygons, and to use the variable polygons as the input ROI of the incremental 3D reconstruction and update module.

[0018] Furthermore, the incremental 3D reconstruction and update module includes a local point cloud densification unit, used to perform dense reconstruction of image sequences within the change region based on the SGM or MVS algorithm, so as to obtain a local point cloud with a grid resolution of ≤5cm.

[0019] Triangular mesh (Mesh) or 3D Gaussian sputtering (3DGS) reconstruction units are used to convert local point clouds into realistic 3D models.

[0020] The graph optimization fusion unit is used to align the local real-world 3D model with the historical global model using a pose graph optimization method with edge-face constraints, and to eliminate seams using a Laplacian-based mesh deformation algorithm to achieve centimeter-level stitching error.

[0021] Furthermore, the spatiotemporal database and version management module includes:

[0022] The hierarchical storage engine is used to store real-world 3D models in layers according to spatial grids, feature categories, and timestamps, and adopts a hybrid architecture of MongoDB and object storage.

[0023] The version chain management unit is used to generate a globally unique version ID for each incremental update and record feature-level differences based on a Merkle tree structure to support quick rollback to any historical version.

[0024] The concurrent locking mechanism is used to apply write locks to the affected regions during model updates and to notify all subscribed nodes via WebSocket event broadcast after the update is completed.

[0025] Furthermore, the visualization interaction and verification platform includes:

[0026] The WebGL 3D rendering engine is used to render real-world 3D models with a range of ≥10km² at a frame rate of at least 30FPS.

[0027] Multi-phase differential overlay units are used to highlight and overlay the excavation range of multiple phases in different colors, intuitively presenting the evolution trajectory and morphological changes of the working face.

[0028] The human-computer interaction tool is used to support operators to manually confirm or correct changes in areas by clicking, selecting boxes, or using voice input, and to feed back the correction results to the intelligent change detection and region extraction module for online retraining.

[0029] Furthermore, the automatic calculation module for mining and stripping quantities includes:

[0030] The volume calculation unit is used to automatically calculate the current mining and stripping volume based on the updated high-precision real-world 3D model using the triangular mesh (TIN) comparison method.

[0031] The progress comparison report unit is used to generate reports and statistical charts on differences in project quantities across multiple periods.

[0032] The interactive query unit supports users in interactively querying model information and engineering quantity data at any location and time period, providing real-time and accurate data support for production scheduling and decision-making.

[0033] The present invention also provides a method for local updating of high-precision maps of open-pit mines. The method adopts the above system and includes the following steps: S1. Collecting multi-source heterogeneous real-scene 3D data at a frequency not lower than the production process cycle;

[0034] S2. Register the current data with the historical real-world 3D model, and use a deep learning network to automatically extract the changed areas;

[0035] S3. Perform local point cloud densification and real-scene 3D reconstruction only on the changed areas to generate a local real-scene 3D model;

[0036] S4. The local real-scene 3D model is seamlessly integrated with the historical global model through a graph optimization algorithm, and a new version of the global real-scene 3D model is generated.

[0037] S5. Store the new version of the global real-scene 3D model into the spatiotemporal database and manually confirm it through a visual interaction and verification platform;

[0038] S6. Based on the updated global real-world 3D model, automatically calculate the current mining and stripping volume, and generate progress comparison reports and statistical charts.

[0039] Furthermore, step S2 further includes:

[0040] A cross-modal fusion network of PointNet++ and U-Net is used to output a change probability map in the joint feature space of point cloud and image. Morphological filtering, clustering and vectorization are performed on the probability map to generate closed change polygons, and the change polygons are used as the ROI for subsequent local reconstruction.

[0041] Furthermore, the graph optimization algorithm in step S4 includes:

[0042] The local model and the historical model are aligned using edge-face constraints through pose graph optimization.

[0043] The Laplacian mesh deformation algorithm is used to eliminate splicing seams, so that the splicing error is ≤5cm.

[0044] Compared with the prior art, the present invention has the following advantages: 1. The update cycle is compressed from "day-level" to "process-level": The present invention uses high-frequency acquisition (10-20Hz) with a production process cycle time of ≥ to complete local mapping within 20 minutes, realizing "updating while mining", which meets the rigid requirement of unmanned mining trucks for timeliness of ≤30 minutes.

[0045] 2. Computation and transmission costs are reduced by an order of magnitude: This invention only performs local reconstruction on 1%–5% of the real change area (ROI), reducing GPU memory usage by 75%, reducing reconstruction time from 2 hours to 12–15 minutes, and saving more than 90% of data upload bandwidth.

[0046] 3. The splicing accuracy reaches the centimeter level, meeting the mining surveying specifications: adopting the dual constraints of "pose diagram optimization + Laplacian mesh deformation", 200 sets of field tests show that the plane error is ≤3cm and the elevation is ≤5cm, which is better than the 5cm limit requirement of the "Open-pit Mine Engineering Surveying Specification" GB / T50228-2022 for 1:500 mining and stripping acceptance maps.

[0047] 4. False detection rate continues to decline, and change detection reliability is ≥88%: The PointNet++-U-Net cross-modal network is integrated, with mIoU ≥0.88; the error correction results are fed back into the training pool online, and the false detection rate is reduced by 8% week-on-week, solving the pain points of missed detections and many false detections in traditional manual mapping.

[0048] 5. Real-time measurement of mining and stripping volume with an accuracy of 1.2%: Based on the incremental TIN comparison method, the current volume is automatically output, and the difference with the belt scale measurement value is only 0.9%–1.2%. No work stoppage is required for measurement, which can reduce the number of surveyors by 2 per shift and save about 600,000 yuan in labor costs per year.

[0049] 6. Version rollback in seconds ensures continuous safety of autonomous driving: Merkle tree differential storage makes rollback files <100kB, and can restore to any historical version in 1.3 seconds; combined with Redis distributed write lock + WebSocket broadcast, the scheduling center, vehicle terminal and Web terminal are consistent, avoiding the security risks caused by the inconsistency between "map and vehicle".

[0050] 7. A single system covers the entire chain of "collection-detection-reconstruction-management-verification-quantity calculation": breaking the previous isolated mode of multiple software serialization and repeated format conversion, forming a mine-specific closed loop, truly upgrading the real-scene 3D model from "offline results" to "production factors", providing a real-time data foundation for digital twin mines, intelligent scheduling, and blasting optimization. Attached Figure Description

[0051] Figure 1 A flowchart illustrating the operation of this invention. Detailed Implementation

[0052] The implementation of the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0053] like Figure 1 As shown, a high-precision map local update system for open-pit mines includes a multi-source heterogeneous data acquisition module, which is used to acquire multi-source real-scene 3D data of the open-pit mine mining face at a frequency no less than the mining production process cycle.

[0054] The intelligent change detection and region extraction module is communicatively connected to the multi-source heterogeneous data acquisition module. It is used to automatically identify and extract areas where the working face is continuously excavated and the terrain is rapidly changing, with a positioning uncertainty threshold of ≤1m, under the point cloud-image-semantic multimodal data fusion framework.

[0055] The incremental 3D reconstruction and update module is communicatively connected to the intelligent change detection and region extraction module. It is used to perform local real-scene 3D model reconstruction only on the changed region, and to use graph optimization and seamless stitching algorithms to fuse the newly generated local 3D model with the historical global 3D model to generate an updated global real-scene 3D model.

[0056] The spatiotemporal database and version management module is used to perform hierarchical storage, version management, and incremental rollback of the global real-scene 3D model.

[0057] A visualization, interaction and verification platform is used to perform 3D visualization, human-computer interaction verification and error statistics on the updated global real-scene 3D model.

[0058] The automatic calculation module for mining and stripping work volume is used to automatically calculate the current mining and stripping volume based on the updated global real-scene 3D model, and generate progress comparison reports and statistical charts.

[0059] The multi-source heterogeneous data acquisition module includes:

[0060] The vehicle-mounted / roadside mobile data acquisition subsystem consists of a lidar, a camera, and a GNSS / IMU integrated navigation unit. It is installed on unmanned mining trucks, patrol vehicles, or fixed pole positions to collect raw point cloud and image data of the work surface and surrounding environment at a frequency of 10Hz-20Hz.

[0061] Hardware: 32-line LiDAR ×1, 4K industrial camera ×2, GNSS / IMU integrated navigation unit ×1, overall protection level IP67;

[0062] Installation location: front bumper of unmanned mining truck, roof of patrol vehicle or 4m high on utility pole, ensuring 360° unobstructed view;

[0063] Data acquisition frequency: 10Hz-20Hz, a single trip can cover a 2km work area;

[0064] Raw data: synchronously recorded point cloud (ROSsensor_msgs / PointCloud2), image (compressed JPEG), GNSS / IMU pose (binaryRMS<2cm).

[0065] The UAV aerial survey subsystem is used to perform oblique photogrammetry on the macroscopic area of ​​the mining face and spoil heap to obtain centimeter-level topographic change data. It adopts an oblique five-camera array with a ground resolution of 1.5cm and a single flight area of ​​0.5km². It uses RTK / PPK fusion, with a plane error of ≤3cm and an elevation error of ≤5cm for image control points. It complements the vehicle-mounted data for updating macroscopic topography and spoil heaps.

[0066] The production operation system interface is used to obtain the real-time operating positions, planned paths, and semantic information of excavators and bulldozers from the mine scheduling system via API, and serves as a prior constraint for change detection. The RESTful API captures WGS84 coordinates, planned paths, and "excavation / piling" semantic tags transmitted back from the excavator and bulldozer's Beidou terminals in real time; as a prior constraint, this reduces the change detection search space by 60%, lowering false detections.

[0067] The intelligent change detection and region extraction module includes:

[0068] The multimodal registration unit is used to complete sub-pixel-level registration between current data and historical models based on the initial GNSS / IMU pose value and the NDT-ICP joint algorithm. Specifically, it is used as follows:

[0069] Input: Current vehicle / drone data + historical global 3D model;

[0070] Method: Initial values ​​provided by GNSS / IMU → NDT-ICP joint fine registration, convergence threshold 5mm;

[0071] Output: Subpixel-level extrinsic parameters, providing pixel-to-voxel level alignment for subsequent deep learning.

[0072] The variable semantic segmentation unit is used to output a variable pixel / voxel level probability map in the joint feature space of point cloud and image by employing a cross-modal deep learning network that integrates PointNet++ and U-Net.

[0073] Network architecture: PointNet++ (point cloud branch) + U-Net (image branch) + cross-modal attention fusion layer;

[0074] Training data: 120GB of manually labeled samples (change / no change binary labels) from three phases of this mine.

[0075] Output: Joint feature space probability map, single card inference time 0.3s, mIoU≥0.88.

[0076] The region aggregation and contour vectorization unit is used to perform morphological filtering, clustering, and Ramer-Douglas-Peucker vectorization on the probability map to generate closed variable polygons, which are then used as the input ROI of the incremental 3D reconstruction and update module. Specifically, morphological opening and closing operations remove speckle; DBSCAN clustering is performed with a minimum cluster size of 20 pixels; and the Ramer-Douglas-Peucker algorithm is used for thinning to generate closed variable polygons (ROIs) with a positioning uncertainty ≤1m.

[0077] The incremental 3D reconstruction and update module includes:

[0078] Local point cloud densification unit is used to perform dense reconstruction of image sequences within a change region based on SGM or MVS algorithms to obtain local point clouds with a grid resolution of ≤5cm; SGM (semi-global matching) or MVS (multi-view stereo) is performed only within the ROI, with a grid resolution of ≤5cm.

[0079] Compared to full-scene reconstruction, GPU memory usage is reduced by 75%, and reconstruction time is reduced from 2 hours to 15 minutes.

[0080] The triangular mesh (Mesh) or 3D Gaussian sputtering (3DGS) reconstruction unit is used to convert local point clouds into realistic 3D models; the triangular mesh (Mesh) + Poisson reconstruction has a face count of <5 million; the 3D Gaussian sputtering (3DGS) explicit radiation field is suitable for WebGL 30FPS rendering; users can switch between them with one click in the configuration interface.

[0081] The graph optimization and fusion unit is used to align the local real-world 3D model with the historical global model using a pose graph optimization method with edge-face constraints, and to eliminate seams using a Laplacian-based mesh deformation algorithm to achieve centimeter-level stitching error.

[0082] Pose graph optimization: Establish edge-face constraints between the local model and the historical model, with vertices representing camera / laser frame poses and edges representing ICP relative poses, using the optimizer g2o;

[0083] Laplacian mesh deformation: Smooth the 0.5m buffer zone in the overlapping area to eliminate seams;

[0084] Splicing error: Based on 200 sets of on-site tests, the horizontal plane error is ≤3cm and the vertical plane error is ≤5cm.

[0085] The spatiotemporal database and version management module includes:

[0086] A hierarchical storage engine is used to store real-world 3D models in layers according to spatial grids, feature categories, and timestamps, and adopts a hybrid architecture of MongoDB + object storage.

[0087] Spatial grid: GeoHash length = 8 (approximately 20m × 40m);

[0088] Element categories: Surface Mesh, Vegetation, Roads, Equipment, Rock Mass;

[0089] Timestamp: UTC milliseconds;

[0090] Architecture: MongoDB stores metadata and grid indexes, and object storage (MinIO / S3) stores binary models, with read / write latency <200ms.

[0091] The version chain management unit is used to generate a globally unique version ID for each incremental update and record feature-level differences based on a Merkle tree structure to support quick rollback to any historical version; each incremental update generates a globally unique UUID version number; feature-level differences (addition, deletion, modification) are recorded based on a Merkle tree, and the difference file is <100kB;

[0092] Supports second-level rollback: only pulls the differences and reverses them, without rewriting the entire model.

[0093] A concurrent locking mechanism is used to apply write locks to affected regions during model updates and to notify all subscribed nodes via WebSocket event broadcast after the update is complete. A write lock (Redis distributed lock) is applied to the affected GeoHash grid, and the update is broadcast via WebSocket to all subscribed nodes (scheduling center, mobile, and web) to ensure consistency across multiple platforms.

[0094] The visual interaction and verification platform includes:

[0095] The WebGL 3D rendering engine is used to render real-world 3D models with a range of ≥10km² at a frame rate of no less than 30FPS; it features self-developed Level-of-Detail (LOD) scheduling, with a frame rate of ≥30FPS for a 10km² scene; and it supports mining-specific tools such as terrain profiles, volumetric selection, and explosion point analysis.

[0096] The multi-phase differential overlay unit is used to highlight and overlay the excavation range of multiple phases with different colors, intuitively presenting the evolution trajectory and morphological changes of the working face; specifically, different versions are dynamically overlaid according to the time axis, and the changed areas are highlighted (red = newly added, blue = removed), making the trajectory clear at a glance.

[0097] Human-computer interaction tools are used to support operators in manually confirming or correcting changed areas by clicking, selecting boxes, or inputting voice, and to feed the correction results back to the intelligent change detection and region extraction module for online retraining; specifically;

[0098] Point-and-click / box-select error correction: The task manager can directly mark "false positive" or "false negative" in the browser;

[0099] Voice input: Supports natural language commands such as "Delete the area 5m north".

[0100] The error correction results are automatically fed back into the training pool, enabling the model to learn incrementally online, and the false positive rate is reduced by 8% week-on-week.

[0101] The automatic calculation module for the amount of mining and stripping work includes:

[0102] The volume calculation unit is used to automatically calculate the current mining and stripping volume based on the updated high-precision real-scene 3D model using the triangular network of irregularities (TIN) comparison method; based on the updated high-precision TIN, it uses the "previous period - current period" elevation difference method to eliminate slope anomalies >60°; it supports custom rock mass density of 2.7t / m³ and directly outputs the tonnage, with the accuracy measured and compared with the total station, the relative error is <1.2%.

[0103] The progress comparison report unit is used to generate reports and statistical charts on differences in project quantities across multiple periods; it automatically generates shift reports, daily reports, and monthly reports, including bar charts of volume, tonnage, planned completion rate, and over- or under-excavation; and it supports one-click export to PDF and Excel.

[0104] The interactive query unit allows users to interactively query model information and engineering quantity data at any location and time period, providing real-time and accurate data support for production scheduling and decision-making. Clicking on any location will bring up historical elevation and volume change curves for that location; it can also interface with the mine scheduling system to provide real-time data services for truck scheduling and blasting design.

[0105] The method for local updating of high-precision maps of open-pit mines includes the following steps:

[0106] S1. Collect multi-source heterogeneous real-world 3D data at a frequency no less than the production cycle time;

[0107] S2. Register the current data with the historical real-world 3D model, and use a deep learning network to automatically extract the changed areas;

[0108] S3. Perform local point cloud densification and real-scene 3D reconstruction only on the changed areas to generate a local real-scene 3D model;

[0109] S4. The local real-scene 3D model is seamlessly integrated with the historical global model through a graph optimization algorithm, and a new version of the global real-scene 3D model is generated.

[0110] S5. Store the new version of the global real-scene 3D model into the spatiotemporal database and manually confirm it through a visual interaction and verification platform;

[0111] S6. Based on the updated global real-world 3D model, automatically calculate the current mining and stripping volume, and generate progress comparison reports and statistical charts.

[0112] Furthermore, step S2 includes:

[0113] A cross-modal fusion network of PointNet++ and U-Net is used to output a change probability map in the joint feature space of point cloud and image. Morphological filtering, clustering and vectorization are performed on the probability map to generate closed change polygons, and the change polygons are used as the ROI for subsequent local reconstruction.

[0114] The graph optimization algorithm in step S4 includes:

[0115] The local model and the historical model are aligned using edge-face constraints through pose graph optimization.

[0116] The Laplacian mesh deformation algorithm is used to eliminate splicing seams, so that the splicing error is ≤5cm.

[0117] The following description uses two specific examples.

[0118] Example 1: Data collection using unmanned mining trucks, specifically including the following steps:

[0119] (1) Data collection: The unmanned mining truck was equipped with the vehicle-mounted subsystem and made two round trips along the designed road, taking 20 minutes to collect 18GB of raw point cloud and 5GB of images.

[0120] (2) Change detection: After registration and PointNet++-U-Net network inference, three change polygons were extracted, with a total area of ​​12,000 m².

[0121] (3) Local reconstruction: SGM densification was performed within the ROI to generate 110 million points of 5cm grid point cloud, and the reconstruction time was 12 minutes;

[0122] (4) Fusion: After image optimization, the stitching error is 2.8cm, and a new version of the global model is generated;

[0123] (5) Quantity calculation: The current mining and stripping volume is 34,500 m³, equivalent to 93,200 tons, which differs from the data of the belt scale at the mine metering station by 0.9%;

[0124] (6) Rollback test: Randomly roll back to the version 7 days ago, which takes 1.3 seconds. The model, reports and difference charts are all restored correctly.

[0125] Example 2: Data collection using drones:

[0126] Drones were used to conduct oblique photography in the spoil heap area, covering a change area of ​​50,000 m². The local reconstruction used the 3DGS radiation field mode, and the web client rendered smoothly at 30 FPS. The work supervisor could input "delete the eastern road area" via voice, and the system would provide real-time feedback and update the model to meet the needs of nighttime scheduling.

Claims

1. A high-precision map local update system for open-pit mines, characterized in that, The system includes: The multi-source heterogeneous data acquisition module is used to acquire multi-source real-scene 3D data of the open-pit mine excavation face at a frequency no less than the cycle time of the mining production process. The intelligent change detection and region extraction module is communicatively connected to the multi-source heterogeneous data acquisition module. It is used to automatically identify and extract areas where the working face is continuously excavated and the terrain is rapidly changing, with a positioning uncertainty threshold of ≤1m, under the point cloud-image-semantic multimodal data fusion framework. The incremental 3D reconstruction and update module is communicatively connected to the intelligent change detection and region extraction module. It is used to perform local real-scene 3D model reconstruction only on the changed region, and to use graph optimization and seamless stitching algorithms to fuse the newly generated local 3D model with the historical global 3D model to generate an updated global real-scene 3D model. The spatiotemporal database and version management module is used to perform hierarchical storage, version management, and incremental rollback of the global real-scene 3D model. A visualization, interaction and verification platform is used to perform 3D visualization, human-computer interaction verification and error statistics on the updated global real-scene 3D model. The automatic calculation module for mining and stripping work volume is used to automatically calculate the current mining and stripping volume based on the updated global real-scene 3D model, and generate progress comparison reports and statistical charts.

2. The open-pit mine high-precision map local update system according to claim 1, characterized in that, The multi-source heterogeneous data acquisition module includes: The vehicle-mounted / roadside mobile data acquisition subsystem consists of a lidar, a camera, and a GNSS / IMU integrated navigation unit. It is installed on unmanned mining trucks, patrol vehicles, or fixed pole positions to collect raw point cloud and image data of the work surface and surrounding environment at a frequency of 10Hz-20Hz. The UAV aerial survey subsystem is used to perform oblique photogrammetry on the macroscopic area of ​​the mining face and spoil heap to obtain centimeter-level topographic change data; The production operation system interface is used to obtain the operating location, planned path, and operation semantic information of excavators and bulldozers in real time from the mine scheduling system through the API interface, and to serve as a prior constraint for change detection.

3. The open-pit mine high-precision map local update system according to claim 1, characterized in that, The intelligent change detection and region extraction module includes: A multimodal registration unit is used to complete sub-pixel-level registration between current data and historical models based on GNSS / IMU pose initial values ​​and the NDT-ICP joint algorithm; The variable semantic segmentation unit is used to output a probability map at the variable pixel / voxel level in the joint feature space of point cloud and image by employing a cross-modal deep learning network that integrates PointNet++ and U-Net. The region aggregation and contour vectorization unit is used to perform morphological filtering, clustering, and Ramer-Douglas-Peucker vectorization on the probability map to generate closed variable polygons, and to use the variable polygons as the input ROI of the incremental 3D reconstruction and update module.

4. The open-pit mine high-precision map local update system according to claim 1, characterized in that, The incremental 3D reconstruction and update module includes: Local point cloud densification unit is used to perform dense reconstruction of image sequences within a changing region based on SGM or MVS algorithms to obtain local point clouds with a grid resolution of ≤5cm. Triangular mesh (Mesh) or 3D Gaussian sputtering (3DGS) reconstruction units are used to convert local point clouds into realistic 3D models; The graph optimization fusion unit is used to align the local real-world 3D model with the historical global model using a pose graph optimization method with edge-face constraints, and to eliminate seams using a Laplacian-based mesh deformation algorithm to achieve centimeter-level stitching error.

5. The high-precision map local update system for open-pit mines according to claim 1, characterized in that, The spatiotemporal database and version management module includes: The hierarchical storage engine is used to store real-world 3D models in layers according to spatial grids, feature categories, and timestamps, and adopts a hybrid architecture of MongoDB + object storage. The version chain management unit is used to generate a globally unique version ID for each incremental update and record feature-level differences based on a Merkle tree structure to support quick rollback to any historical version. The concurrent locking mechanism is used to apply write locks to the affected regions during model updates and to notify all subscribed nodes via WebSocket event broadcast after the update is completed.

6. The open-pit mine high-precision map local update system according to claim 1, characterized in that, The visualization interaction and verification platform includes: The WebGL 3D rendering engine is used to render real-world 3D models with a range of ≥10km² at a frame rate of no less than 30FPS. Multi-phase differential overlay unit is used to highlight and overlay the excavation range of multiple phases with different colors, intuitively presenting the evolution trajectory and morphological changes of the working face; The human-computer interaction tool is used to support operators to manually confirm or correct changes in areas by clicking, selecting boxes, or using voice input, and to feed back the correction results to the intelligent change detection and region extraction module for online retraining.

7. The open-pit mine high-precision map local update system according to claim 1, characterized in that, The automatic calculation module for the amount of mining and stripping work includes: The volume calculation unit is used to automatically calculate the current mining and stripping volume based on the updated high-precision real-scene 3D model and the triangular irregular network (TIN) comparison method. The progress comparison report unit is used to generate multi-phase project quantity difference reports and statistical charts; The interactive query unit supports users in interactively querying model information and engineering quantity data at any location and time period, providing real-time and accurate data support for production scheduling and decision-making.

8. A method for local updating of high-precision maps of open-pit mines, characterized in that, The system described in any one of claims 1-7 is specifically used in a method comprising the following steps: S1. Acquiring multi-source heterogeneous real-world 3D data at a frequency not less than the production cycle time; S2. Register the current data with the historical real-world 3D model, and use a deep learning network to automatically extract the changed areas; S3. Perform local point cloud densification and real-scene 3D reconstruction only on the changed areas to generate a local real-scene 3D model; S4. The local real-scene 3D model is seamlessly integrated with the historical global model through a graph optimization algorithm, and a new version of the global real-scene 3D model is generated. S5. Store the new version of the global real-scene 3D model into the spatiotemporal database and manually confirm it through a visual interaction and verification platform; S6. Based on the updated global real-world 3D model, automatically calculate the current mining and stripping volume, and generate progress comparison reports and statistical charts.

9. The method for local updating of high-precision maps of open-pit mines according to claim 8, characterized in that, Step S2 further includes: A cross-modal fusion network of PointNet++ and U-Net is used to output a change probability map in the joint feature space of point cloud and image. Morphological filtering, clustering and vectorization are performed on the probability map to generate closed change polygons, and the change polygons are used as the ROI for subsequent local reconstruction.

10. The method for local updating of high-precision maps of open-pit mines according to claim 8, characterized in that, The graph optimization algorithm in step S4 includes: The local model and the historical model are aligned using edge-face constraints through pose graph optimization. The Laplacian mesh deformation algorithm is used to eliminate splicing seams, so that the splicing error is ≤5cm.

Images