Mine safety monitoring method and device based on unmanned aerial vehicle scanning mapping and 3DGS scene reconstruction, equipment and medium

Abstract

The application discloses a mine safety monitoring method and device based on unmanned aerial vehicle scanning mapping and 3DGS scene reconstruction, equipment and medium, the method comprises the following steps: acquiring multi-view image data, laser point cloud data and unmanned aerial vehicle attitude data, and preprocessing to obtain a standardized mine observation data set, generating a three-dimensional scene model of a mine monitoring area based on the standardized mine observation data set, and labeling a safety risk area in the three-dimensional scene model, extracting safety risk parameters in the three-dimensional scene model based on the safety risk area, and determining the risk level corresponding to each safety risk area based on the safety risk parameters, sending the early warning command and monitoring strategy corresponding to each risk level to the mine control system to monitor the safety of the mine monitoring area and dynamically warn, thereby solving the problem of low risk identification accuracy and identification lag caused by low traditional mine modeling accuracy, greatly improving the real-time and reliability of mine safety monitoring.

Description

Technical Field

[0001] This invention relates to the field of mine safety technology, and in particular to a mine safety monitoring method, device, equipment and medium based on UAV scanning mapping and 3DGS scene reconstruction. Background Technology

[0002] Mine safety is the cornerstone of industrial development, impacting the safety of people and property as well as the sustainability of the ecological environment. However, the mine environment is complex and ever-changing, with high slopes, goafs, and ore piles prone to safety hazards such as cracks, deformation, landslides, and water accumulation. Traditional safety monitoring relies primarily on manual inspections and fixed sensors, which suffers from low efficiency, narrow coverage, delayed risk identification, and poor reliability in harsh environments such as dust and vibration. In recent years, while UAV remote sensing technology has been applied to mine mapping, the traditional 3D models it generates often suffer from insufficient accuracy, slow reconstruction speed, and difficulty in dynamically quantifying subtle risk features such as millimeter-level cracks.

[0003] Existing reconstruction methods based on photogrammetry or traditional point clouds face challenges in complex, large-scale, highly dynamic, and dusty environments like mines. These challenges include coarse reconstruction models, significant loss of detail, and low computational efficiency, hindering the timely and accurate identification of early safety risks. Furthermore, existing monitoring systems typically stop at risk identification, failing to effectively integrate with mine control systems and thus preventing the formation of an automated "perception-decision-response" closed loop. This results in delayed early warning responses and missed opportunities for optimal intervention. The field of mine safety urgently needs a technology capable of rapid, high-precision, and automated scene reconstruction and risk identification. This technology should deeply integrate advanced sensing methods with intelligent reconstruction algorithms and automated control systems to fundamentally improve the real-time performance, accuracy, and proactiveness of mine safety monitoring. Summary of the Invention

[0004] The main objective of this invention is to provide a method, device, equipment, and medium for mine safety monitoring based on UAV scanning mapping and 3DGS scene reconstruction. This invention aims to solve the technical problems of low accuracy in traditional mine modeling, which leads to low accuracy and lag in risk identification, and the failure to effectively link with the mine control system, resulting in delayed risk warning response.

[0005] To achieve the above objectives, this invention provides a mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction. The method is applied to a UAV scanning system deployed in a mine monitoring area, and includes the following steps:

[0006] Acquire multi-view image data, laser point cloud data, and drone attitude data collected by the drone;

[0007] Based on the UAV attitude data, the multi-view image data and the laser point cloud data are preprocessed to obtain a standardized mine observation dataset. The preprocessing includes data cleaning and spatiotemporal alignment of data.

[0008] The standardized mine observation dataset is input into the optimized 3DGS model to generate a three-dimensional scene model of the mine monitoring area, and safety risk areas are marked in the three-dimensional scene model.

[0009] Based on the security risk areas, security risk parameters are extracted from the 3D scene model, and the risk level corresponding to each security risk area is determined based on the security risk parameters.

[0010] The warning instructions and monitoring strategies corresponding to each risk level are sent to the mine control system to conduct safety monitoring and dynamic warnings for the mine monitoring area.

[0011] Optionally, before acquiring the multi-view image data, laser point cloud data, and UAV attitude data collected by the UAV, the following steps are also included:

[0012] A scanning strategy is generated based on the topographic feature data of the mine monitoring area. The topographic feature data includes contour line data and mining range data. The scanning strategy includes: a Z-shaped dense scanning strategy for high slope areas, a spiral scanning strategy for ore pile areas, and a bidirectional parallel scanning strategy for transportation roads.

[0013] A scan path is generated based on the aforementioned scan strategy;

[0014] Based on the scanning path, multiple drones are controlled to scan the mine monitoring area and monitor the drone attitude data of each drone. The drones are equipped with industrial cameras, lidar and attitude sensors. During the scanning process, the drones collect multi-view image data and laser point cloud data.

[0015] Optionally, the preprocessing of the multi-view image data and the laser point cloud data based on the UAV attitude data to obtain a standardized mine observation dataset includes:

[0016] The multi-view image data is subjected to dust denoising processing based on an adaptive bilateral filtering algorithm to obtain candidate images;

[0017] A depth-guided atmospheric scattering model is constructed based on the depth information of laser point cloud data. The dark channel transmittance of the candidate image is weighted and corrected using point cloud depth constraints. The local atmospheric light value in the non-uniform dust environment of the mine is calculated. Based on the local atmospheric light value, the candidate image is mapped into a clear target image.

[0018] Based on the spatiotemporal alignment relationship between the target clear image and the laser point cloud data, dynamic interference features that cause displacement between consecutive frames are identified and eliminated through reprojection error analysis and semantic consistency verification, thereby obtaining static mine image data and static mine point cloud data.

[0019] Based on the UAV attitude data, the static mine image data and the static mine point cloud data are spatiotemporally aligned to obtain a standardized mine observation dataset.

[0020] Optionally, before inputting the standardized mine observation dataset into the optimized 3DGS model, the method further includes:

[0021] Based on the local normal vector distribution characteristics of laser point cloud data, an anisotropic geometric constraint term is constructed; the anisotropic geometric constraint term is used to initialize the covariance matrix of the Gaussian kernel in the original 3DGS model, so that the principal axis direction of the Gaussian kernel is orthogonal to the normal direction of the mine slope surface, and the first 3DGS model is obtained.

[0022] Redundant Gaussian kernels are removed based on the spatial overlap and feature similarity of adjacent Gaussian kernels in the first 3DGS model to obtain the second 3DGS model.

[0023] Based on the constraints of mine safety characteristics and the risk area characteristics of image data in historical standardized mine observation data, the second 3DGS model is optimized by Gaussian kernel weight to obtain the third 3DGS model.

[0024] The third 3DGS model is optimized and trained based on the labeled data of the security risk area to obtain the optimized 3DGS model.

[0025] Optionally, the third 3DGS model is optimized and trained based on the labeled data of the security risk area to obtain an optimized 3DGS model, including:

[0026] Based on the density distribution characteristics of point cloud data and the texture distribution characteristics of image data in the historical standardized mine observation data, the missing point cloud regions and missing image regions in the historical standardized mine observation data are determined.

[0027] One or more complete missing regions are determined based on the missing regions of the point cloud and the missing regions of the image.

[0028] Based on the region boundary of the complete missing region, multiple valid region ranges are selected from the domain of the complete missing region;

[0029] Based on the regional features of the effective region, the Gaussian kernel parameters of the third 3DGS model are optimized to construct a joint loss function that includes geometric continuity constraints and texture consistency constraints.

[0030] Based on the joint loss function, the Gaussian kernel generated in the missing region is iteratively optimized by gradient descent using the features of the effective region range, minimizing the geometric tearing error at the boundary of the missing region, and obtaining the fourth 3DGS model.

[0031] The fourth 3DGS model is optimized and trained based on the labeled data of the security risk area to obtain the optimized 3DGS model.

[0032] Optionally, the safety risk parameters include slope gradient, crack size, and water accumulation area; the extraction of safety risk parameters from the three-dimensional scene model based on the safety risk area includes:

[0033] Based on the safety risk area, a plane is fitted to the point cloud of the slope surface in the three-dimensional scene model. The angle between the plane normal vector and the vertical direction is calculated to obtain the slope.

[0034] Based on the security risk area, the 3D scene model is sampled along the crack annotation path, the length between sampling points and the width of the point cloud on both sides of the crack are calculated, and the crack size is obtained.

[0035] Based on the safety risk area, the two-dimensional projected area of ​​the water accumulation area marked in the three-dimensional scene model is extracted to obtain the water accumulation area.

[0036] Optionally, after sending the early warning instructions and monitoring strategies corresponding to each risk level to the mine control system, the method further includes:

[0037] In response to the processing results fed back by the mine control system, multiple drones are controlled to scan the safety risk areas in the mine monitoring area based on the processing results;

[0038] The 3D scene model is updated based on the data scanning results, and the safety risk parameters are re-extracted based on the updated 3D scene model.

[0039] Determine whether all safety risk parameters are within the safety threshold range;

[0040] If all safety risk parameters are within the safety threshold range, the risk is considered eliminated.

[0041] If at least one safety risk parameter is not within the safety threshold range, the risk level of the safety risk area is updated, and the early warning instructions and monitoring strategies are regenerated based on the updated risk level. Then, the process returns to the step of sending the early warning instructions and monitoring strategies corresponding to each risk level to the mine control system to perform safety monitoring and dynamic early warning for the mine monitoring area.

[0042] Furthermore, to achieve the above objectives, this invention also proposes a mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction, wherein the mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction includes:

[0043] The data acquisition module is used to acquire multi-view image data, laser point cloud data, and UAV attitude data collected by the UAV;

[0044] The data processing module is used to preprocess the multi-view image data and the laser point cloud data based on the UAV attitude data to obtain a standardized mine observation dataset. The preprocessing includes data cleaning and data spatiotemporal alignment.

[0045] The mine scene reconstruction module is used to input the standardized mine observation dataset into the optimized 3DGS model, generate a three-dimensional scene model of the mine monitoring area, and mark the safety risk areas in the three-dimensional scene model;

[0046] The risk identification module is used to extract safety risk parameters from the three-dimensional scene model based on the safety risk areas, and to determine the risk level corresponding to each safety risk area based on the safety risk parameters.

[0047] The risk response module is used to send early warning instructions and monitoring strategies corresponding to each risk level to the mine control system in order to conduct safety monitoring and dynamic early warning of the mine monitoring area.

[0048] Furthermore, to achieve the above objectives, this application also proposes a mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction. The device includes: a memory, a processor, and a mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction stored in the memory. The processor is used to run the mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction. The computer program is configured to implement the steps of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described above.

[0049] In addition, to achieve the above objectives, this application also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described above.

[0050] This invention acquires multi-view image data, laser point cloud data, and UAV attitude data collected by a UAV. Based on the UAV attitude data, it preprocesses the multi-view image data and laser point cloud data to obtain a standardized mine observation dataset. Preprocessing includes data cleaning and spatiotemporal alignment. The standardized mine observation dataset is then input into an optimized 3DGS model to generate a 3D scene model of the mine monitoring area. Safety risk areas are marked in the 3D scene model. Safety risk parameters are extracted from the 3D scene model based on these safety risk areas, and the risk level corresponding to each safety risk area is determined based on these parameters. Early warning instructions and monitoring strategies corresponding to each risk level are sent to the mine control system for safety monitoring and dynamic early warning of the mine monitoring area. This invention preprocesses multi-view image data and laser point cloud data based on UAV attitude data, achieving unification of image and point cloud data in time and space dimensions. Based on the optimized 3DGS model, a three-dimensional scene model of the mine monitoring area is created, thereby improving the accuracy and completeness of mine scene reconstruction, fully restoring the geometric shape and texture features of the mine's static structure, and fully exploring the geometric features and quantitative features of safety risks of the mine's static structure. Early warning instructions and monitoring strategies corresponding to each risk level are sent to the mine control system to conduct safety monitoring and dynamic early warning of the mine monitoring area, forming an automated closed loop of perception-decision-response, greatly improving the speed of risk identification and early warning response, and fundamentally improving the real-time performance, accuracy, and proactivity of mine safety monitoring. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of the structure of a mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction, which is part of the hardware operating environment of the embodiment of the present invention.

[0053] Figure 2 This is a flowchart illustrating the first embodiment of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction of the present invention.

[0054] Figure 3 This is a flowchart illustrating the second embodiment of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction of the present invention.

[0055] Figure 4 This is a flowchart illustrating the third embodiment of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction of the present invention.

[0056] Figure 5 This is a flowchart illustrating the fourth embodiment of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction of the present invention.

[0057] Figure 6 This is a structural block diagram of the first embodiment of the mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction of the present invention.

[0058] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0059] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0060] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a mine safety monitoring equipment based on UAV scanning mapping and 3DGS scene reconstruction, which is part of the hardware operating environment of the embodiment of the present invention.

[0061] like Figure 1 As shown, the mine safety monitoring equipment based on UAV scanning mapping and 3DGS scene reconstruction may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk storage device. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0062] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on mine safety monitoring equipment based on UAV scanning mapping and 3DGS scene reconstruction. It may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0063] like Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction.

[0064] exist Figure 1 In the mine safety monitoring equipment based on UAV scanning mapping and 3DGS scene reconstruction shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the mine safety monitoring equipment based on UAV scanning mapping and 3DGS scene reconstruction of the present invention can be set in the mine safety monitoring equipment based on UAV scanning mapping and 3DGS scene reconstruction. The mine safety monitoring equipment based on UAV scanning mapping and 3DGS scene reconstruction calls the mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction stored in the memory 1005 through the processor 1001, and executes the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction provided in the embodiment of the present invention.

[0065] This invention provides a mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction of the present invention.

[0066] In this embodiment, the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction includes the following steps:

[0067] Step S10: Acquire multi-view image data, laser point cloud data and UAV attitude data collected by the UAV.

[0068] It should be noted that this embodiment applies to a drone scanning system deployed in the mine monitoring area. This system is suitable for real-time monitoring and early warning of safety risks such as high slope instability, ore pile collapse, and water accumulation on transport roads in open-pit mines. The drone scanning system can be an integrated system consisting of one or more drones equipped with dust-resistant industrial cameras, LiDAR, and attitude sensors, along with a ground control terminal, data transmission module, and power supply module. The mine monitoring area can be a key area in the mine requiring focused safety monitoring, typically including critical areas prone to safety risks (such as slope landslides, ore pile collapses, and crack propagation), such as high slopes, ore piles, tunnel entrances, tailings ponds, work platforms, and transport roads.

[0069] It should be understood that the executing entity of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or a terminal electronic device capable of performing the above functions. The following description uses a mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction (hereinafter referred to as the monitoring device) as an example to illustrate this embodiment and the following embodiments.

[0070] It should be noted that multi-view image data can be a series of images taken by a drone flying along a preset route from different angles (such as orthogonal, 45° tilt, and side view) of the mine monitoring area using a high-definition camera. This data includes two-dimensional visual information such as texture, color, and contour of the mine scene, serving as a texture data source for 3D reconstruction. Laser point cloud data can be generated by a lidar device onboard a drone emitting laser pulses towards the mine monitoring area. By receiving the reflected pulses, the three-dimensional coordinates (X, Y, Z) of the target points are calculated. The point cloud collection formed by the convergence of numerous target points can accurately represent the three-dimensional geometric shape of the mine scene, serving as a geometric data source for 3D reconstruction.

[0071] It should be noted that the attitude data of the UAV can be the state data of the UAV recorded by the built-in GPS module, IMU (inertial measurement unit) and other devices during the flight. It mainly includes GPS coordinates (longitude, latitude, altitude), heading angle, pitch angle, roll angle and corresponding collection timestamp, which is used to realize spatiotemporal alignment of multi-source data.

[0072] In some embodiments, the monitoring equipment can define the boundary range of the monitoring area based on the needs of mine monitoring, plan the flight route of the UAV in combination with the mine terrain (such as slope and altitude difference), and control the UAV to perform data scanning and collection based on the corresponding scanning method for different types of areas, ensuring that there are no blind spots in the monitoring area, and setting the flight route overlap rate to 30%-50% (to ensure the accuracy of image texture stitching).

[0073] In some embodiments, the monitoring device can perform focal length calibration and exposure parameter setting (adjusting ISO and shutter speed according to lighting conditions) on the image acquisition device of the UAV, perform ranging accuracy calibration on the lidar device, and perform positioning and attitude accuracy calibration on the GPS and IMU devices to ensure that the acquisition accuracy of each device meets the requirements.

[0074] Furthermore, in order to improve data collection efficiency and ensure data quality, the following may be included before step S10:

[0075] Step S101: Generate a scanning strategy based on the terrain feature data of the mine monitoring area. The terrain feature data includes contour line data and mining range data. The scanning strategy includes: a Z-shaped dense scanning strategy for high slope areas, a spiral scanning strategy for ore pile areas, and a bidirectional parallel scanning strategy for transportation roads.

[0076] Step S102: Generate a scan path based on the scan strategy;

[0077] Step S103: Based on the scanning path, control multiple drones to scan the mine monitoring area and monitor the drone attitude data of each drone. The drones are equipped with industrial cameras, lidar and attitude sensors. During the scanning process, the drones collect multi-view image data and laser point cloud data.

[0078] It should be noted that topographic feature data is used to characterize the topography and mining status of the mine monitoring area, including contour line data and mining range data. It can also be supplemented with data such as slope, elevation difference, and landform type to formulate differentiated scanning strategies.

[0079] It should be noted that the Z-shaped dense scanning strategy can be a scanning mode designed for areas with complex terrain, steep slopes, and high safety risks (such as high slopes). The UAV flies along a "Z"-shaped trajectory, and achieves dense data collection by reducing the spacing between flight paths and increasing the scanning overlap rate, ensuring that no details of complex terrain are missed.

[0080] The spiral scanning strategy can be a scanning mode designed for areas with a three-dimensional cone shape or irregular convex shape (such as a ore pile). The UAV takes the center point of the area as the center and gradually increases or decreases the flight altitude in a spiral trajectory to achieve a full-range, blind-spot-free circumferential scan of the ore pile surface.

[0081] The bidirectional parallel scanning strategy can be a scanning mode designed for areas with flat terrain and long strip distribution (such as transportation roads). The UAV first flies parallel to the axis of the area in one direction to scan, and then flies parallel to the opposite direction to supplement the scan, ensuring that the data of the road surface and the edges on both sides are complete.

[0082] It should be noted that high slope areas refer to steep slopes (typically >45°) in mines, prone to landslides, crack propagation, and other safety risks. These are key high-risk areas for mine monitoring, requiring extremely high scanning accuracy and density. Ore pile areas can be three-dimensional areas formed by the accumulation of ore, slag, etc., in a mine. They are often irregularly conical in shape with significant surface undulations, requiring comprehensive scanning to accurately obtain key parameters such as volume and slope angle. Transport roads are passageways used for transporting ore, equipment, and personnel in a mine. They are often long and narrow with gentle terrain. The core of the scanning is to obtain information on the road surface flatness, width, and side boundaries.

[0083] In practice, the monitoring equipment adopts a "zonal scanning + key re-scanning" path strategy based on the contour lines and mining range data in the mine terrain: "Z-shaped dense scanning" is used in high slope areas, with a horizontal overlap of 80% and a vertical overlap of 70%; "spiral scanning" is used in ore pile areas, with the radius increasing by 0.5m each time; and "bidirectional parallel scanning" is used for transportation roads.

[0084] In some embodiments, the industrial camera carried by the drone has an IP67 dustproof and waterproof rating, a lidar ranging accuracy of ≤±2cm, and an attitude sensor (IMU) sampling frequency of ≥200Hz; it simultaneously acquires multi-view images with a resolution of ≥5000×3000 pixels, a laser point cloud point density of ≥100 points / m², and a drone real-time attitude position error of ≤±5cm, and realizes real-time data transmission through a 5G repeater.

[0085] Step S20: Based on the UAV attitude data, preprocess the multi-view image data and the laser point cloud data to obtain a standardized mine observation dataset. The preprocessing includes data cleaning and spatiotemporal alignment.

[0086] It should be noted that the standardized mine observation dataset can be a standardized dataset that meets the requirements of subsequent 3D reconstruction in terms of data format, spatiotemporal reference, data quality, and data accuracy after data cleaning and spatiotemporal alignment. It includes static mine image data and static mine point cloud data.

[0087] In some embodiments, the monitoring equipment can perform data cleaning on the image data and point cloud data respectively, and then perform timestamp matching and spatial coordinate transformation on the cleaned image data and point cloud data based on UAV attitude data to achieve data spatiotemporal alignment. This integrates the cleaned and spatiotemporally aligned static mine image data and static mine point cloud data, adds data description information (acquisition time, equipment parameters, preprocessing process, spatiotemporal reference), and encapsulates it into a standardized mine observation dataset according to a preset format.

[0088] Step S30: Input the standardized mine observation dataset into the optimized 3DGS model to generate a three-dimensional scene model of the mine monitoring area, and mark the safety risk areas in the three-dimensional scene model.

[0089] It should be noted that the 3D scene model can be a complete 3D digital model of the mine monitoring area reconstructed by 3DGS model based on a standardized mine observation dataset. It can accurately restore the geometric shape and texture features of the static structure of the mine (slope, ore pile, roadway entrance, etc.) for subsequent safety risk analysis.

[0090] Safety risk areas can be areas within the mine monitoring area that have potential safety hazards. These mainly include cracked areas (slopes, ore pile cracks), water accumulation areas (work platforms, roadway water accumulation), high slope deformation areas, and ore pile collapse risk areas, which are key targets for mine safety monitoring.

[0091] It should be understood that this embodiment can mark the location, range and type of safety risk areas in the 3D scene model by combining automatic identification by semantic segmentation algorithm and manual verification, and store the annotation information in the model attribute library, thereby realizing the annotation of safety risk areas in the 3D scene model.

[0092] It is understood that this embodiment can adapt and optimize the traditional 3DGS (3D Gaussian Splatting) algorithm to suit the characteristics of mining scenarios, resulting in an optimized 3DGS model. The optimization directions include: laser point cloud-guided Gaussian kernel initialization (reducing redundancy in homogeneous areas), adding mine safety feature constraints (enhancing the reconstruction accuracy of risk areas), neighborhood Gaussian kernel interpolation (supplementing local data loss caused by dust), and multi-machine collaborative scanning data fusion (eliminating blind spots caused by viewpoint obstruction), which has the ability to reconstruct efficiently and with high precision in mining scenarios.

[0093] In practice, the monitoring equipment uses a 3DGS scene reconstruction algorithm for mine environment optimization. It inputs a standardized mine observation dataset into the optimized 3DGS model to generate a high-precision three-dimensional scene model of the mine area, and simultaneously marks safety risk feature areas such as cracks, slope deformation areas, and water accumulation areas.

[0094] In some embodiments, the monitoring equipment can input a standardized mine observation dataset (static images and static point clouds) into an optimized 3DGS model. The model is iteratively optimized based on Gaussian kernel parameters (position, color, scale, etc.) to fit the three-dimensional surface morphology and texture features of the mine scene. The accuracy is monitored in real time during the reconstruction process to ensure that the generated three-dimensional scene model meets the preset indicators. The reconstructed model is post-processed to remove abnormal structures (such as floating points and structural fracture areas) in the model, optimize the model texture stitching effect, and generate a complete and high-precision three-dimensional scene model of the mine monitoring area, thereby completely restoring the geometric morphology and texture features of the static structure of the mine.

[0095] In some embodiments, the monitoring equipment can employ the U-Net semantic segmentation network, inputting static image data from a standardized dataset to automatically segment two-dimensional ranges of safety risk areas such as cracks and water accumulation areas. These ranges are then mapped to a three-dimensional scene model through spatiotemporal correlation to initially determine the three-dimensional coordinates of the risk areas. Based on expert-annotated data combined with multi-angle renderings of the three-dimensional scene model, the automatically segmented risk areas are reconfirmed, with a focus on identifying complex risk areas such as high slope deformation zones (areas prone to missed or false detections due to automatic segmentation), correcting any missed or false detections. After standardizing and organizing the annotation information of the risk areas, it is stored in the attribute library of the three-dimensional scene model, establishing a correlation mapping between the geometric information of the three-dimensional model and the risk attribute information, thus balancing the efficiency and accuracy of risk area identification.

[0096] Step S40: Extract safety risk parameters from the 3D scene model based on the safety risk areas, and determine the risk level corresponding to each safety risk area based on the safety risk parameters.

[0097] It should be noted that safety risk parameters can be key indicators that can quantify the degree of risk extracted from the safety risk areas of the 3D scene model. The core parameters are different for different types of risk areas, such as crack areas (crack width, length, depth, propagation rate), water accumulation areas (water accumulation area, depth, duration), high slope deformation areas (deformation amount, deformation rate, slope gradient), and ore pile areas (pile height, slope angle, stability coefficient), etc.

[0098] Risk levels can be determined by classifying the degree of danger in each safety risk area based on extracted safety risk parameters and in conjunction with industry standards for mine safety management or preset thresholds. These levels can be divided into four categories: low risk, medium risk, high risk, and extremely high risk, and are used to formulate early warning instructions and monitoring strategies.

[0099] In practice, the monitoring equipment extracts safety risk parameters from the three-dimensional scene model, including slope gradient, crack length / width, and water accumulation area. The safety risk parameters are compared with preset safety thresholds to determine the risk level. The risk level is set into three different risk level states: red warning, yellow warning, and safe.

[0100] In some embodiments, the monitoring equipment can extract corresponding core parameters for different types of risk areas based on the geometric information and attribute library annotation information of the 3D scene model and using a 3D feature extraction algorithm: for crack areas, the width and length are extracted by fitting the model surface contour, and the depth is estimated by point cloud density analysis; for water accumulation areas, the water accumulation area and depth are extracted by analyzing the model elevation difference; for high slope deformation areas, the deformation amount and deformation rate are extracted by comparing the model with the historical benchmark model; for ore pile areas, parameters such as pile height and slope angle are extracted by calculating the model volume and fitting the contour; the extracted parameters are standardized, and the units are unified (such as width: mm, area: m², deformation amount: cm), and outliers are removed (through statistical analysis) to ensure the accuracy of the parameters.

[0101] Furthermore, in order to accurately extract safety risk parameters and achieve precise quantification of multi-dimensional risks, step S40 above may include:

[0102] Step S401: Based on the safety risk area, fit a plane to the point cloud of the slope surface in the three-dimensional scene model, calculate the angle between the plane normal vector and the vertical direction, and obtain the slope.

[0103] Step S402: Based on the security risk area, sample the three-dimensional scene model along the crack annotation path, calculate the length between sampling points and the width of the point cloud on both sides of the crack, and obtain the crack size;

[0104] Step S403: Based on the safety risk area, extract the two-dimensional projected area of ​​the water accumulation area from the three-dimensional scene model to obtain the water accumulation area.

[0105] In the specific implementation, risk parameters are extracted from the 3D scene model:

[0106] The slope gradient is calculated by fitting a plane to the slope surface point cloud in the model, and the angle between the plane's normal vector and the vertical direction is calculated. The slope point cloud subset is defined as follows: ;

[0107] in, This represents the k-th sampling point in the slope point cloud. The three-dimensional spatial coordinate components;

[0108] The fitted plane equation is: The coefficients are solved using the least squares method, and the least squares objective function is:

[0109]

[0110] The constraints are as follows (to avoid coefficient redundancy):

[0111] The unit normal vector of the plane is: ;

[0112] The vertical unit vector is: ;

[0113] The angle between the plane unit normal vector and the vertical unit vector satisfy slope gradient The calculation formula is: ;

[0114] Crack length / width is calculated by sampling along the crack's marked path; the sampling point set along the crack's marked path is set as follows: (The distance between adjacent sampling points is 0.1m), the coordinates of the kth sampling point are: The direction vector of the crack is The direction vector perpendicular to the crack direction and parallel to the ground is Represents the cross product; take the points on both sides of the crack edge. The crack width at the kth sampling point is... for: Overall average width of cracks The calculation formula is:

[0115]

[0116] The water accumulation area is extracted by showing the two-dimensional projected area of ​​the marked water accumulation zone. Preset safety thresholds may include: a red warning for slope gradient > 45°, and a yellow warning for slope gradient ≤ 45°; a red warning for crack length > 3m or width > 5mm, and a yellow warning for crack length ≤ 3m or crack width ≤ 5mm (1m ≤ crack length ≤ 3m or 2mm ≤ width ≤ 5mm); a yellow warning for water accumulation area > 10m²; all parameters within the above threshold ranges are considered safe.

[0117] Step S50: Send the early warning instructions and monitoring strategies corresponding to each risk level to the mine control system to conduct safety monitoring and dynamic early warning of the mine monitoring area.

[0118] It should be noted that the early warning instructions can be instructions formulated based on the risk level of each safety risk area to remind mine managers to take countermeasures. Different risk levels correspond to different intensities of early warnings (such as low risk corresponding to a warning, medium risk corresponding to a warning of concern, high risk corresponding to an emergency warning, and extremely high risk corresponding to a production stoppage warning). The instructions include the warning area, risk type, risk level, and suggested countermeasures.

[0119] Monitoring strategies can be differentiated data collection and analysis plans for areas with different risk levels. The core is to increase the frequency and accuracy of monitoring in high-risk areas and maintain routine monitoring in low-risk areas in order to achieve optimal resource allocation and efficient monitoring. This includes monitoring cycles (e.g., once a day in high-risk areas and once a month in low-risk areas), monitoring accuracy requirements, key monitoring parameters, and other related content.

[0120] The mine control system can be the core command system for mine safety management. It integrates functions such as data receiving module, early warning release module, monitoring and dispatching module, and emergency response module. It can receive early warning instructions and monitoring strategies, release early warning information to relevant management personnel and operators, and dispatch monitoring equipment to carry out subsequent monitoring work.

[0121] In practice, the monitoring equipment can send early warning instructions corresponding to the risk level to the mine control system via wireless communication, and link on-site audible and visual alarms, equipment start-up and shutdown or personnel evacuation devices to achieve closed-loop monitoring of mine safety.

[0122] In some embodiments, the monitoring equipment can generate differentiated early warning instructions based on the risk level of each safety risk area: a "prompt warning" is generated for low-risk areas (recorded only in the system to remind managers to pay attention regularly); a "pay attention warning" is generated for medium-risk areas (sent SMS / APP reminders to on-site managers, requiring weekly checks); an "emergency warning" is generated for high-risk areas (sent real-time warnings to managers and the dispatch center, requiring on-site verification and the development of control measures within 24 hours); and a "shutdown warning" is generated for extremely high-risk areas (immediately sent the highest-level warning to the mine dispatch center and safety management department, requiring immediate cessation of operations in and around the area and activation of emergency response procedures).

[0123] In some embodiments, the monitoring equipment can formulate a matching monitoring strategy: for high-risk / extremely high-risk areas, the monitoring cycle is set to 1-3 days, using a high-precision scanning mode (improved image resolution and increased point cloud density), focusing on monitoring changes in risk parameters (such as crack propagation and increased deformation); for medium-risk areas, the monitoring cycle is 1 week, maintaining conventional scanning accuracy; for low-risk areas, the monitoring cycle is 1-3 months, using a conventional monitoring mode, and optimizing the allocation of monitoring resources.

[0124] Furthermore, in order to achieve a closed-loop dynamic early warning system for mine safety monitoring and improve mine safety, the above step S50 may include:

[0125] Step S501: In response to the processing results fed back by the mine control system, control multiple drones to perform data scanning on the safety risk areas in the mine monitoring area based on the processing results;

[0126] Step S502: Update the 3D scene model based on the data scanning results, and re-extract the safety risk parameters based on the updated 3D scene model;

[0127] Step S503: Determine whether all safety risk parameters are within the safety threshold range;

[0128] Step S504: If all safety risk parameters are within the safety threshold range, the risk is determined to be eliminated;

[0129] Step S505: If at least one safety risk parameter is not within the safety threshold range, the risk level of the safety risk area is updated, and the early warning instruction and monitoring strategy are regenerated based on the updated risk level. Then, the process returns to the step of sending the early warning instruction and monitoring strategy corresponding to each risk level to the mine control system to perform safety monitoring and dynamic early warning of the mine monitoring area.

[0130] In practical implementation, after receiving the handling results from the mine control system, the monitoring equipment controls the drone to re-scan the risk area; based on the re-scan data, the 3DGS three-dimensional scene model is updated, the risk parameters are recalculated, and the handling effect is verified; if the risk parameters still exceed the threshold, the early warning command is upgraded until the risk is eliminated, forming a closed-loop control. Let the risk parameters before handling (slope / crack width / water accumulation area) be... The corresponding parameters after treatment are: The formula for calculating the treatment effectiveness index E is:

[0131]

[0132] Judgment rule: If and , If a preset safety threshold is set, the action is deemed "effective"; if or If so, it is determined that "the handling did not meet the standards"; if If the condition is not met or the risk is aggravated, the warning instruction is upgraded and the drone is controlled to rescan until the risk is eliminated, thus forming a closed-loop control.

[0133] This embodiment acquires multi-view image data, laser point cloud data, and UAV attitude data collected by a UAV. Based on the UAV attitude data, it preprocesses the multi-view image data and laser point cloud data to obtain a standardized mine observation dataset. Preprocessing includes data cleaning and spatiotemporal alignment. The standardized mine observation dataset is then input into an optimized 3DGS model to generate a 3D scene model of the mine monitoring area. Safety risk areas are marked in the 3D scene model. Safety risk parameters are extracted from the 3D scene model based on these safety risk areas, and the risk level corresponding to each safety risk area is determined based on these parameters. Early warning instructions and monitoring strategies corresponding to each risk level are sent to the mine control system for safety monitoring and dynamic early warning of the mine monitoring area. In this embodiment, multi-view image data and laser point cloud data are preprocessed based on UAV attitude data, achieving unification of image and point cloud data in time and space dimensions. Based on the optimized 3DGS model, a three-dimensional scene model of the mine monitoring area is performed, thereby improving the accuracy and completeness of the mine scene reconstruction, fully restoring the geometric shape and texture features of the mine's static structure, and fully exploring the geometric features and quantitative features of safety risks of the mine's static structure. Early warning instructions and monitoring strategies corresponding to each risk level are sent to the mine control system to conduct safety monitoring and dynamic early warning of the mine monitoring area, forming an automated closed loop of perception-decision-response, greatly improving the speed of risk identification and early warning response, and fundamentally improving the real-time performance, accuracy and initiative of mine safety monitoring.

[0134] refer to Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction of the present invention.

[0135] Based on the first embodiment described above, in this embodiment, step S20 further includes:

[0136] Step S201: Perform dust denoising processing on the multi-view image data based on the adaptive bilateral filtering algorithm to obtain candidate images.

[0137] It should be noted that the adaptive bilateral filtering algorithm is an image filtering algorithm that takes into account both noise removal and edge preservation. By adaptively adjusting the filter kernel parameters, it adapts to the distribution characteristics of dust noise in the mine, achieving dust noise reduction without blurring the edges of the mine structure.

[0138] The candidate images are intermediate images obtained after the initial dust denoising of multi-view images through adaptive bilateral filtering. They may still have some dust scattering interference and need to be further sharpened.

[0139] It is understood that the dust noise reduction processing in this embodiment can be a process that removes dust noise from images and restores the real texture of the mining scene by using techniques such as filtering and scattering correction to address the problems of image blurring and reduced contrast caused by the high dust environment of the mine.

[0140] In practical implementation, the monitoring equipment can adaptively adjust the filter kernel size (e.g., 3×3 to 7×7), spatial standard deviation (e.g., 1.5 to 3.0), and grayscale standard deviation (e.g., 0.1 to 0.3) based on the dust concentration differences in the mine images (e.g., determined by the image grayscale standard deviation). The filter kernel is enlarged in areas with high dust concentration to ensure effective noise removal. The multi-view images are divided into scanning areas (e.g., high slopes, mine piles, roads), and the filter parameters are optimized according to the dust distribution characteristics of different areas (e.g., higher dust concentration in mine pile areas) to avoid incomplete noise reduction or blurred edges caused by uniform parameters. Adaptive bilateral filtering is performed on each multi-view image to remove high-frequency dust noise from the image, while retaining key structural features such as slope edges and mine pile corners. The candidate image set is output and associated with the timestamp of the corresponding UAV attitude data.

[0141] Step S202: Construct a depth-guided atmospheric scattering model based on the depth information of the laser point cloud data, use the point cloud depth constraint to perform weighted correction on the dark channel transmittance of the candidate image, calculate the local atmospheric light value in the non-uniform dust environment of the mine, and map the candidate image into a clear target image based on the local atmospheric light value.

[0142] It should be noted that the depth-guided atmospheric scattering model can be a scattering correction model that is built based on the physical principles of atmospheric scattering and introduces laser point cloud depth information as a constraint. The core is to quantify the atmospheric scattering intensity at different spatial locations through depth information (the farther the distance, the stronger the scattering), so as to achieve accurate adaptation to non-uniform scattering environments.

[0143] Dark channel transmittance reflects the degree to which each pixel in an image is affected by atmospheric scattering. The closer the value is to 1, the less affected it is by scattering, and the closer the value is to 0, the more affected it is by scattering. It is the core parameter for image dehazing and descattering processing.

[0144] Point cloud depth constraints can utilize the precise three-dimensional depth information of laser point cloud data to constrain and correct the calculation results of the dark channel transmittance of the image, avoiding transmittance estimation errors caused by single image information.

[0145] Local atmospheric light values ​​can be calculated because the atmospheric scattering intensity varies in different areas under non-uniform dust environments. Therefore, the image is divided into multiple local areas, and the atmospheric light intensity (i.e., the true light source intensity in the area that is not affected by scattering) is calculated separately for each area, rather than using a globally uniform atmospheric light value.

[0146] In the specific implementation, the original laser point cloud data is preprocessed (denoising, simplification, coordinate system unification) to extract depth information of the spatial region corresponding to the candidate image, generating a depth map (mapping the three-dimensional depth information to the two-dimensional image coordinate system). Based on the classical atmospheric scattering model, the laser point cloud depth information is introduced to construct a mapping relationship between depth and scattering intensity (the greater the depth, the greater the scattering intensity coefficient), forming a depth-guided atmospheric scattering model. First, the initial transmittance of the candidate image is calculated using the traditional dark channel method. Then, the initial transmittance is weighted and corrected using the laser point cloud depth map—increasing the transmittance correction weight for regions with greater depth (stronger scattering) to enhance scattering suppression, and decreasing the correction weight for regions with less depth (weaker scattering). To avoid over-correction leading to detail distortion, an adaptive block-segmentation strategy is adopted, dividing the candidate image into multiple local regions (the block size is adaptively adjusted according to the image resolution and the complexity of the mining scene). For each local region, the top X% (e.g., the top 0.1%) of bright pixels in grayscale value are selected, and the average grayscale value of these bright pixels is calculated as the local atmospheric light value of the region. The corrected dark channel transmittance and local atmospheric light value are substituted into the depth-guided atmospheric scattering model to solve the true grayscale value of each pixel in reverse. The resulting image is then stitched together to generate a clear target image, thus accurately adapting to the non-uniform dust environment of the mine. This solves the problem that traditional global atmospheric scattering models cannot cope with local scattering differences, significantly improving the image descattering and dehazing effects.

[0147] In some embodiments, the monitoring device can take the local minimum value of an 8×8 pixel window for each candidate image to generate a corresponding dark channel image, highlighting dust scattering areas (where the pixel value of the dark channel is relatively high). Based on statistical analysis of the dark channel image, a dust scattering threshold is set (calibrated according to historical dust data from the mine), and areas with severe scattering are screened out. The atmospheric scattering coefficient of this area is calculated using the least squares method to ensure that the coefficient matches the scattering characteristics of mine dust. Based on the atmospheric scattering physical model, the calculated atmospheric scattering coefficient is used to correct the scattering of the candidate image, and the target clear image without dust interference is derived in reverse. The sharpening effect is verified by calculating the contrast and edge gradient between the target clear image and the candidate image. If the contrast improvement is insufficient (not reaching the preset threshold), the dark channel window size and scattering coefficient calculation parameters are readjusted, and the processing procedure is repeated. The dark channel prior mechanism can be an image sharpening algorithm based on the atmospheric scattering physical model. The core assumption is that there is a dark channel with a pixel value close to 0 in a local area of ​​the fog-free (low dust) image, which is used to estimate the atmospheric scattering coefficient and restore the clear image. Dark channel images can be grayscale images generated by taking the minimum value of candidate images according to a preset window. They can highlight the grayscale characteristics of dust scattering areas and provide a core basis for calculating atmospheric scattering coefficients.

[0148] Step S203: Based on the spatiotemporal alignment relationship between the target clear image and the laser point cloud data, through reprojection error analysis and semantic consistency verification, identify and eliminate dynamic interference features that cause displacement between consecutive frames to obtain static mine image data and static mine point cloud data.

[0149] It should be noted that dynamic interference features can be objects in motion in the mine scanning scene, including mine cars, workers, and construction machinery. Their presence will interfere with the accuracy of the three-dimensional reconstruction of the static structure of the mine and must be removed from the collected data.

[0150] It should be noted that static mine image data can be a collection of clear images of the target after removing dynamic interference features, containing only the texture information of the static structure of the mine (slopes, ore piles, roads, etc.), which is used for texture mapping in subsequent 3D reconstruction. Static mine point cloud data can be a collection of laser point cloud data after removing dynamic interference features, containing only the 3D geometric information of the static structure of the mine, providing a precise geometric benchmark for 3D reconstruction.

[0151] In practical implementation, the monitoring equipment can further optimize the spatiotemporal alignment accuracy between the clear target image and the laser point cloud data based on UAV attitude data and LiDAR extrinsic parameters, ensuring that the coordinate systems of the two are identical and the timestamps are synchronized, establishing a precise mapping relationship between pixels and point clouds; feature extraction is performed on the clear target image (extracting visual features) and the laser point cloud data (extracting 3D key points, normal vectors, and other geometric features) respectively, and corresponding feature pairs are obtained through feature descriptor matching; the feature points on the laser point cloud side are projected onto the image coordinate system, and the reprojection error between each projection point and the corresponding feature point on the image side is calculated; an error threshold is set ( Based on the accuracy requirements of the mining scene, features with errors exceeding the threshold are marked as suspected dynamic interference features. A semantic segmentation model is used to perform semantic annotation on the clear target image and laser point cloud data respectively to obtain semantic segmentation results. Semantic verification is performed on the region where the suspected dynamic interference features are located. If the semantics on the image side is "moving vehicle" but the semantics on the point cloud side cannot match, or if the semantic categories of the same region conflict, then the feature is confirmed as a dynamic interference feature. The confirmed dynamic interference features are removed from the image and point cloud data. The remaining data after removal are stitched and integrated to generate static mining image data and static mining point cloud data.

[0152] Step S204: Based on the UAV attitude data, perform spatiotemporal alignment of the static mine image data and the static mine point cloud data to obtain a standardized mine observation dataset.

[0153] Understandably, this embodiment employs an adaptive bilateral filter ensemble dark channel prior algorithm to reduce dust noise in the images, removing image blurring caused by mine dust; it uses the YOLOv8 target detection algorithm to identify and remove dynamic interference features in the images and point clouds, such as mine trucks, personnel, and construction machinery, while preserving the static mine structure; and it uses timestamp matching and coordinate transformation based on UAV attitude data to achieve spatiotemporal alignment between the images and point clouds, forming a standardized mine observation dataset.

[0154] This embodiment effectively improves image quality through dust noise reduction and sharpening processing, achieving precise removal of dust noise while preserving the edges of the mine structure. The dark channel prior mechanism completely eliminates dust scattering interference, restoring the true texture of the scene and solving the image blurring problem in high-dust mine environments, providing high-quality texture data for subsequent target detection and 3D reconstruction. By accurately identifying and eliminating dynamic interference in the mine scene, it ensures that static data only contains the core static structure of the mine, avoiding distortion of the reconstruction model caused by dynamic interference features. Based on timestamp matching and spatial coordinate transformation of UAV attitude data, precise correlation between imagery and point cloud data is achieved, providing a solid benchmark for multi-source data fusion and 3D reconstruction, avoiding a decrease in reconstruction accuracy due to spatiotemporal misalignment, and reducing the preprocessing workload for subsequent 3D reconstruction.

[0155] refer to Figure 4 , Figure 4 This is a flowchart illustrating the third embodiment of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction of the present invention.

[0156] Based on the above embodiments, in this embodiment, before step S30, the method further includes:

[0157] Step S31: Based on the local normal vector distribution characteristics of the laser point cloud data, construct anisotropic geometric constraint terms; use the anisotropic geometric constraint terms to initialize the covariance matrix of the Gaussian kernel in the original 3DGS model, so that the principal axis direction of the Gaussian kernel is orthogonal to the normal direction of the mine slope surface, and obtain the first 3DGS model.

[0158] It should be noted that historical standardized mine observation data can be a standardized dataset obtained from past mine monitoring after preprocessing, including historical static images, point cloud data, and associated pose information, used for pre-optimization training of the 3DGS model. The original 3DGS model can be a basic 3D Gaussian sputtering model that has not been adapted and optimized for the mine scene, possessing general 3D reconstruction capabilities, but without adjusting parameters for the characteristics of mine data and safety monitoring requirements.

[0159] In practical implementation, the monitoring equipment can collect multiple sets of historical standardized mine observation data from different areas of the mine, screen out qualified point cloud data, and construct a point cloud density analysis dataset after unifying the coordinate benchmark. Using the spatial voxel partitioning method, the historical point cloud data is divided into grids according to a preset voxel size, the number of point clouds in each grid is counted, and the point cloud density distribution characteristics of each area are extracted (e.g., high density in high slope detail areas and low density in flat road areas). Based on the extracted density characteristics, an adaptive density matching algorithm is used to adjust the initial distribution of Gaussian kernels in the original 3DGS model—deploying more Gaussian kernels in high-density areas of the point cloud and reducing the number of Gaussian kernels in low-density areas, so that the Gaussian kernel distribution accurately matches the density characteristics of the mine point cloud. After the distribution optimization is completed, the Gaussian kernel distribution parameters are fixed, and the first 3DGS model adapted to the density characteristics of the mine point cloud is output.

[0160] Step S32: Based on the spatial overlap and feature similarity of adjacent Gaussian kernels in the first 3DGS model, redundant Gaussian kernels are removed to obtain the second 3DGS model.

[0161] It should be noted that the first 3DGS model can be an intermediate model obtained after Gaussian kernel distribution optimization. Its Gaussian kernel distribution is adapted to the density characteristics of the mine point cloud data, improving data adaptability. Redundant Gaussian kernels can be those in the first 3DGS model that have highly overlapping spatial locations and extremely high feature similarity. Their presence increases the model's computational load, reduces reconstruction efficiency, and has no additional reconstruction value. The second 3DGS model can be the model after removing redundant Gaussian kernels, which reduces model complexity and improves computational efficiency while retaining core reconstruction capabilities.

[0162] In its implementation, the monitoring equipment can calculate the spatial overlap (based on the proportion of intersection of Gaussian kernel spatial ranges) and feature similarity (based on the cosine similarity of feature vectors such as color and scale) of any two adjacent Gaussian kernels in the first 3DGS model. Combining the requirements for mine reconstruction accuracy and computational efficiency, it sets overlap and similarity thresholds (e.g., overlap > 60% and similarity > 85% are considered redundant). It then iterates through all Gaussian kernels, removes those that meet the redundancy criteria, and prioritizes retaining Gaussian kernels corresponding to key mine structures (such as slope edges and ore pile corners) to avoid losing core features. After redundancy removal, it fine-tunes the spatial distribution of the remaining Gaussian kernels to ensure no reconstruction blind spots and outputs a lightweight second 3DGS model.

[0163] Step S33: Based on the mine safety feature constraints and the risk area features of the image data in the historical standardized mine observation data, the second 3DGS model is optimized with Gaussian kernel weights to obtain the third 3DGS model.

[0164] It should be noted that mine safety feature constraints can be feature limitations set to meet the needs of mine safety monitoring, covering the reconstruction accuracy requirements of key safety features such as slope edges, cracks, and water accumulation areas. Risk area features can be texture, grayscale, and contour information of safety risk areas (such as cracks and high slope deformation areas) extracted from historical image data, used to strengthen the model's reconstruction weight for such areas. The third 3DGS model refers to the model obtained after Gaussian kernel weight optimization, which assigns higher weights to the Gaussian kernel of mine safety risk areas, enhancing the reconstruction accuracy of key areas.

[0165] In practical implementation, monitoring equipment can construct a set of safety features based on mine safety feature types (such as edge continuity, crack texture, and grayscale features of water accumulation areas); extract features such as texture, contour, and grayscale gradient of risk areas from historical image data to establish a risk area feature library; assign basic high weights to Gaussian kernels corresponding to key safety features based on mine safety feature constraints; for features in the risk area feature library, use feature matching algorithms to locate the corresponding Gaussian kernels in the second 3DGS model, further enhance the reconstruction weight of such Gaussian kernels, and strengthen the feature expression of risk areas; adjust the weight parameters of the target Gaussian kernels through gradient descent algorithm so that the model prioritizes safety features and risk areas during reconstruction; after completing the weight adjustment, verify the model's preliminary reconstruction effect on risk areas to ensure the effectiveness of weight optimization, and output the third 3DGS model.

[0166] Step S34: Optimize and train the third 3DGS model based on the labeled data of the security risk area to obtain the optimized 3DGS model.

[0167] In practical implementation, the monitoring equipment can divide the historical standardized mine observation dataset into a training set and a validation set (e.g., 7:3) and simultaneously import the corresponding safety risk area annotation data (including risk area location and type labels). Key parameters such as the loss function (using a weighted sum of reconstruction error loss and risk area feature loss), learning rate (using an adaptive learning rate strategy), and number of iterations are set for model training. Training is started using the third 3DGS model as the initial model. During training, the model's reconstruction accuracy and risk area identification accuracy are monitored in real time using the validation set. If overfitting occurs (validation set error increases), an early stopping strategy is used to terminate training and adjust parameters. After training, the model's performance is evaluated using a historical standardized dataset that was not used in training, focusing on verifying the risk area reconstruction accuracy and overall scene integrity. Once the evaluation is successful, the final optimized 3DGS model is output.

[0168] In some embodiments, the monitoring equipment can perform mine-specific optimization of traditional 3DGS algorithms:

[0169] In the initialization phase, laser point cloud-guided Gaussian kernel distribution is used to reduce Gaussian kernel redundancy in large homogeneous areas of the mine. This is specifically achieved by constructing an objective function, where the set of sampling points of the laser point cloud in a local region is defined as follows:

[0170]

[0171] in, This represents the set of sampling points in a local region of the laser point cloud. Represents the first point in the point cloud One sampling point, Indicates the first The coordinate components of each sampling point in three-dimensional space The index variable represents the sampling point. This indicates the total number of sampling points in the laser point cloud within a local area.

[0172] The Gaussian kernel set of 3DGS is:

[0173]

[0174] in, Denotes the Gaussian kernel set, Indicates the first A Gaussian kernel, Coordinates of the Gaussian kernel center For the covariance matrix, For transparency;

[0175] The objective function is:

[0176]

[0177] In the objective function, the first term is the distance-weighted error between the laser point cloud and the Gaussian kernel, and the second term is the regularization term. Indicates the number of Gaussian kernels. Let represent the penalty coefficient for the regularization term, and constrain the local density of the laser point cloud. point No new Gaussian kernel is added when the threshold of homogeneous regions is reached.

[0178] During the model rendering stage, mine safety feature constraints were incorporated. High-weight Gaussian kernels were assigned to cracks in areas of abrupt grayscale changes and water accumulation in areas of low grayscale values ​​in the imagery to enhance the accuracy of risk feature reconstruction. The generated 3D scene model met the following requirements: overall reconstruction error ≤ ±5cm, minimum crack width ≥ 2mm, and slope gradient calculation error ≤ ±0.5°. Safety risk areas were labeled using a combination of semantic segmentation and manual verification: cracks and water accumulation areas were automatically segmented using a U-Net network, and high slope deformation areas were manually verified. The labeling results were stored in the model attribute database.

[0179] This embodiment improves model adaptability through Gaussian kernel distribution optimization, ensuring precise matching between the model's Gaussian kernel distribution and the density characteristics of mine point clouds. This avoids the problem of insufficient reconstruction in high-density areas or resource waste in low-density areas by general models, laying the foundation for high-quality reconstruction. Redundant Gaussian kernel removal achieves model lightweighting, reducing computational complexity and subsequent reconstruction time and resource consumption while retaining core reconstruction capabilities, thus improving overall process efficiency. Gaussian kernel weight optimization strengthens risk area reconstruction. By adjusting weights, the model prioritizes mine safety features and risk areas, significantly improving the reconstruction accuracy of these key areas and providing more accurate 3D data for subsequent risk identification. Training based on historical mine annotation data allows the model to fully learn mine scene features, avoiding reconstruction biases common to general 3DGS models in mine scenarios and enhancing the model's practicality and reliability.

[0180] refer to Figure 5 , Figure 5 This is a flowchart illustrating the fourth embodiment of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction of the present invention.

[0181] Based on the above embodiments, in this embodiment, step S34 further includes:

[0182] Step S341: Based on the density distribution characteristics of point cloud data and the texture distribution characteristics of image data in the historical standardized mine observation data, determine the missing point cloud regions and missing image regions in the historical standardized mine observation data.

[0183] It should be noted that missing point cloud areas can be regions in historical standardized mine observation data where the point cloud data density is lower than a preset threshold due to scanning blind spots, occlusion, dust interference, etc., resulting in incomplete 3D geometric information. Missing image areas can be regions in historical standardized mine observation data where image texture information is blurred or missing due to insufficient lighting, dust occlusion, residual dynamic interference features, etc., making it impossible to effectively extract features.

[0184] In practical implementation, the monitoring equipment can spatially divide the point cloud data in the historical standardized data into voxels, count the number of points in each voxel, set a density threshold (based on the density calibration of point cloud from normal mine scanning), and determine the voxel areas with a density lower than the threshold as point cloud missing areas; for the image data in the historical standardized data, the texture gradient and information entropy of each pixel area are calculated, a texture clarity threshold is set, and pixel areas with a texture gradient lower than the threshold or excessive information entropy (blurred areas) are determined as image missing areas; based on the spatial coordinate reference of the UAV attitude data, the point cloud missing areas and the image missing areas are mapped to the same geographic coordinate system to generate spatial distribution maps of the two types of missing areas.

[0185] Step S342: Determine one or more complete missing regions based on the missing regions of the point cloud and the missing regions of the image.

[0186] It should be noted that the complete missing region can be the spatial intersection of the missing point cloud region and the missing image region. This region lacks both geometric and textural information and is the core and difficult region for model reconstruction.

[0187] In practical implementation, the monitoring equipment can use spatial geometric operations to calculate the spatial intersection of the missing areas in the point cloud and the missing areas in the image, and obtain the complete missing areas that lack both geometric and texture information; use edge detection algorithms (such as the Canny algorithm) to extract the boundary coordinates of the complete missing areas, and clarify their spatial range and shape; classify the complete missing areas according to their area size and location (such as whether they are located in high-risk areas such as high slopes), and prioritize the processing of complete missing areas within risk areas.

[0188] Step S343: Based on the region boundary of the complete missing region, filter multiple valid region ranges from the domain of the complete missing region.

[0189] It should be noted that the effective region can be the neighborhood of the complete missing region, a region with sufficient point cloud density, clear image texture, and complete feature information. Its features can be used to transfer and complete the model parameters of the complete missing region.

[0190] In practical implementation, the monitoring equipment can start from the boundary of the complete missing area, set the neighborhood search radius (5-10m, adaptively adjusted according to the size of the missing area) and the screening conditions (point cloud density ≥ normal threshold, image texture gradient ≥ clear threshold); traverse all areas within the search radius, filter out the effective area range that meets the conditions, and ensure that each complete missing area matches at least 3-5 effective areas to avoid the completion deviation caused by single feature transfer; extract the point cloud density distribution features and image texture features (such as grayscale histogram, edge contour) of each effective area range to construct an effective area feature library.

[0191] Step S344: Optimize the Gaussian kernel parameters of the third 3DGS model based on the regional features of the effective region range, and construct a joint loss function that includes geometric continuity constraints and texture consistency constraints;

[0192] Step S345: Based on the joint loss function, use the features of the effective region range to perform gradient descent iterative optimization on the Gaussian kernel generated in the missing region, minimize the geometric tearing error at the boundary of the missing region, and obtain the fourth 3DGS model.

[0193] It should be noted that the fourth 3DGS model can be an intermediate model obtained by optimizing the Gaussian kernel parameters through effective regional feature transfer for complete missing regions and filling in the model features of the missing regions, thus solving the problem of insufficient reconstruction of missing regions by the original model.

[0194] In practical implementation, the monitoring equipment can calculate the similarity between the boundary features (such as boundary curvature and orientation) of the complete missing region and the features of each effective region, and select the 2-3 effective regions with the highest similarity as feature transfer sources; based on the Gaussian kernel parameters (density, scale, weight) of the feature transfer source regions, the interpolation algorithm is used to adjust the Gaussian kernel parameters corresponding to the complete missing region in the third 3DGS model to generate a Gaussian kernel configuration that adapts to the features of the missing region; the optimized Gaussian kernel parameters of the missing region are integrated with the original parameters of the third 3DGS model to generate a fourth 3DGS model; by comparing the feature consistency between the model output after the missing region is filled and the features of the neighboring normal region, the filling effect is verified. If the consistency is insufficient, the effective regions are re-selected and the parameters are optimized.

[0195] Step S346: Optimize and train the fourth 3DGS model based on the labeled data of the security risk area to obtain the optimized 3DGS model.

[0196] It is understandable that this embodiment uses "neighborhood Gaussian kernel interpolation" to supplement local data loss caused by mine dust: for example, the location, color, and scale of Gaussian kernel parameters within a 5m radius around the missing area can be weighted and calculated to generate the Gaussian kernel for the missing area, ensuring the integrity of the reconstructed model.

[0197] In some embodiments, for viewpoint occlusion in dynamic mine scanning, such as ore piles occluding slopes, the monitoring equipment can adopt a "multi-machine collaborative supplementary scanning" strategy: deploy 2-3 drones to simultaneously scan the occluded area from different perspectives, integrate the supplementary scanning data into the 3DGS model, and avoid reconstruction blind spots.

[0198] For example, a multi-drone collaborative scanning strategy can be adopted, deploying 2-3 drones to simultaneously scan occluded areas from different perspectives, integrating the scanned data into the 3DGS model to avoid reconstruction blind spots. Let the drones... The position vector is The attitude angle matrix is (by roll angle) Pitch angle Yaw angle (Computationally generated), drone The point cloud coordinates in the local coordinate system are Coordinates converted to the mine's global coordinate system for:

[0199]

[0200] Based on the ranging accuracy of UAV lidar Point cloud weights ≤ ±2cm The weight calculation formula is:

[0201]

[0202] Merged point cloud coordinates for:

[0203]

[0204] This fusion model avoids reconstruction blind spots and ensures the integrity of regional coverage of the 3DGS model.

[0205] This embodiment improves the accuracy of data quality assessment by identifying missing regions, accurately locating missing regions in point clouds and images, providing clear targets for subsequent model parameter completion, avoiding blind reconstruction of missing regions, and improving overall reconstruction quality. Key missing regions are selected through spatial intersection, prioritizing the handling of missing areas within risk regions to ensure the model's reconstruction accuracy for core safety monitoring areas. Effective region selection provides a reliable basis for parameter optimization: feature support from multiple effective regions avoids the bias of single feature transfer, providing high-quality feature references for Gaussian kernel parameter optimization of missing regions. Gaussian kernel parameter optimization fills model feature gaps: addressing the issue of insufficient reconstruction of missing regions in the 3DGS model, it enhances the model's adaptability to complex mining scenarios.

[0206] Furthermore, this embodiment of the invention also proposes a computer-readable storage medium storing a mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction. When the mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction is executed by a processor, it implements the steps of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described above.

[0207] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0208] The aforementioned computer-readable storage medium may be included in mine safety monitoring equipment based on UAV scanning mapping and 3DGS scene reconstruction; or it may exist independently and not be assembled into mine safety monitoring equipment based on UAV scanning mapping and 3DGS scene reconstruction.

[0209] Furthermore, this invention also proposes a computer program product, including a mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction. When the mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction is executed by a processor, it implements the steps of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described above.

[0210] The specific implementation of the computer program product of the present invention is basically the same as the embodiments of the above-mentioned mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction, and will not be repeated here.

[0211] Reference Figure 6 , Figure 6 This is a structural block diagram of the first embodiment of the mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction of the present invention.

[0212] like Figure 6 As shown, the mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction proposed in this embodiment of the invention includes:

[0213] Data acquisition module 10 is used to acquire multi-view image data, laser point cloud data and UAV attitude data collected by the UAV;

[0214] Data processing module 20 is used to preprocess the multi-view image data and the laser point cloud data based on the UAV attitude data to obtain a standardized mine observation dataset. The preprocessing includes data cleaning and data spatiotemporal alignment.

[0215] The mine scene reconstruction module 30 is used to input the standardized mine observation dataset into the optimized 3DGS model, generate a three-dimensional scene model of the mine monitoring area, and mark the safety risk areas in the three-dimensional scene model;

[0216] The risk identification module 40 is used to extract safety risk parameters from the three-dimensional scene model based on the safety risk area, and to determine the risk level corresponding to each safety risk area based on the safety risk parameters.

[0217] The risk response module 50 is used to send early warning instructions and monitoring strategies corresponding to each risk level to the mine control system in order to conduct safety monitoring and dynamic early warning of the mine monitoring area.

[0218] This embodiment acquires multi-view image data, laser point cloud data, and UAV attitude data collected by a UAV. Based on the UAV attitude data, it preprocesses the multi-view image data and laser point cloud data to obtain a standardized mine observation dataset. Preprocessing includes data cleaning and spatiotemporal alignment. The standardized mine observation dataset is then input into an optimized 3DGS model to generate a 3D scene model of the mine monitoring area. Safety risk areas are marked in the 3D scene model. Safety risk parameters are extracted from the 3D scene model based on these safety risk areas, and the risk level corresponding to each safety risk area is determined based on these parameters. Early warning instructions and monitoring strategies corresponding to each risk level are sent to the mine control system for safety monitoring and dynamic early warning of the mine monitoring area. In this embodiment, multi-view image data and laser point cloud data are preprocessed based on UAV attitude data, achieving unification of image and point cloud data in time and space dimensions. Based on the optimized 3DGS model, a three-dimensional scene model of the mine monitoring area is performed, thereby improving the accuracy and completeness of the mine scene reconstruction, fully restoring the geometric shape and texture features of the mine's static structure, and fully exploring the geometric features and quantitative features of safety risks of the mine's static structure. Early warning instructions and monitoring strategies corresponding to each risk level are sent to the mine control system to conduct safety monitoring and dynamic early warning of the mine monitoring area, forming an automated closed loop of perception-decision-response, greatly improving the speed of risk identification and early warning response, and fundamentally improving the real-time performance, accuracy and initiative of mine safety monitoring.

[0219] The mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction provided in this application adopts the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction in the above embodiments, which can solve the technical problems of mine safety monitoring based on UAV scanning mapping and 3DGS scene reconstruction. Compared with the prior art, the beneficial effects of the mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction provided in this application are the same as the beneficial effects of the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction provided in the above embodiments, and other technical features in the mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction are the same as the features disclosed in the methods of the above embodiments, and will not be repeated here.

[0220] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0221] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0222] In addition, for technical details not described in detail in this embodiment, please refer to the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction provided in any embodiment of the present invention, which will not be repeated here.

[0223] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0224] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0225] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0226] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction, wherein the method is applied to a UAV scanning system deployed in the mine monitoring area, characterized in that, The method includes: Acquire multi-view image data, laser point cloud data, and drone attitude data collected by the drone; Based on the UAV attitude data, the multi-view image data and the laser point cloud data are preprocessed to obtain a standardized mine observation dataset. The preprocessing includes data cleaning and spatiotemporal alignment of data. The standardized mine observation dataset is input into the optimized 3DGS model to generate a three-dimensional scene model of the mine monitoring area, and safety risk areas are marked in the three-dimensional scene model. Based on the security risk areas, security risk parameters are extracted from the 3D scene model, and the risk level corresponding to each security risk area is determined based on the security risk parameters. The warning instructions and monitoring strategies corresponding to each risk level are sent to the mine control system to conduct safety monitoring and dynamic warnings for the mine monitoring area. Before inputting the standardized mine observation dataset into the optimized 3DGS model, the process also includes: Based on the local normal vector distribution characteristics of laser point cloud data, an anisotropic geometric constraint term is constructed; the anisotropic geometric constraint term is used to initialize the covariance matrix of the Gaussian kernel in the original 3DGS model, so that the principal axis direction of the Gaussian kernel is orthogonal to the normal direction of the mine slope surface, and a first 3DGS model is obtained. The first 3DGS model is an intermediate model obtained after Gaussian kernel distribution optimization, and its Gaussian kernel distribution has been adapted to the density characteristics of the mine point cloud data. Redundant Gaussian kernels are removed based on the spatial overlap and feature similarity of adjacent Gaussian kernels in the first 3DGS model to obtain the second 3DGS model. Based on the mine safety feature constraints and the risk area features of image data in historical standardized mine observation data, the second 3DGS model is optimized with Gaussian kernel weights to obtain the third 3DGS model. The mine safety feature constraints are feature restrictions set for the needs of mine safety monitoring, covering the reconstruction accuracy requirements of slope edges, cracks, and water accumulation areas. The risk area features are the texture, grayscale, and outline of the safety risk area, which are used to strengthen the model's reconstruction weight for the safety risk area. Based on the density distribution characteristics of point cloud data and the texture distribution characteristics of image data in the historical standardized mine observation data, the missing point cloud regions and missing image regions in the historical standardized mine observation data are determined. Based on the missing point cloud region and the missing image region, one or more complete missing regions are determined. The complete missing region is the spatial intersection of the missing point cloud region and the missing image region, and this region lacks both geometric and texture information. Based on the region boundary of the complete missing region, multiple valid region ranges are selected from the neighborhood of the complete missing region. The valid region range is the region within the neighborhood of the complete missing region that can be used to transfer and complete the model parameters of the complete missing region. Based on the regional features of the effective region, the Gaussian kernel parameters of the third 3DGS model are optimized to construct a joint loss function that includes geometric continuity constraints and texture consistency constraints. Based on the joint loss function, gradient descent iterative optimization is performed on the Gaussian kernel generated in the missing region using the features of the effective region range to minimize the geometric tearing error at the boundary of the missing region, thus obtaining the fourth 3DGS model. The fourth 3DGS model is an intermediate model obtained by optimizing the Gaussian kernel parameters through effective region feature transfer for the complete missing region and filling in the model features of the missing region, which is used to solve the problem of insufficient reconstruction of the missing region by the original model. The fourth 3DGS model is optimized and trained based on the labeled data of the security risk area to obtain the optimized 3DGS model.

2. The mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described in claim 1, characterized in that, Before acquiring the multi-view image data, laser point cloud data, and UAV attitude data collected by the UAV, the process also includes: A scanning strategy is generated based on the terrain feature data of the mine monitoring area. The terrain feature data includes contour line data and mining range data. The scanning strategy includes: a Z-shaped dense scanning strategy for high slope areas, a spiral scanning strategy for ore pile areas, and a bidirectional parallel scanning strategy for transportation roads. A scan path is generated based on the aforementioned scan strategy; Based on the scanning path, multiple drones are controlled to scan the mine monitoring area and monitor the drone attitude data of each drone. The drones are equipped with industrial cameras, lidar and attitude sensors. During the scanning process, the drones collect multi-view image data and laser point cloud data.

3. The mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described in claim 1, characterized in that, The preprocessing of the multi-view image data and the laser point cloud data based on the UAV attitude data to obtain a standardized mine observation dataset includes: The multi-view image data is subjected to dust denoising processing based on an adaptive bilateral filtering algorithm to obtain candidate images; A depth-guided atmospheric scattering model is constructed based on the depth information of laser point cloud data. The dark channel transmittance of the candidate image is weighted and corrected using point cloud depth constraints. The local atmospheric light value in the non-uniform dust environment of the mine is calculated. Based on the local atmospheric light value, the candidate image is mapped into a clear target image. Based on the spatiotemporal alignment relationship between the target clear image and the laser point cloud data, dynamic interference features that cause displacement between consecutive frames are identified and eliminated through reprojection error analysis and semantic consistency verification, thereby obtaining static mine image data and static mine point cloud data. Based on the UAV attitude data, the static mine image data and the static mine point cloud data are spatiotemporally aligned to obtain a standardized mine observation dataset.

4. The mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described in claim 1, characterized in that, The safety risk parameters include slope gradient, crack size, and water accumulation area; The step of extracting security risk parameters from the 3D scene model based on the security risk region includes: Based on the safety risk area, a plane is fitted to the point cloud of the slope surface in the three-dimensional scene model. The angle between the plane normal vector and the vertical direction is calculated to obtain the slope. Based on the security risk area, the 3D scene model is sampled along the crack annotation path, the length between sampling points and the width of the point cloud on both sides of the crack are calculated, and the crack size is obtained. Based on the safety risk area, the two-dimensional projected area of ​​the water accumulation area marked in the three-dimensional scene model is extracted to obtain the water accumulation area.

5. The mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described in claim 1, characterized in that, After sending the early warning instructions and monitoring strategies corresponding to each risk level to the mine control system, the process also includes: In response to the processing results fed back by the mine control system, multiple drones are controlled to scan the safety risk areas in the mine monitoring area based on the processing results; The 3D scene model is updated based on the data scanning results, and the safety risk parameters are re-extracted based on the updated 3D scene model. Determine whether all safety risk parameters are within the safety threshold range; If all safety risk parameters are within the safety threshold range, the risk is considered eliminated. If at least one safety risk parameter is not within the safety threshold range, the risk level of the safety risk area is updated, and the early warning instructions and monitoring strategies are regenerated based on the updated risk level. Then, the process returns to the step of sending the early warning instructions and monitoring strategies corresponding to each risk level to the mine control system to perform safety monitoring and dynamic early warning for the mine monitoring area.

6. A mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction, applying the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described in any one of claims 1 to 5, characterized in that, The device includes: The data acquisition module is used to acquire multi-view image data, laser point cloud data, and UAV attitude data collected by the UAV; The data processing module is used to preprocess the multi-view image data and the laser point cloud data based on the UAV attitude data to obtain a standardized mine observation dataset. The preprocessing includes data cleaning and data spatiotemporal alignment. The mine scene reconstruction module is used to input the standardized mine observation dataset into the optimized 3DGS model, generate a three-dimensional scene model of the mine monitoring area, and mark the safety risk areas in the three-dimensional scene model; The risk identification module is used to extract safety risk parameters from the three-dimensional scene model based on the safety risk areas, and to determine the risk level corresponding to each safety risk area based on the safety risk parameters. The risk response module is used to send early warning instructions and monitoring strategies corresponding to each risk level to the mine control system in order to conduct safety monitoring and dynamic early warning of the mine monitoring area.

7. A mine safety monitoring device based on UAV scanning mapping and 3DGS scene reconstruction, characterized in that, The mine safety monitoring equipment based on UAV scanning mapping and 3DGS scene reconstruction includes: a memory, a processor, and a mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction stored in the memory. The processor is used to run the mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction. The mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction is configured to implement the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction. When the mine safety monitoring program based on UAV scanning mapping and 3DGS scene reconstruction is executed by a processor, it implements the mine safety monitoring method based on UAV scanning mapping and 3DGS scene reconstruction as described in any one of claims 1 to 5.

Images