Underground mine live-action three-dimensional modeling system and method based on unmanned aerial vehicle SLAM and multi-sensor fusion

Abstract

The invention provides an underground mine live-action three-dimensional modeling system and method based on unmanned aerial vehicle SLAM and multi-sensor fusion. The underground mine live-action three-dimensional modeling system comprises a hardware platform and a software platform. The hardware platform takes an unmanned aerial vehicle as a main body, and integrates a visual sensor, a laser radar, an IMU, a self-adaptive lighting system, a visual laser acquisition control point device, a roof control point identification camera and a wireless communication module; the wireless communication module is used for transmitting the acquired data back to a software system in real time and transmitting remote control instructions among the components; the software platform comprises a front-end SLAM module and a modeling and rendering module; the front-end SLAM module adopts a vision-laser radar-IMU tight coupling SLAM algorithm and fuses control point data acquired by a top plate control point identification camera to realize real-time point cloud registration and map construction of absolute coordinate constraint in a GPS-free environment; the core of the modeling and rendering module is a Mesh and 3DGS double-model parallel generation technology, and the modeling and rendering module has the parallel operation capability of synchronously completing Mesh model generation and 3DGS model training.

Description

Technical Field

[0001] This invention relates to the field of mine surveying and 3D reality modeling technology, specifically to a system and method for 3D reality modeling of underground mines based on UAV SLAM and multi-sensor fusion. Background Technology

[0002] Currently, the creation of realistic models in underground mines mainly relies on handheld SLAM devices. When collecting control point data, a total station is needed to transfer control points from the roadway roof to the ground floor, or a visual laser is used for aiming and data collection. This method has several drawbacks: First, it is inefficient, as the handheld device has a limited range of movement and a long data collection cycle; second, positioning accuracy is difficult to guarantee and is easily affected by human error; third, it poses high safety risks to personnel, as the underground mine environment is complex and contains hazards such as collapses and toxic gases, putting workers at great risk; fourth, it cannot adapt to the dynamic production needs of mines, requiring complete re-collection of data for model updates, which is costly. Furthermore, the lack of GPS signal coverage in underground mines further limits the application of traditional positioning and modeling technologies.

[0003] Therefore, a system and method for real-world 3D modeling of underground mines based on UAV SLAM and multi-sensor fusion is proposed to solve the above problems. Summary of the Invention

[0004] The main objective of this invention is to address the problems existing in the prior art and provide a system and method for real-world 3D modeling of underground mines based on UAV SLAM and multi-sensor fusion.

[0005] The technical solution of the present invention is as follows: a 3D modeling system for underground mine scenes based on UAV SLAM and multi-sensor fusion, including a hardware platform and a software platform;

[0006] The hardware platform includes a drone body, which integrates a visual sensor, lidar, IMU, adaptive lighting system, visual laser acquisition control point device, top control point recognition camera and wireless communication module.

[0007] The roof control point recognition camera is used to detect and identify the identifiable control points on the roof of the roadway; the adaptive lighting system is used to supplement lighting in complex lighting environments in the roadway; the visual laser acquisition control point device is used to accurately acquire roof control point data; and the wireless communication module is used to realize the real-time transmission of acquired data back to the software system and the transmission of remote control commands between various components.

[0008] The software platform includes a front-end SLAM module and a modeling and rendering module;

[0009] The front-end SLAM module adopts a vision-LiDAR-IMU tightly coupled SLAM algorithm, which integrates control point data collected by the top control point recognition camera to achieve real-time point cloud registration and map construction without absolute coordinate constraints in a GPS environment.

[0010] The core of the modeling and rendering module is the parallel generation technology of Mesh and 3DGS dual models, which has the ability to simultaneously complete the parallel computing of Mesh model generation and 3DGS model training.

[0011] Furthermore, the main body of the drone supports remote control flight, is equipped with obstacle avoidance sensors and an emergency processing module, and can realize real-time data transmission.

[0012] When a collision warning is issued during flight, the aircraft will automatically hover or fly around to avoid obstacles. If a signal is interrupted, it will autonomously return to the takeoff point without requiring personnel to enter dangerous tunnel areas.

[0013] Furthermore, the main body of the drone and all integrated sensor components meet the IP65 dustproof and waterproof protection level, are suitable for the high dust and humid environment of underground mines, and have an operating temperature range of -10℃ to 60℃, which can withstand extreme temperature fluctuations in mine tunnels.

[0014] Furthermore, the wireless communication module adopts a hybrid architecture of Mesh self-organizing network + microwave communication. In deep tunnels without base station coverage, the data transmission distance is ≥500m and the transmission rate is ≥10Mbps. It supports real-time data transmission and zero-delay interaction of control commands.

[0015] This invention also provides a method for three-dimensional modeling of underground mine scenes based on UAV SLAM and multi-sensor fusion. Using the above system, the specific method includes the following steps: S1, Establishing absolute positioning benchmarks: Accurately measure the absolute coordinates of at least 3 uniformly arranged marked control points on the top of the roadway using a total station and affix labels. Based on the roadway cross-sectional dimensions and length, plan the optimal flight path including obstacle avoidance nodes and complete sensor calibration, communication link and UAV attitude self-check.

[0016] S2. Integrated data acquisition: UAV takes off → IMU initial attitude calibration + preset control point matching initial positioning → flight along the flight path → real-time SLAM mapping, integrating top control point data → top control point identification → automatic control point acquisition → real-time data transmission → anomaly handling → safe landing process operation, integrating adaptive lighting processing and dust interference suppression during the acquisition process;

[0017] S3. Dual Model Generation and Control Point Constraint Stitching: Mesh model generation and 3DGS model training are completed in parallel. Multi-segment model fusion is achieved through the control point constraint stitching method of multi-segment data acquisition.

[0018] S4. Incremental Update: Based on the local incremental update mechanism of the underground mine scene model, the model identifies the changed areas by comparing the old and new models, and only incrementally collects and updates the model for the changed areas.

[0019] Furthermore, in step S1, the top plate control points are used to establish the absolute coordinate system of the underground mine. The known absolute coordinates of the control points are added to the SLAM optimization map as strong constraints. Based on the control point recognition quality, which is evaluated by a recognition success rate of ≥95% and a single-point positioning error of ≤5cm, the constraint weights are dynamically adjusted to achieve positioning drift suppression.

[0020] Furthermore, in step S2, the adaptive lighting processing automatically adjusts the lighting system parameters and camera exposure parameters according to the ambient light intensity;

[0021] Dust interference suppression processing employs anti-interference point cloud filtering algorithms and image dehazing algorithms to ensure the data quality of laser scanning and image acquisition.

[0022] Furthermore, the Mesh model generation process in step S3 is as follows: point cloud denoising → point cloud registration under control point constraints → triangular mesh generation → texture mapping → model optimization.

[0023] Absolute coordinate constraints of top plate control points are incorporated into the 3DGS model training process to ensure that the model's positioning accuracy is consistent with that of the Mesh model.

[0024] The multi-segment model fusion uses the top plate control points as a reference, and eliminates the cumulative error of multi-segment data acquisition through the control point constraint splicing method, with a splicing error ≤8cm.

[0025] Furthermore, the core of the local incremental update mechanism in step S4 is: through the detection results of geometric deviation ≥10cm or volume change ≥1m³ between the old and new models, the area of ​​change in the mining environment is accurately identified, and only the area of ​​change and the outer 5m range are re-flyed for data acquisition and local model update, without the need for overall re-mining and reconstruction.

[0026] Compared with the prior art, the present invention has the following advantages: 1. The present invention uses a drone as a mobile platform, laying the hardware foundation for autonomous operation. Its multi-sensor integrated design solves the problems of limited mobility and scattered data collection of traditional handheld devices. The path adaptive planning technology in the construction method automatically generates the optimal collection path by analyzing the roadway environment data in real time, avoiding repeated collection and path redundancy. Combined with the multi-sensor synchronous collection mechanism, the data collection process is transformed from a discrete mode of manual movement-single point collection-one-by-one recording to a continuous mode of autonomous cruise-synchronous collection-direct data transmission. At the same time, the automatic control point matching unit of the system eliminates manual mapping operations. The construction method further optimizes the response speed of the control point registration algorithm, which greatly improves the registration efficiency. The two work together to significantly shorten the overall data collection cycle compared with the traditional method, thereby reducing the time cost of operation.

[0027] 2. This invention constructs a multi-source fusion positioning core consisting of LiDAR, inertial navigation, and visual sensors. By complementing data from multiple sensors, it solves the positioning deviation problem of a single sensor in the absence of GPS. The error compensation model in the construction method performs targeted algorithm corrections for specific defects such as cumulative drift of inertial navigation and point cloud noise of LiDAR. At the same time, it enhances the feature point matching accuracy of point cloud registration, reducing the consistency error of multi-source data fusion to ≤1cm. The two work together to achieve dual protection of hardware complementarity and algorithm correction, ultimately achieving ultra-high precision positioning of ≤3cm in the absence of GPS signal, ensuring the coordinate uniformity, geometric accuracy, and detail reproduction of the real scene model, and meeting the high-precision specifications for underground mine surveying.

[0028] 3. The system of this invention adopts an unmanned operation mode, which fundamentally avoids workers from entering high-risk roadways and solves the core safety hazards of traditional manual operation, such as direct exposure to collapse and toxic gases. At the same time, the environmental monitoring and early warning module of the construction method is deeply integrated with the unmanned mobile platform to collect environmental data such as gas concentration and roof stability in the roadway in real time. Once the risk threshold is detected to exceed the standard, the platform is immediately triggered to automatically evacuate through the remote emergency control mechanism. This forms a full-process safety protection system that isolates risks through unmanned operation, monitors and warns of risks in real time, and avoids risks through emergency control. It completely eliminates the risk of personnel casualties and complies with the safety management specifications for high-risk environment operations in underground mines.

[0029] 4. The incremental data update module of this invention establishes the core idea of ​​local updates rather than overall reconstruction, solving the high cost problem of full data collection and repetitive modeling in traditional model updates. At the same time, the construction method adopts a multi-time series data comparison algorithm to accurately identify local change areas such as roadway deformation and equipment displacement in dynamic mine production by comparing real-time collected data with historical model data at the pixel level. Combined with the local update triggering mechanism, it only triggers targeted collection of change areas to avoid invalid data collection. The two work together to reduce the amount of data for model updates by more than 70%, significantly reducing equipment wear and manpower costs for modeling and maintenance, while shortening the update cycle and ensuring that the model can reflect the current status of mine production in real time and adapt to dynamic production needs.

[0030] 5. The multi-sensor fusion positioning scheme of this invention breaks the dependence of traditional technology on GPS signals. It constructs a positioning foundation in the absence of GPS by using environmental perception of lidar, motion state monitoring of inertial navigation, and feature point matching of visual sensors. At the same time, the construction method optimizes the inertial navigation drift correction algorithm for the special needs of the environment without satellite signals and improves the positioning stability in complex tunnel environments (such as darkness and dust obstruction) through visual feature point enhancement matching technology, which greatly improves the continuity of positioning. It solves the pain points of positioning interruption and sharp drop in accuracy of traditional positioning technology in the absence of GPS. The system and method work together to achieve continuous, stable and high-precision positioning in the absence of satellite signals, covering various complex tunnel scenarios in underground mines, and breaking through the application limitations of traditional technology. Attached Figure Description

[0031] Figure 1 This is a system architecture diagram of the present invention;

[0032] Figure 2 This is a flowchart of the present invention. Detailed Implementation

[0033] The implementation of the technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0034] like Figure 1 As shown, the underground mine real-scene 3D modeling system based on UAV SLAM and multi-sensor fusion includes a hardware platform and a software platform.

[0035] The hardware platform is based on a drone and integrates a visual sensor, lidar, IMU, adaptive lighting system, visual laser acquisition control point device, top control point recognition camera and wireless communication module;

[0036] The roof control point recognition camera is used to detect and identify the identifiable control points on the roof of the roadway, with a recognition success rate of ≥95% and a single-point positioning error of ≤5cm. It supports stable recognition in environments with dust concentrations of ≤100mg / m³. The adaptive lighting system is used to supplement lighting in complex lighting environments in the roadway. The visual laser acquisition control point device is used to accurately acquire roof control point data. The wireless communication module is used to realize the real-time transmission of acquired data back to the software system and the transmission of remote control commands between various components.

[0037] The software platform includes a front-end SLAM module and a modeling and rendering module;

[0038] The front-end SLAM module adopts a vision-LiDAR-IMU tightly coupled SLAM algorithm, which integrates control point data collected by the top control point recognition camera to achieve real-time point cloud registration and map construction without absolute coordinate constraints in a GPS environment.

[0039] The core of the modeling and rendering module is the parallel generation technology of Mesh and 3DGS dual models, which has the ability to simultaneously complete the parallel computing of Mesh model generation and 3DGS model training.

[0040] Priority is also given to the data preprocessing module, which integrates anti-interference point cloud filtering algorithms, image dehazing algorithms, and point cloud noise reduction functions to suppress interference and optimize the quality of raw data collected by LiDAR and vision sensors.

[0041] The system hardware adopts a multi-sensor integrated design, and the software supports the fully automated process of data preprocessing, SLAM mapping, and parallel generation of dual models. It can flexibly adapt to mine roadway scenarios with different cross-sectional dimensions and lengths. The Mesh model meets the precise analysis needs of engineering surveying, while the 3DGS model is adapted for applications such as visualization and virtual simulation. The parallel output of dual models can support multiple application scenarios such as mine excavation planning, equipment installation and positioning, surrounding rock deformation monitoring, and emergency rescue route planning, providing comprehensive technical support for the digital and intelligent construction of underground mines.

[0042] Furthermore, the drone body supports remote control flight, is equipped with obstacle avoidance sensors and emergency response modules, and can realize real-time data transmission.

[0043] Preferably, the obstacle avoidance sensor is a fusion module of lidar and millimeter-wave radar, which can detect static obstacles and dynamic interference within a range of 0.5-50m in front, with an obstacle avoidance response time of ≤100ms and a flight path deviation of ≤30cm.

[0044] When a collision warning is issued during flight, the aircraft will automatically hover or fly around to avoid obstacles. If a signal is interrupted, it will autonomously return to the takeoff point without requiring personnel to enter dangerous tunnel areas.

[0045] The drone supports remote control flight and is equipped with a lidar + millimeter-wave radar fusion obstacle avoidance module, which can detect obstacles within a range of 0.5-50m in real time. The obstacle avoidance response is fast and accurate. Combined with the emergency handling mechanism of automatic hovering / detour after collision warning and autonomous return after signal interruption, it eliminates the need for operators to enter dangerous roadway areas, fundamentally avoiding safety risks such as rockfalls, collapses, and toxic and harmful gases in deep mines, and significantly improving the safety factor of mine surveying operations.

[0046] Furthermore, the main body of the drone and all integrated sensor components meet the IP65 dustproof and waterproof protection level, which is suitable for the high dust and humid environment of underground mines, and the working temperature range is -10℃ to 60℃, which can withstand the extreme temperature fluctuations in mine tunnels.

[0047] Preferably, the drone is equipped with a high-capacity explosion-proof battery with a flight time of ≥90 minutes, supporting continuous data collection in a single run through a tunnel of ≥2km without the need for mid-flight return for charging. Combined with optimal flight path planning including obstacle avoidance nodes, the data collection cycle is significantly shortened.

[0048] The drone body and sensor components have a protection capability of ≥5J energy impact, and adopt a high-strength polycarbonate shell and silicone buffer structure, which is suitable for impact scenarios such as rockfalls in tunnels and equipment vibrations.

[0049] The drone body and sensor components meet the IP65 dustproof and waterproof rating, have an operating temperature range of 10℃ to 60℃, and are equipped with ≥5J energy impact protection, making them resistant to harsh environments such as high dust, humidity, extreme temperature fluctuations, and rockfall impacts in mines.

[0050] Furthermore, the wireless communication module adopts a hybrid architecture of Mesh self-organizing network + microwave communication. In deep tunnels without base station coverage, the data transmission distance is ≥500m and the transmission rate is ≥10Mbps. It supports real-time data back transmission and zero-delay interaction of control commands. The hybrid architecture of Mesh self-organizing network + microwave communication enables real-time data back transmission in deep tunnels (transmission rate ≥10Mbps, distance ≥500m), avoiding redundant operations of secondary data import.

[0051] like Figure 2 As shown, the method for 3D modeling of underground mine scenes based on UAV SLAM and multi-sensor fusion, based on the above system, specifically includes the following steps:

[0052] S1. Establishment of absolute positioning benchmark: Accurately measure the absolute coordinates of at least 3 uniformly distributed marked control points on the roof of the tunnel using a total station and affix labels. Based on the tunnel cross-section size and length, plan the optimal flight path including obstacle avoidance nodes and complete sensor calibration, communication link and UAV attitude self-check.

[0053] S2. Integrated data acquisition: UAV takes off → IMU initial attitude calibration + preset control point matching initial positioning → flight along the flight path → real-time SLAM mapping, integrating top control point data → top control point identification → automatic control point acquisition → real-time data transmission → anomaly handling → safe landing process operation, integrating adaptive lighting processing and dust interference suppression during the acquisition process;

[0054] S3. Dual Model Generation and Control Point Constraint Stitching: Mesh model generation and 3DGS model training are completed in parallel. Multi-segment model fusion is achieved through the control point constraint stitching method of multi-segment data acquisition.

[0055] S4. Incremental Update: Based on the local incremental update mechanism of the underground mine scene model, the model identifies the changed areas by comparing the old and new models, and only incrementally collects and updates the model for the changed areas.

[0056] Furthermore, the roof control points in step S1 are used to establish the absolute coordinate system of the underground mine. The known absolute coordinates of the control points are added to the SLAM optimization map as strong constraints. Based on the control point recognition quality, the recognition quality is evaluated by a recognition success rate of ≥95% and a single-point positioning error of ≤5cm. The constraint weights are dynamically adjusted to achieve positioning drift suppression.

[0057] Furthermore, in step S2, the adaptive lighting processing automatically adjusts the lighting system parameters and camera exposure parameters based on the ambient light intensity;

[0058] Dust interference suppression processing employs anti-interference point cloud filtering algorithms and image dehazing algorithms to ensure the data quality of laser scanning and image acquisition.

[0059] Furthermore, the Mesh model generation process in step S3 is as follows: point cloud denoising → point cloud registration under control point constraints → triangular mesh generation → texture mapping → model optimization;

[0060] Absolute coordinate constraints of top plate control points are incorporated into the 3DGS model training process to ensure that the model's positioning accuracy is consistent with that of the Mesh model.

[0061] Multi-segment model fusion uses the top plate control points as a reference. The cumulative error of multi-segment data acquisition is eliminated by the control point constraint splicing method, and the splicing error is ≤8cm.

[0062] Preferably, the absolute positioning error of the overall real-scene 3D model is ≤10cm, the planar accuracy is ≤8cm, and the elevation accuracy is ≤12cm, which meets the accuracy requirements for underground mine excavation planning and safety monitoring.

[0063] Furthermore, the core of the local incremental update mechanism in step S4 is to accurately identify the areas of change in the mining environment by using the detection results of geometric deviation ≥10cm or volume change ≥1m³ between the old and new models.

[0064] Preferably, the comparison between the new and old models adopts a point cloud registration algorithm based on feature point matching and AI semantic segmentation technology, with a comparison accuracy of ≥98%, which can accurately identify areas of change such as roadway deformation, equipment displacement, and surrounding rock spalling;

[0065] Only the changed area and an outer 5m range are re-acquired and the model is locally updated; no overall re-acquisition and reconstruction is required.

[0066] The local incremental update mechanism uses precise detection of geometric deviation ≥10cm or volume change ≥1m³ to re-update only the changed area and an outer 5m range, eliminating the need for overall re-mining and reconstruction. This significantly reduces the cost of repetitive operations and improves the efficiency of dynamic monitoring of the mine environment.

[0067] The following will use three specific underground mines as examples for detailed explanation:

[0068] Example 1: Realistic Modeling of a High-Dust, High-Gas Underground Coal Mine

[0069] Implementation Scenario: An underground anthracite coal mine with a main mining roadway cross-section of 5m×4m and a total length of 3km. The dust concentration in the roadway is maintained at 80-100mg / m³ year-round, posing a risk of gas leakage. Some areas are humid and the temperature fluctuates between -5℃ and 45℃. Traditional handheld SLAM data acquisition results in data distortion due to dust interference, and manual entry poses risks of gas poisoning and collapse. Therefore, high-precision modeling and safety monitoring without human intervention are required.

[0070] System configuration adaptation:

[0071] Hardware platform:

[0072] The drone body is equipped with an IP65 dustproof and waterproof shell and a 5J energy impact protection structure, making it suitable for high dust and rock impact environments.

[0073] Sensor integration: a tightly coupled module of visual sensor + LiDAR + IMU, paired with a top control point recognition camera and a LiDAR + millimeter-wave radar fusion obstacle avoidance module;

[0074] Auxiliary components: Adaptive lighting system (supports 0-1000 lux light adjustment), explosion-proof high-capacity battery (105 minutes of continuous use), Mesh + microwave hybrid communication module (transmission distance 600m);

[0075] Software platform:

[0076] Front-end SLAM: Vision-LiDAR-IMU tightly coupled algorithm, incorporating anti-interference point cloud filtering and image dehazing algorithms;

[0077] Modeling module: Mesh and 3DGS dual models are generated in parallel, and control point constraint stitching is supported.

[0078] Specific implementation steps:

[0079] S1. Establishing the absolute positioning benchmark: Use a total station to measure 4 evenly distributed marked control points on the roof, with an absolute coordinate accuracy of ±2cm, and affix highly reflective labels; plan the optimal path with 12 obstacle avoidance nodes (avoiding roadway support pillars and water accumulation areas), and complete sensor calibration and communication link testing.

[0080] S2, Integrated Data Acquisition:

[0081] After the UAV takes off, the IMU completes the initial attitude calibration (error ≤ 0.1°) and matches the preset control points to achieve initial positioning;

[0082] While flying along the path, the adaptive lighting system adjusts the exposure parameters according to the light intensity of the alley (minimum 30 lux), and the anti-interference point cloud filtering algorithm suppresses dust interference in real time;

[0083] The top control point identification camera automatically captures the tags, and the visual laser acquisition device accurately collects data (single point positioning error ≤ 4cm). The data is transmitted back in real time through the Mesh network (transmission rate 12Mbps).

[0084] When the aircraft flew to a distance of 2.5km, it detected that the gas concentration exceeded the standard by ≥0.8%. The emergency response module triggered automatic hovering → obstacle avoidance → return to a safe area. The remaining tunnel data was then collected in segments.

[0085] S3. Dual Model Generation and Fusion: Mesh model (triangular mesh density 0.05m) and 3DGS model are generated in parallel. Three data segments are stitched together based on four top plate control points, with a stitching error of ≤6cm. The overall model has an absolute positioning error of ≤9cm, a planar accuracy of ≤7cm, and an elevation accuracy of ≤11cm.

[0086] S4. Incremental Update: Dynamic monitoring is carried out once a month. By comparing the old and new models through AI semantic segmentation technology, two areas of rock spalling in the roadway (geometric deviation ≥12cm) are identified. Only the area and the surrounding 5m range are re-sampled, and the update cycle is shortened to 8 hours.

[0087] Implementation results:

[0088] Safety aspects: No personnel enter the tunnels throughout the entire process, avoiding the risks of gas poisoning and collapse, increasing the safety factor by 100%;

[0089] In terms of accuracy: the data acquisition efficiency in dusty environments is ≥95%, and the model accuracy meets the requirements of coal mine mining planning and gas drainage borehole positioning.

[0090] In terms of efficiency: data collection in a 3km alley took 2.5 hours (compared to 12 hours for traditional handheld SLAM), reducing the collection cycle by 79%;

[0091] Environmental adaptability: Stable operation within the range of 80-100mg / m³ dust and -5℃ to 45℃, with a 100% failure rate.

[0092] This embodiment adopts an unmanned operation mode to fundamentally avoid workers entering high-risk roadways, solving the core safety hazards of traditional manual operation, such as direct exposure to collapse and toxic gases. At the same time, the environmental monitoring and early warning module of the construction method is deeply integrated with the unmanned mobile platform to collect environmental data such as gas concentration and roof stability in the roadway in real time. Once the risk threshold is detected to exceed the standard, the platform is immediately triggered to automatically evacuate through the remote emergency control mechanism, forming a full-process safety protection system of unmanned operation to isolate risks, real-time monitoring and early warning of risks, and emergency control to avoid risks, completely eliminating the risk of personnel casualties and complying with the safety management specifications for high-risk environment operations in underground mines.

[0093] Example 2: Modeling of Long-Distance Complex Cross-Section Underground Metal Mines

[0094] Implementation scenario: An underground iron mine with a main transport roadway of 5km in length and irregular cross-sectional dimensions (gradually changing from 3m×3m to 6m×5m). There are 3 sharp bends (radius of curvature ≤10m), 2 waterlogged areas (water depth ≤30cm), and some sections with residual fallen rocks. It is necessary to achieve long-distance continuous modeling and equipment installation positioning adaptation. Traditional methods cannot meet the requirements due to path redundancy and large multi-segment splicing errors.

[0095] System configuration adaptation:

[0096] Hardware platform:

[0097] Drone body: high-strength polycarbonate shell + silicone cushioning structure, flight time of 120 minutes, supports continuous data collection up to 2.5km in a single run;

[0098] Sensor integration: obstacle avoidance module, LiDAR + millimeter-wave radar, detection range 0.5-50m, obstacle avoidance response time ≤90ms; top control point recognition camera (supports low light recognition, recognition success rate ≥96% when illumination ≤50lux).

[0099] Communication module: Mesh self-organizing network + microwave hybrid architecture, transmission rate of 15Mbps, transmission distance of 550m in areas without base station coverage;

[0100] Software platform:

[0101] Path planning: Adaptive path algorithm (supports dynamic adjustment of flight altitude and speed across irregular cross-sections);

[0102] Modeling module: Multi-segment model control point constraint splicing algorithm, splicing error ≤8cm, overall positioning error ≤10cm.

[0103] Implementation steps

[0104] S1. Baseline Establishment and Path Planning: Six control points are evenly distributed on the top plate, and the absolute coordinates are measured with a total station. Based on the gradual change characteristics of the cross section, a segmented path is planned, consisting of cruising on straight sections (speed 3m / s), low-speed detouring on sharp curves (speed 1m / s), and elevated flight in waterlogged areas (height 5m). Eight obstacle avoidance nodes are set (for rockfall areas).

[0105] S2. Segmented Data Acquisition:

[0106] The drone collected data in two segments (2.5km each). After takeoff, the first segment completed IMU calibration and control point matching, and flew autonomously along the path. The adaptive lighting system compensated for the illumination in the dark areas of the alley, and the anti-interference algorithm suppressed the interference of water reflection.

[0107] The collected data is transmitted back in real time via the Mesh network. The second segment of data collection is initialized based on the control point at the end of the first segment to avoid accumulated errors.

[0108] When encountering a falling rock obstacle (3m away), the obstacle avoidance module triggers a detour (path offset ≤25cm), and there is no data collection interruption.

[0109] S3, Multi-segment Model Fusion:

[0110] Two tunnel sections are generated in parallel: a Mesh model (for equipment installation and positioning) and a 3DGS model (for visualization).

[0111] Using six control points on the top plate as constraint benchmarks, the two model segments were spliced ​​together using a point cloud registration algorithm, with a splicing error of ≤7.5cm.

[0112] The final model has a planar accuracy of ≤7.8cm and an elevation accuracy of ≤11.5cm, which meets the ±10cm accuracy requirement for the installation and positioning of mining equipment.

[0113] S4. Subsequent maintenance and updates: A full tunnel scan is conducted every quarter. By comparing the old and new models, areas where equipment has shifted (geometric deviation ≥10cm) are identified. Only these areas are re-sampled and updated, with each update taking 1.5 hours.

[0114] Implementation results:

[0115] Adaptability: Successfully adapts to complex scenarios such as irregular cross-sections, sharp bends, and water accumulation, with no blind spots in data collection;

[0116] Efficiency: Total data collection time for a 5km tunnel was 4 hours (compared to 20 hours using traditional methods), representing an 80% increase in efficiency;

[0117] Accuracy: Multi-segment splicing error ≤8cm, meeting equipment installation and positioning requirements, no secondary calibration required;

[0118] Cost: The long-range design avoids mid-journey return, reducing equipment wear and tear by 40%.

[0119] This embodiment system constructs a multi-source fusion positioning core consisting of LiDAR, inertial navigation, and visual sensors. By complementing data from multiple sensors, it solves the positioning deviation problem of a single sensor in the absence of GPS. The error compensation model in the construction method performs targeted algorithm corrections for specific defects such as cumulative drift of inertial navigation and point cloud noise of LiDAR. At the same time, it enhances the feature point matching accuracy of point cloud registration, reducing the consistency error of multi-source data fusion to ≤1cm. The two work together to achieve dual protection of hardware complementarity and algorithm correction, ultimately achieving ultra-high precision positioning of ≤3cm in the absence of GPS signal, ensuring the coordinate uniformity, geometric accuracy, and detail reproduction of the real scene model, and meeting the high-precision specifications for underground mine surveying.

[0120] Example 3: Dynamic Monitoring Modeling of Underground Mine Working Face

[0121] Implementation scenario: An underground copper mine with a fast mining face advancement speed (150m per month), a tunnel length of 2km, and a cross-section of 4m×3.5m. The real-world model needs to be updated monthly to guide mining planning. Traditional whole-body remining methods are time-consuming and costly, and the surrounding rock deformation risk of the working face is high, so it is necessary to accurately identify the changing areas.

[0122] System configuration adaptation:

[0123] Hardware platform:

[0124] Drone body: Lightweight design (weight ≤3kg), flight time 90 minutes, supports close-range, close-to-the-wall flight (≥0.8m from the tunnel wall).

[0125] Sensor integration: high-resolution visual sensor + LiDAR (point cloud density ≥100 points / cm²), top control point recognition camera (recognition success rate ≥98%).

[0126] Communication module: Mesh self-organizing network communication, transmission rate of 12Mbps, supporting stable transmission in complex electromagnetic environments at mining faces;

[0127] Software platform:

[0128] Core algorithms: AI semantic segmentation technology (accuracy of changing region recognition ≥99%), local incremental update triggering mechanism (geometric deviation ≥10cm or volume change ≥1m³).

[0129] Modeling module: Parallel generation of dual models, supporting seamless integration of local models and historical models.

[0130] Implementation steps:

[0131] S1. Initial benchmark establishment: Set up 3 control points (600m apart) on the roof of the roadway, measure the absolute coordinates with a total station, plan the reciprocating acquisition path along the direction of the mining face, and complete the sensor calibration.

[0132] S2, Initial Modeling (First Month):

[0133] The drone flies along the path, collects data from the entire tunnel, and incorporates control point constraints through a tightly coupled SLAM algorithm to generate an initial Mesh+3DGS dual model with a positioning error of ≤9cm.

[0134] The initial data collection took 2 hours, and the model was used as a benchmark for mining planning.

[0135] S3, Monthly Incremental Update (Second Month):

[0136] The drone only collects data on newly added mining faces (150m) and suspected deformation areas of existing roadways (predicted based on historical data);

[0137] After data preprocessing, AI semantic segmentation technology was used to compare the old and new models, accurately identifying 3 areas of surrounding rock deformation (geometric deviation 12-18cm) and 2 areas of equipment displacement (volume change ≥1.2m³), with an accuracy rate of 99.2%.

[0138] Only the changed area and the outer 5m range (a total of 200m section) were resampled, and the sampling time was 40 minutes.

[0139] S4, Model Update and Fusion:

[0140] Incremental data collection is integrated with historical models to update the Mesh model (for deformation analysis) and the 3DGS model (for planning visualization).

[0141] The updated model has an overall positioning error of ≤10cm and a measurement accuracy of ≤8cm in the deformation area, meeting the dynamic adjustment requirements of mining planning.

[0142] Implementation results:

[0143] Dynamic adaptation: accurately identifies areas of change such as surrounding rock deformation and equipment displacement, with an accuracy rate of ≥99%;

[0144] Efficiency: Incremental updates reduce time by 75% compared to overall resampling (40 minutes vs. 2 hours);

[0145] Costs: Data collection volume reduced by 80%, labor and equipment depreciation costs reduced by 65%;

[0146] Safety: Reduces the time drones spend in high-risk mining faces, lowering operational risks by 50%.

[0147] This embodiment's incremental data update module establishes the core idea of ​​local updates rather than overall reconstruction, solving the high cost problem of traditional model updates involving full data collection and repetitive modeling. Simultaneously, the construction method employs a multi-time-series data comparison algorithm to accurately identify localized changes in mine dynamics, such as roadway deformation and equipment displacement, by comparing real-time collected data with historical model data at the pixel level. Combined with a local update triggering mechanism, targeted data collection is triggered only in the changed areas, avoiding invalid data collection. The combined effect reduces the amount of data for model updates by more than 70%, significantly reducing equipment wear and manpower costs for modeling and maintenance, while shortening the update cycle. This ensures the model can reflect the mine's current production status in real time and adapt to dynamic production needs.

Claims

1. A 3D modeling system for underground mines based on UAV SLAM and multi-sensor fusion, characterized in that: This includes both hardware and software platforms; The hardware platform includes a drone body, which integrates a visual sensor, lidar, IMU, adaptive lighting system, visual laser acquisition control point device, top control point recognition camera and wireless communication module. The roof control point recognition camera is used to detect and identify the identifiable control points on the roof of the roadway; the adaptive lighting system is used to supplement lighting in complex lighting environments in the roadway; the visual laser acquisition control point device is used to accurately acquire roof control point data; and the wireless communication module is used to realize the real-time transmission of acquired data back to the software system and the transmission of remote control commands between various components. The software platform includes a front-end SLAM module and a modeling and rendering module; The front-end SLAM module adopts a vision-LiDAR-IMU tightly coupled SLAM algorithm, which integrates control point data collected by the top control point recognition camera to achieve real-time point cloud registration and map construction without absolute coordinate constraints in a GPS environment. The modeling and rendering module is used for parallel generation of Mesh and 3DGS dual models, enabling parallel computation of simultaneously generating Mesh models and training 3DGS models.

2. The underground mine real-scene 3D modeling system based on UAV SLAM and multi-sensor fusion as described in claim 1, characterized in that: The main body of the UAV supports remote control flight and is equipped with obstacle avoidance sensors and an emergency processing module. The emergency processing module is used to enable the UAV to automatically hover or fly around obstacles when a collision warning occurs during flight, and to autonomously return to the takeoff point when a signal interruption occurs.

3. The underground mine real-scene 3D modeling system based on UAV SLAM and multi-sensor fusion as described in claim 1, characterized in that: The main body of the drone and all integrated sensor components meet the IP65 dustproof and waterproof protection level, and the operating temperature range is -10℃ to 60℃.

4. The 3D modeling system for underground mines based on UAV SLAM and multi-sensor fusion as described in claim 1, characterized in that: The wireless communication module adopts a hybrid architecture of Mesh self-organizing network and microwave communication. In deep tunnels without base station coverage, the data transmission distance is ≥500m and the transmission rate is ≥10Mbps. It supports real-time data transmission and zero-delay interaction of control commands.

5. A method for 3D modeling of underground mine scenes based on UAV SLAM and multi-sensor fusion, characterized by: The system according to any one of claims 1-4 specifically includes the following steps: S1. Establishment of absolute positioning benchmark: Accurately measure the absolute coordinates of at least 3 uniformly distributed marked control points on the roof of the tunnel using a total station and affix labels. Based on the tunnel cross-section size and length, plan the optimal flight path including obstacle avoidance nodes and complete sensor calibration, communication link and UAV attitude self-check. S2. Integrated data acquisition: UAV takes off → IMU initial attitude calibration + preset control point matching initial positioning → flight along the flight path → real-time SLAM mapping, integrating top control point data → top control point identification → automatic control point acquisition → real-time data transmission → anomaly handling → safe landing process operation, integrating adaptive lighting processing and dust interference suppression during the acquisition process; S3. Dual Model Generation and Control Point Constraint Stitching: Mesh model generation and 3DGS model training are completed in parallel. Multi-segment model fusion is achieved through the control point constraint stitching method of multi-segment data acquisition. S4. Incremental Update: Based on the local incremental update mechanism of the underground mine scene model, the model identifies the changed areas by comparing the old and new models, and only incrementally collects and updates the model for the changed areas.

6. The method for 3D modeling of underground mine scenes based on UAV SLAM and multi-sensor fusion according to claim 5, characterized in that: In step S1, the top plate control points are used to establish the absolute coordinate system of the underground mine. The known absolute coordinates of the control points are added to the SLAM optimization map as strong constraints. Based on the control point recognition quality, which is evaluated by a recognition success rate of ≥95% and a single-point positioning error of ≤5cm, the constraint weights are dynamically adjusted to achieve positioning drift suppression.

7. The method for 3D modeling of underground mine scenes based on UAV SLAM and multi-sensor fusion as described in claim 5, characterized in that: The adaptive lighting processing in step S2 automatically adjusts the lighting system parameters and camera exposure parameters according to the ambient light intensity. Dust interference suppression processing employs anti-interference point cloud filtering algorithms and image dehazing algorithms to ensure the data quality of laser scanning and image acquisition.

8. The method for 3D modeling of underground mine scenes based on UAV SLAM and multi-sensor fusion according to claim 5, characterized in that: The Mesh model generation process in step S3 is as follows: point cloud denoising → point cloud registration under control point constraints → triangular mesh generation → texture mapping → model optimization. Absolute coordinate constraints of top plate control points are incorporated into the 3DGS model training process to ensure that the model's positioning accuracy is consistent with that of the Mesh model. The multi-segment model fusion uses the top plate control points as a reference, and eliminates the cumulative error of multi-segment data acquisition through the control point constraint splicing method, with a splicing error ≤8cm.

9. The method for 3D modeling of underground mine scenes based on UAV SLAM and multi-sensor fusion according to claim 5, characterized in that: The core of the local incremental update mechanism in step S4 is: by using the detection results of geometric deviation ≥10cm or volume change ≥1m³ between the old and new models, the area of ​​change in the mining environment is accurately identified, and re-flying data collection and local model update are performed only on the area of ​​change and an outer 5m range.

Images