Surface mine live-action modeling system and method
By combining intelligent drone data collection and 3DGS rendering technology with a closed-loop data quality control system, the problems of low efficiency and uncontrollable quality in open-pit mine modeling have been solved. This has enabled the generation of high-precision, real-time interactive 3D models, which are suitable for dynamic management and safe production in open-pit mines.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA NO 15 METALLURGICAL CONSTR GRP
- Filing Date
- 2026-02-04
- Publication Date
- 2026-05-01
AI Technical Summary
Existing open-pit mine modeling technology is inefficient, slow to update, has uncontrollable model quality, presents a contradiction between rendering efficiency and accuracy, lacks an automated quality control loop, and is difficult to generate high-precision, real-time interactive 3D models.
By employing intelligent drone data acquisition, 3DGS rendering technology, and closed-loop data quality control, combined with multi-sensor data fusion, we achieve automated data acquisition, real-time transmission, and efficient modeling. We construct a high-precision model using a 3D Gaussian ellipsoid and introduce geological semantic constraints.
It generates efficient and automated high-precision realistic 3D models, supports real-time interaction, solves the problems of low efficiency and uncontrollable quality in traditional modeling, and improves the geometric accuracy and geological rationality of the models.
Smart Images
Figure CN121962497A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of mine surveying, computer vision and 3D reality modeling technology, specifically a real-scene modeling system and method for open-pit mines. Background Technology
[0002] With the increasing demands for safe production, efficient resource extraction, and environmental protection in open-pit mines, the reliance on high-precision, high-timeliness realistic 3D models is deepening. Traditional mine modeling mainly relies on manual ground surveying or airborne lidar, which suffers from low efficiency, high cost, and long model update cycles, making it difficult to meet the needs of modern dynamic mine management.
[0003] In recent years, UAV oblique photogrammetry technology has been widely used. However, this method has the following inherent drawbacks in mining applications:
[0004] First, the existing modeling process has a low degree of automation. Flight mission planning, data collection, and recovery rely heavily on manual operation and judgment, making it difficult to achieve high-frequency, routine data updates and respond promptly to dynamic changes in the mine.
[0005] Secondly, the quality of models generated by existing modeling methods is overly dependent on scene textures. For typical mining scenes with weak textures and many repetitive features, problems such as model voids, distortions, and geometric distortions can easily occur, making it difficult to accurately depict the fine structure of key facilities such as steps and slopes.
[0006] Furthermore, existing technologies suffer from an inherent contradiction between rendering efficiency and model accuracy. While ensuring high accuracy, the massive triangular mesh models generated by oblique photogrammetry result in huge data volumes, making it difficult to achieve real-time, smooth rendering and interaction on web or mobile devices, which severely restricts the immediate application and sharing of the models.
[0007] Furthermore, the data acquisition, transmission and processing are relatively independent of each other, lacking an integrated and automated quality control loop, which leads to uncontrollable quality of the final model and low reliability of the modeling results.
[0008] 3DGS, an emerging radiation field rendering technology proposed in 2023, represents a scene using a series of Gaussian ellipsoids in three-dimensional space. It can train high-quality, real-time-renderable 3D models from a sparse set of input images. Compared to traditional Neural Radiation Field (NeRF), 3DGS boasts significantly faster training and rendering speeds. However, current research and applications of this technology primarily focus on small-scale, enclosed indoor and outdoor scenes. There are currently no systematically adapted, optimized, and engineered solutions for application scenarios such as open-pit mines—large-scale environments with complex terrain and geology, rapid dynamic changes, and unique texture features.
[0009] Therefore, there is an urgent need in this field for an open-pit mine real-scene modeling system and method that can achieve full automation, high precision, high efficiency, and generate high-fidelity models that support real-time interaction, in order to overcome the many shortcomings of existing technologies. Summary of the Invention
[0010] The main objective of this invention is to solve the problems existing in the prior art and provide an open-pit mine real-scene modeling system and method.
[0011] The technical solution of the present invention is as follows: an open-pit mine real-scene modeling system, comprising...
[0012] Intelligent data acquisition unit: used to automatically acquire oblique photogrammetric image data and laser point cloud data of the target mining area according to preset tasks;
[0013] Data communication and transmission unit: used to stably, in real time and completely transmit the data collected by the intelligent data acquisition unit back to the local data processing center;
[0014] Data preprocessing and quality assessment unit: used to automatically preprocess the returned data and assess the data quality based on preset quality indicators, forming a closed loop of quality control;
[0015] Real-scene 3D modeling engine unit: Based on 3D Gaussian technology and integrating multi-source data and mine semantic constraints, it can quickly generate high-precision and real-time rendering real-scene 3D models of open-pit mines.
[0016] Furthermore, the intelligent data acquisition unit includes:
[0017] At least one drone airport deployed at the mining site for the automatic take-off and landing, charging, and data relay of the drones;
[0018] At least one drone is equipped with a multi-sensor payload, which includes a high-definition visible light camera, an RTK module, and a lidar.
[0019] The mission planning module is used to automatically generate optimized flight paths and sensor parameters based on the digital elevation model, geological map, mining plan map, and real-time meteorological data of the target mining area.
[0020] Furthermore, the data communication and transmission unit includes:
[0021] Intelligent transmission module: Deployed at the edge preprocessing gateway, it supports compression, packetization and adaptive transmission of collected data;
[0022] Local Area Network (LAN): Deploy a local area network to achieve efficient data backhaul;
[0023] Local data center / server: Receives and stores data transmitted back from drones in real time.
[0024] Furthermore, the data preprocessing and quality assessment unit includes:
[0025] Data preprocessing module: used to perform uniform illumination and color matching, distortion correction on the returned raw image data, and denoising and registration on the point cloud data;
[0026] Quality assessment module: Used to automatically assess data quality based on indicators such as sharpness, overlap, and color consistency, through a trained convolutional neural network model;
[0027] Quality control closed-loop module: When the evaluation data is unqualified, it automatically generates a reflight or reacquisition task instruction and sends it to the intelligent data acquisition unit.
[0028] Furthermore, the real-scene 3D modeling engine unit includes:
[0029] Initial Spatial Point Cloud Generation Module: Used to recover the camera pose from oblique photogrammetric images using SfM and generate sparse point clouds;
[0030] 3DGS scene initialization module: Constructs an initial distribution of a 3D Gaussian ellipsoid based on the initial point cloud;
[0031] Iterative training module: continuously optimizes the Gaussian parameters through differentiable rendering and parameter optimization;
[0032] 3DGS Model Output Module: Generates the final model that supports real-time rendering on both web and mobile devices.
[0033] The present invention also provides a method for real-scene modeling of open-pit mines. Using the above system, the specific method includes the following steps: S1, collecting oblique photogrammetric image data and laser point cloud data of the target mining area to construct the initial spatial point cloud of the mining area;
[0034] S2. Based on the mining area DEM, mining status and real-time meteorological data, generate an adaptive flight acquisition mission and transmit data back through the communication link;
[0035] S3. Perform preprocessing and quality assessment on the returned data; if it fails to meet the requirements, trigger the re-sampling mechanism.
[0036] S4. Initialize a 3D Gaussian scene using the initial spatial point cloud, and perform iterative training through differentiable rendering and parameter optimization.
[0037] S5. Introduce geological semantic constraints during training to enhance the geological rationality and continuity of the model;
[0038] S6. Outputs the trained 3DGS model, supporting high frame rate rendering and interaction on both web and mobile devices.
[0039] Furthermore, step S1 involves constructing the initial spatial point cloud of the mining area, including:
[0040] Control a drone equipped with multiple sensor payloads to execute a preset route and a preset number of flights over the target mining area to obtain oblique photogrammetry image sequences and laser point cloud datasets;
[0041] Feature extraction and matching are performed on the oblique photographic image sequence, and camera parameters are calculated using motion reconstruction structure to generate sparse point clouds;
[0042] Trajectory and point cloud calculations are performed on the laser point cloud dataset to obtain a high-precision three-dimensional coordinate point cloud.
[0043] Furthermore, step S4 involves iterative training through differentiable rendering and parameter optimization, including:
[0044] 3D Gaussian initialization: Based on the initial spatial point cloud, a 3D Gaussian ellipsoid is created for each 3D point, with parameters including position, covariance matrix, opacity and spherical harmonic coefficients;
[0045] Differentiable rendering: 3D Gaussian projection is applied to a 2D image plane, and alpha blending is performed using a tile-based rasterizer.
[0046] Parameter optimization: Calculate the loss function between the rendered image and the real image, and update the Gaussian parameters through backpropagation;
[0047] Adaptive density control: Periodically perform cloning, splitting, and pruning operations to optimize the distribution of Gaussian points and model details.
[0048] Compared with existing technologies, this invention has the following advantages: it combines automated UAV data acquisition, high-bandwidth data transmission, and cutting-edge 3DGS rendering technology for application in the field of mine real-world modeling. Specifically, it achieves unmanned and adaptive data acquisition through UAV airports and intelligent task planning, solving the problems of low efficiency and slow updates in traditional methods. By constructing a closed-loop system including data quality control, it ensures the high quality and reliability of input data. Through innovative deep integration of laser point clouds and the 3DGS training process, it significantly improves the geometric accuracy and training speed of the model in weakly textured areas of mines using precise geometric priors. By introducing semantic constraints related to mines, the generated model not only possesses visual fidelity but also contains professional geological logic. Ultimately, this invention can efficiently and automatically generate high-precision real-world 3D models that support real-time interaction, providing a powerful data foundation for mine digital twins, safe production, and intelligent decision-making. Attached Figure Description
[0049] Figure 1 This is a schematic diagram of the overall architecture of the system of the present invention; Figure 2 This is a flowchart of the open-pit mine real-scene modeling method of the present invention. Detailed Implementation
[0050] 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.
[0051] like Figure 1 As shown, the open-pit mine real-scene modeling system of this invention realizes fully automatic, high-precision, and real-time construction of open-pit mine real-scene models by constructing a technical system that integrates intelligent data acquisition, reliable transmission, quality control closed loop, and high-performance modeling.
[0052] The system mainly consists of four units: intelligent data acquisition unit, data communication and transmission unit, data preprocessing and quality assessment unit, and realistic 3D modeling engine unit. Each unit will be described in detail below.
[0053] Intelligent data acquisition unit: This is the data source of the system. Its core consists of a drone airport deployed at the mining site, a swarm of drones equipped with multiple sensors, and a mission planning module.
[0054] The mission planning module integrates and analyzes digital elevation models, geological maps, mining plans, and real-time meteorological information obtained from weather stations in the mining area to automatically generate flight paths and sensor parameters that can avoid obstacles, adapt to terrain undulations, and optimize data collection efficiency.
[0055] Data communication and transmission unit: This unit ensures the reliability of the data link from the field to the data processing center. The system deploys a dedicated local area network within the mining area.
[0056] The intelligent transmission module, deployed on drones and edge gateways, performs real-time or near-real-time segmentation and compression of large-capacity imagery and point cloud data, and features adaptive transmission and breakpoint resume capabilities. Data transmission is completed via a local area network, ensuring complete and continuous transmission to the local data processing center.
[0057] Data Preprocessing and Quality Assessment Unit: This unit is deployed on an edge server or in the cloud. Its data preprocessing module performs operations such as light and color homogenization and lens distortion correction on the returned raw images, and performs noise reduction and image position registration on the laser point cloud data.
[0058] Subsequently, the quality assessment module uses a convolutional neural network model trained based on deep learning to automatically evaluate key indicators such as clarity, overlap, and color consistency of the preprocessed data.
[0059] If the evaluation results do not meet the preset quality threshold, the quality control closed-loop module will automatically generate a re-flight or re-collection task instruction for the non-conforming area and send it to the intelligent data acquisition unit for execution, thereby forming a strict, end-to-end quality control closed loop.
[0060] Real-scene 3D modeling engine unit: This unit is the core innovation of this invention. It is built based on 3DGS technology and has been deeply optimized for mining scenes.
[0061] Its workflow begins with the initial spatial point cloud generation module. This module first uses Structure for Motion Restoration (SfM) technology to recover the camera pose from oblique photogrammetric images and generate a sparse point cloud. To improve the accuracy of the initial geometry, especially in weakly textured regions, this invention precisely registers and fuses the SfM sparse point cloud with the high-precision 3D point cloud acquired by LiDAR, forming a denser and more accurate initial spatial point cloud.
[0062] Subsequently, the process enters the core training loop of 3DGS. The 3D Gaussian scene initialization module initializes millions of three-dimensional Gaussian ellipsoids based on the above-mentioned fused point cloud. Each Gaussian ellipsoid has parameters such as position, covariance, opacity, and spherical harmonic coefficients.
[0063] In the iterative training module, differentiable rendering and parameter optimization are performed:
[0064] Differentiable rendering: For each training viewpoint, all 3D Gaussian ellipsoids are projected onto a 2D image plane and rendered through a fast, tile-based rasterizer to synthesize a prediction image.
[0065] Parameter optimization: Calculate the loss function between the predicted image and the real captured image, and use the backpropagation algorithm to iteratively optimize the parameters of all Gaussians, so that the rendering result continuously approaches the real photo.
[0066] During the optimization process, the system periodically performs adaptive density control to dynamically manage the distribution of Gaussian points. It performs cloning or splitting operations on Gaussians that are too large or too small to increase detail, and performs pruning operations on Gaussians with too low opacity to optimize storage and computation efficiency.
[0067] like Figure 2 As shown, based on the above system, the open-pit mine real-scene modeling method of this embodiment includes the following steps:
[0068] S1. Collect oblique photogrammetric image data and laser point cloud data of the target mining area to construct the initial spatial point cloud of the mining area;
[0069] S2. Based on the mining area DEM, mining status and real-time meteorological data, generate an adaptive flight acquisition mission and transmit data back through the communication link;
[0070] S3. Perform preprocessing and quality assessment on the returned data; if it fails to meet the requirements, trigger the re-sampling mechanism.
[0071] S4. Initialize a 3D Gaussian scene using the initial spatial point cloud, and perform iterative training through differentiable rendering and parameter optimization.
[0072] S5. Introduce geological semantic constraints during training to enhance the geological rationality and continuity of the model;
[0073] S6. Outputs the trained 3DGS model, supporting high frame rate rendering and interaction on both web and mobile devices.
[0074] Furthermore, step S1 involves constructing the initial spatial point cloud of the mining area, including:
[0075] Control a drone equipped with multiple sensor payloads to execute a preset route and a preset number of flights over the target mining area to obtain oblique photogrammetry image sequences and laser point cloud datasets;
[0076] Feature extraction and matching are performed on the oblique photographic image sequence, and camera parameters are calculated using motion reconstruction structure to generate sparse point clouds;
[0077] Trajectory and point cloud calculations are performed on the laser point cloud dataset to obtain a high-precision three-dimensional coordinate point cloud.
[0078] Furthermore, step S4 involves iterative training through differentiable rendering and parameter optimization, including:
[0079] 3D Gaussian initialization: Based on the initial spatial point cloud, a 3D Gaussian ellipsoid is created for each 3D point, with parameters including position, covariance matrix, opacity and spherical harmonic coefficients;
[0080] Differentiable rendering: 3D Gaussian projection is applied to a 2D image plane, and alpha blending is performed using a tile-based rasterizer.
[0081] Parameter optimization: Calculate the loss function between the rendered image and the real image, and update the Gaussian parameters through backpropagation;
[0082] Adaptive density control: Periodically perform cloning, splitting, and pruning operations to optimize the distribution of Gaussian points and model details.
[0083] To fully illustrate the specific implementation methods and technical effects of the present invention, the following uses a large iron ore open-pit mine as an example to describe in detail the complete real-scene modeling operation cycle.
[0084] The iron ore mine is approximately 2.5 kilometers long from east to west and 1.5 kilometers wide from north to south, with multiple mining benches already formed, and the highest slope exceeding 180 meters. The mine is currently in a period of high-intensity mining, and the surface morphology is changing rapidly. Its core requirements are:
[0085] Efficient dynamic updates: A high-precision 3D model of the entire mining area needs to be generated at least once a week to meet the needs of production scheduling and dynamic reserve calculation.
[0086] Detailed slope monitoring: It is necessary to accurately represent the rock mass structure and minute deformations of steep slopes to provide data support for slope stability analysis.
[0087] Instant Model Application: The generated models must be able to load and interact smoothly on ordinary office computers and mobile devices in various departments such as geology, surveying, and safety, eliminating the reliance on professional graphics workstations.
[0088] The system deployment details in this embodiment are as follows: Intelligent data acquisition terminal: A fully automated drone airport is deployed at two commanding heights on the north and south sides of the mining area, forming dual-base station coverage to ensure no blind spots. The airport is equipped with constant temperature and humidity, automatic charging, and 4G / 5G network communication functions.
[0089] Drones and payload: Equipped with two long-endurance mapping drones. Each drone carries:
[0090] GNSS RTK module: Provides centimeter-level real-time positioning.
[0091] 36-megapixel full-frame tilt camera: used to acquire high-definition images from five perspectives (downward, front, back, left, and right).
[0092] Lightweight LiDAR: with a scanning frequency of 240 lines / second, it directly acquires three-dimensional point clouds of the earth's surface.
[0093] Communication network: Data transmission is achieved by utilizing the existing local area network in the mining area.
[0094] Software and Platform: Deploy the open-pit mine real-scene modeling cloud platform described in this invention. This platform integrates all modules for task planning, data transmission, quality assessment, 3DGS modeling engine, and model publishing.
[0095] The specific implementation process is as follows: Phase 1: Adaptive task planning and fully automated data acquisition
[0096] Task Planning: The platform's task planning module imports the latest DEM (Digital Elevation Model) of the mining area, the mining plan red line map, and real-time wind speed and cloud cover data from meteorological stations. The module automatically analyzes the data using algorithms to identify newly added mining areas and key slope monitoring areas for the week. To avoid dust interference, the system schedules flight missions for the early morning when meteorological data is most favorable.
[0097] Output: Two optimized flight paths are generated. The main flight path covers the entire mining area, with a flight altitude of 180 meters, 80% overlap in the heading direction, and 70% overlap in the lateral direction, using a terrain-following flight mode. The other is a refined inspection flight path, targeting high slope areas, with the flight altitude reduced to 80 meters and the overlap increased to 85%-80% to capture more refined texture and geometric information.
[0098] Automated Execution: On the morning of the mission day, the drones at the northern airport automatically take off at a preset time to perform the main data collection task. The platform monitors the drone status and data cache in real time. When flying to the dusty area in the middle of the mining site, the edge preprocessing gateway initiates a real-time lossy compression algorithm for the image data, significantly reducing the transmission pressure while ensuring feature point extraction. The data is stably transmitted to the local data processing center via the local area network.
[0099] Phase Two: Closed-Loop Quality Control and Data Preprocessing
[0100] Automated preprocessing: The local modeling platform performs parallel preprocessing on the 2500 original images returned, including:
[0101] Image lighting and color uniformity: Eliminate color differences in images taken under different conditions and with different lighting conditions.
[0102] Lens distortion correction: Corrects image edge distortion based on camera calibration parameters.
[0103] Point cloud denoising and registration: The LiDAR point cloud is filtered to remove moving noise points such as birds and vehicles, and then it is calculated with RTK / IMU trajectory data to generate an accurate absolute coordinate point cloud.
[0104] Intelligent quality assessment and closed-loop: The quality assessment module calls the pre-trained ResNet-50 convolutional neural network model to scan and analyze the pre-processed image dataset.
[0105] The assessment revealed that the model identified 35 images in the southwest corner of the mining area where the image sharpness index was below the preset threshold due to instantaneous dust. Additionally, the image overlap in this area was slightly insufficient due to the drone's obstacle avoidance maneuvers.
[0106] Triggering the closed loop: The quality control closed-loop module immediately generates a "targeted" supplementary flight task, which targets only the non-compliant areas and plans denser flight paths. The command is issued to the UAV at the southern airport via the platform. After completing its own main flight path task, the UAV automatically takes over and executes this supplementary flight task, successfully acquiring qualified data. The entire process requires no manual intervention.
[0107] Phase 3: Generation of Reality Models Integrating 3DGS
[0108] High-precision initial point cloud generation: The platform first performs SfM calculations on all qualified images to generate a sparse point cloud containing 5.5 million points.
[0109] Subsequently, the multi-source data fusion module registered and fused the SfM sparse point cloud with 120 million high-precision 3D points acquired by LiDAR. The LiDAR point cloud provided accurate geometric support for SfM in weakly textured and shadowed regions, forming an initial spatial point cloud with excellent density and accuracy.
[0110] 3DGS Model Training and Semantic Constraint Injection:
[0111] Initialization: Based on the fused point cloud, the 3D Gaussian scene initialization module creates approximately 900,000 3D Gaussian ellipsoids as the initial distribution.
[0112] Iterative Training: The iterative training module is launched and training is performed on a single NVIDIA A800 GPU server.
[0113] Differentiable rendering: In each iteration step, a 3D Gaussian is projected onto 2D to synthesize an image from the training viewpoint.
[0114] Parameter optimization: Calculate the L1 loss and D-SSIM loss between the synthetic image and the ground image, and optimize all Gaussian parameters through backpropagation.
[0115] Adaptive density control: Density control is performed once every 1000 iterations. Large Gaussians with excessively large gradient accumulation are split; Gaussians with too small a scale are cloned to fill the gaps; and Gaussians with opacity approaching zero are pruned. After 3 hours of training, the number of Gaussians stabilizes at approximately 12 million.
[0116] Semantic Constraint Enhancement: During training, the semantic constraint module loads the mine's "lithological zoning vector map." The module calculates the lithological unit to which each Gaussian point belongs and adds a "semantic consistency constraint term" to the loss function. This constraint term encourages sets of Gaussian points within the same lithological unit to have more similar distributions of spherical harmonic coefficients and covariance matrices. This results in more natural color and texture transitions in the final model within the same lithological strata and effectively reduces model noise and distortion caused by texture loss in homogeneous rock wall areas, enhancing the model's geological plausibility.
[0117] Modeling Results and Comparative Analysis: After training, the 3DGS model output module generated the final realistic 3D model. This model was loaded onto the web using the Three.js engine, and on a typical office computer, the rendering frame rate remained stable at 55-60fps, achieving truly real-time and smooth interaction.
[0118] The model generated by this invention has an overwhelming advantage in detail over models generated by existing technologies:
[0119] Stepped areas: Traditional oblique photogrammetry models show obvious "streaking" and blurring at the edges of steps. In contrast, the step boundaries of the model in this invention are clear and sharp, and even the drilling holes and track marks left by the drilling rig can be clearly identified on the steps.
[0120] Slope area: Traditional models show numerous holes and distortions on rock slopes. The model of this invention completely reconstructs the rock mass structure of the slope, and the undulations of the rock strata and the orientation and dip angle of the main joints and fissures are realistically reproduced.
[0121] Weak texture areas: In large slag heaps, traditional models are severely distorted.
[0122] Data efficiency: Traditional oblique photogrammetry generates OSGB format model data of approximately 45GB. The total size of the 3DGS model file output by this invention is only 1.3GB, reducing the data volume by approximately 97%, which greatly facilitates network transmission and terminal loading.
[0123] This embodiment fully verifies the practicality and advancement of the present invention. The generated realistic 3D model has been used as the core foundation of the "mine digital twin" and applied to:
[0124] Production Management: The system automatically calculates the total amount of mining and stripping and ore reserves every week with an accuracy of over 95%, providing a precise basis for production planning.
[0125] Safety Management: By comparing models from two consecutive weeks, the system automatically identifies and warns of localized minor deformation areas on the slope, providing unprecedented data granularity for slope stability analysis.
[0126] Collaborative sharing: The lightweight model allows personnel from different departments such as geology, surveying, production, and safety to view, annotate, and conduct collaborative consultations in real time on a browser, breaking down the barriers between professional software and data silos.
[0127] This invention successfully solves the core pain points encountered by the mine in the process of intelligent transformation, such as slow data updates, poor model quality, and high application threshold.
Claims
1. A real-scene modeling system for open-pit mines, characterized in that, include Intelligent data acquisition unit: used to automatically acquire oblique photogrammetric image data and laser point cloud data of the target mining area according to preset tasks; Data communication and transmission unit: used to stably, in real time and completely transmit the data collected by the intelligent data acquisition unit back to the local data processing center; Data preprocessing and quality assessment unit: used to automatically preprocess the returned data and assess the data quality based on preset quality indicators, forming a closed loop of quality control; Real-scene 3D modeling engine unit: Based on 3D Gaussian technology and integrating multi-source data and mine semantic constraints, it can quickly generate high-precision and real-time rendering real-scene 3D models of open-pit mines.
2. The open-pit mine real-scene modeling system according to claim 1, characterized in that, The intelligent data acquisition unit includes: At least one drone airport deployed at the mining site for the automatic take-off and landing, charging, and data relay of the drones; At least one drone is equipped with a multi-sensor payload, which includes a high-definition visible light camera, an RTK module, and a lidar. The mission planning module is used to automatically generate optimized flight paths and sensor parameters based on the digital elevation model, geological map, mining plan map, and real-time meteorological data of the target mining area.
3. The open-pit mine real-scene modeling system according to claim 1, characterized in that, The data communication and transmission unit includes: Intelligent transmission module: Deployed at the edge preprocessing gateway, it supports compression, packetization and adaptive transmission of collected data; Local Area Network (LAN): Deploy a local area network to achieve efficient data backhaul; Local data center / server: Receives and stores data transmitted back from drones in real time.
4. The open-pit mine real-scene modeling system according to claim 1, characterized in that, The data preprocessing and quality assessment unit includes: Data preprocessing module: used to perform uniform illumination and color matching, distortion correction on the returned raw image data, and denoising and registration on the point cloud data; Quality assessment module: Used to automatically assess data quality based on indicators such as sharpness, overlap, and color consistency, through a trained convolutional neural network model; Quality control closed-loop module: When the evaluation data is unqualified, it automatically generates a reflight or reacquisition task instruction and sends it to the intelligent data acquisition unit.
5. The open-pit mine real-scene modeling system according to claim 1, characterized in that, The real-scene 3D modeling engine unit includes: Initial Spatial Point Cloud Generation Module: Used to recover the camera pose from oblique photogrammetric images using SfM and generate sparse point clouds; 3DGS scene initialization module: Constructs an initial distribution of a 3D Gaussian ellipsoid based on the initial point cloud; Iterative training module: continuously optimizes the Gaussian parameters through differentiable rendering and parameter optimization; 3DGS Model Output Module: Generates the final model that supports real-time rendering on both web and mobile devices.
6. A method for real-world modeling of open-pit mines, characterized in that, The system described in any one of claims 1-5 includes the following steps: S1, collecting oblique photographic image data and laser point cloud data of the target mining area to construct an initial spatial point cloud of the mining area; S2. Based on the mining area DEM, mining status and real-time meteorological data, generate an adaptive flight acquisition mission and transmit data back through the communication link; S3. Perform preprocessing and quality assessment on the returned data; if it fails to meet the requirements, trigger the re-sampling mechanism. S4. Initialize a 3D Gaussian scene using the initial spatial point cloud, and perform iterative training through differentiable rendering and parameter optimization. S5. Introduce geological semantic constraints during training to enhance the geological rationality and continuity of the model; S6. Outputs the trained 3DGS model, supporting high frame rate rendering and interaction on both web and mobile devices.
7. The method for real-scene modeling of an open-pit mine according to claim 6, characterized in that, Step S1 involves constructing the initial spatial point cloud of the mining area, including: Control a drone equipped with multiple sensor payloads to execute a preset route and a preset number of flights over the target mining area to obtain oblique photogrammetry image sequences and laser point cloud datasets; Feature extraction and matching are performed on the oblique photographic image sequence, and camera parameters are calculated using motion reconstruction structure to generate sparse point clouds; Trajectory and point cloud calculations are performed on the laser point cloud dataset to obtain a high-precision three-dimensional coordinate point cloud.
8. The method for real-scene modeling of an open-pit mine according to claim 6, characterized in that, Step S4 involves iterative training using differentiable rendering and parameter optimization, including: 3D Gaussian initialization: Based on the initial spatial point cloud, a 3D Gaussian ellipsoid is created for each 3D point, with parameters including position, covariance matrix, opacity and spherical harmonic coefficients; Differentiable rendering: 3D Gaussian projection is applied to a 2D image plane, and alpha blending is performed using a tile-based rasterizer. Parameter optimization: Calculate the loss function between the rendered image and the real image, and update the Gaussian parameters through backpropagation; Adaptive density control: Periodically perform cloning, splitting, and pruning operations to optimize the distribution of Gaussian points and model details.