Mine area grid map generation method and system for unmanned mine card, terminal and medium

Abstract

The invention relates to the field of unmanned driving, and particularly provides a mining area grid map generation method and system for an unmanned mine card, a terminal and a medium, and the method comprises the steps: collecting the static environment data of a loading area through a multi-source sensor disposed on the mine card, and generating an initial point cloud map through denoising, time synchronization and SLAM fusion processing based on graph optimization; and an initial grid map is exported through automatic point cloud rasterization classification and boundary extraction. On the basis, establishing a real-time updating mechanism: continuously collecting dynamic operation data, extracting current loading boundary characteristics, and carrying out geometric comparison and area change judgment on the current loading boundary characteristics and an initial boundary; and if the change exceeds a threshold value, triggering an incremental updating process, only performing local recalculation and map coverage on a change area, and finally outputting an updated grid map. According to the invention, high-precision automatic construction and efficient real-time maintenance of the grid map of the mining area are realized, and the safety and efficiency of unmanned mine card operation are improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving, and specifically to a method, system, terminal, and medium for generating a grid map of a mining area for an autonomous mining truck. Background Technology

[0002] In mining operations, accurate mapping of the loading area is crucial. As a key area for production, the accuracy of the map information in the loading area directly affects the scheduling of transport vehicles, route planning, and the efficiency and safety of the entire mining operation. Existing methods for creating mine maps have several shortcomings when dealing with the complex and dynamically changing loading area. Firstly, maps generated by traditional methods often lack sufficient precision, failing to accurately reflect subtle terrain changes, equipment locations, and drivable zones. This makes it easy for unmanned or manually driven vehicles to deviate from their paths and increase the risk of collisions when operating in the loading area. Secondly, existing methods are not timely or efficient in updating loading area maps. Because the terrain of the loading area constantly changes during excavation and loading operations, traditional methods struggle to quickly capture these changes and update the map in a timely manner, leading to a disconnect between the map and the actual situation, further hindering the smooth operation of mining operations. Summary of the Invention

[0003] To address the aforementioned issues, this invention provides a method, system, terminal, and medium for generating a grid map of a mining area for unmanned mining trucks. Through multi-source sensor fusion and incremental dynamic updates, it achieves high-precision automatic construction and efficient real-time maintenance of the mining area grid map, thereby improving the safety and efficiency of unmanned mining truck operations.

[0004] The present invention provides a method for generating a grid map of a mining area for unmanned mining trucks, as well as a system, terminal, and medium for executing the method. The method includes the following steps: By deploying multi-source sensors on unmanned vehicles in the mining area, static environmental data of the loading area is collected before equipment operation; the static environmental data includes first GNSS data, first IMU data, and first point cloud data generated by lidar scanning; The first point cloud data is denoised, then the first GNSS data and the first IMU data are fused together, and then a graph-optimized SLAM mapping algorithm is used to generate an initial point cloud map in the world coordinate system. Based on the map editing platform, the initial point cloud map is loaded, the initial map boundary is drawn according to the boundary features of the point cloud map of the mining area loading area, and the initial raster map is exported. Based on the initial grid map, a real-time update mechanism is established. This mechanism is configured to: continuously collect dynamic operation data of the target area of ​​the loading point through multi-source sensors during the operation, extract the current loading boundary features based on the dynamic operation data, and then compare the current loading boundary features with the boundary on which the initial grid map is based to determine whether the loading area has changed. If the change exceeds a set threshold, an incremental map update process is triggered to update the initial grid map.

[0005] As can be seen from the above technical solutions, this application has the following advantages: First, by deploying multi-source sensors and performing deep fusion processing, it can automatically generate high-precision initial point cloud maps and raster maps in the world coordinate system, accurately reflecting the terrain, boundaries, and drivable areas of the loading area, thereby effectively reducing the path planning deviation and collision risk of unmanned mining trucks; Second, through the established real-time update mechanism, it can continuously perceive the dynamic changes in the loading area, and automatically trigger the incremental update process when the terrain changes exceed the threshold. This method only recalculates and updates the changed areas, rather than reconstructing the entire map, which greatly improves the update efficiency, reduces the computational overhead, and ensures the real-time consistency between the map and the actual situation, providing a reliable guarantee for continuous, efficient, and safe production operations in the mining area. Attached Figure Description

[0006] To more clearly illustrate the technical solution of this application, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0007] Figure 1 This is a schematic diagram of a method for generating a grid map of a mining area for unmanned mining trucks, provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the installation location of a multi-source sensor.

[0009] Figure 3 This is a schematic block diagram of a mining area grid map generation system for unmanned mining trucks, provided as an embodiment of the present invention.

[0010] Figure 4 This is a schematic diagram of the structure of a terminal provided in an embodiment of the present invention. Detailed Implementation

[0011] To make the purpose, features, and advantages of this application more apparent and understandable, specific embodiments and accompanying drawings will be used to clearly and completely describe the technical solution protected by this application. Obviously, the embodiments described below are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this application and in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0013] Figure 1 This is a schematic flowchart illustrating a method for generating a raster map of a mining area for unmanned mining trucks, provided in an embodiment of the present invention. Figure 1 The executing entity can be a mining area grid map generation system for unmanned mining trucks. The mining area grid map generation method for unmanned mining trucks provided in this embodiment of the invention is executed by a computer device; correspondingly, the mining area grid map generation system for unmanned mining trucks runs on the computer device. Depending on different needs, the order of the steps in this flowchart can be changed, and some steps can be omitted.

[0014] like Figure 1 As shown, the method includes the following steps.

[0015] S1, by using multi-source sensors deployed on unmanned vehicles in the mining area, static environmental data of the loading area is collected before the equipment operates; the static environmental data includes at least first GNSS data, first IMU data, and first point cloud data generated by lidar scanning.

[0016] S2, the first point cloud data is denoised, then the first GNSS data and the first IMU data are fused, and then a graph-optimized SLAM mapping algorithm is used to generate an initial point cloud map in the world coordinate system.

[0017] S3. Based on the map editing platform, load the initial point cloud map, draw the initial map boundary according to the boundary features of the point cloud map of the mining area loading area, and export the initial raster map.

[0018] S4. Based on the initial grid map, a real-time update mechanism is established. This mechanism is configured to: continuously collect dynamic operation data of the target area of ​​the loading point through multi-source sensors during the operation, extract the current loading boundary features based on the dynamic operation data, and then compare the current loading boundary features with the boundary on which the initial grid map is based to determine whether the loading area has changed. If the change exceeds a set threshold, an incremental map update process is triggered to update the initial grid map.

[0019] Furthermore, as a refinement and extension of the specific implementation of the above embodiments, in order to fully illustrate the specific implementation process of this embodiment, another embodiment of a method for generating a grid map of a mining area for unmanned mining trucks is provided, which includes the following steps.

[0020] S100, static environment data acquisition.

[0021] Before the equipment operates, multi-source sensors deployed on unmanned vehicles in the mining area collect complete static environmental information of the loading area.

[0022] First, sensor units, including lidar, GNSS receiver, IMU (inertial measurement unit), and cameras, are rigidly installed and precisely calibrated at suitable locations on the unmanned mining truck. This embodiment primarily uses data collected by lidar, GNSS receiver, and IMU.

[0023] LiDAR is typically mounted on the roof of the vehicle to obtain an unobstructed, wide field of view, used to acquire high-precision 3D point cloud data of the loading area and its surrounding environment. A GNSS receiver can be mounted on the vehicle's antenna mount to receive satellite signals and determine the vehicle's absolute geographic coordinates. An IMU is usually installed near the vehicle's center of gravity; its physical position relative to the LiDAR and GNSS receivers needs to be precisely measured, and it is used to measure the vehicle's three-axis angular velocity and three-axis acceleration. Figure 2 This is a schematic diagram of the installation location of a multi-source sensor; the right side is the front of the vehicle.

[0024] Before the equipment starts operating, the data acquisition system of the unmanned mining truck is activated. After the system is powered on, each sensor is initialized and self-tested to ensure that all sensors are working properly and in a ready state.

[0025] The unmanned mining truck is controlled to travel at a low speed along a preset path that ensures complete coverage and scanning of the entire loading area. During travel, data from all sensors is simultaneously triggered and recorded.

[0026] The lidar continuously scans the surrounding environment at a preset first frequency, generating a series of first point cloud data frames with synchronized timestamps. These point cloud data accurately record the terrain undulations, the location and shape of fixed facilities in the loading area before the operation.

[0027] The GNSS receiver records the vehicle's first GNSS data at a second frequency higher than that of the lidar. This data includes at least latitude, longitude, elevation, and timestamp information, providing a global positioning reference for the system.

[0028] The IMU also records the vehicle's first IMU data at a high frequency, which includes triaxial angular velocity and triaxial acceleration to capture high-frequency attitude changes of the vehicle during the scan.

[0029] The data acquisition system packages all the raw data collected by the above sensors, including the first point cloud data, the first GNSS data, and the first IMU data, together with their precise timestamps, into a static environment data package, and stores it in the vehicle's solid-state drive or transmits it to the edge computing server through the vehicle network.

[0030] S200 generates the initial point cloud map in the world coordinate system.

[0031] First, the first point cloud data collected in step S100 is denoised. Then, the denoised first point cloud data is fused with the first GNSS data and the first IMU data collected in step S100. Finally, a graph-optimized SLAM mapping algorithm is used to generate an initial point cloud map in the world coordinate system. The specific steps include the following steps.

[0032] S201. A statistical filtering algorithm is used to calculate the average distance and standard deviation between each point and its neighboring points in the first point cloud data, and remove all outliers whose distance exceeds the standard deviation range. Then, a radius filtering algorithm is used to traverse each point in the current point cloud, count the number of its neighboring points within a certain radius, and remove noise points whose number of neighboring points is lower than a preset threshold, thereby obtaining the denoised point cloud.

[0033] The raw first point cloud data acquired by the lidar is denoised to improve data quality. This step involves performing statistical filtering and radius filtering sequentially.

[0034] For statistical filtering, core parameters are set based on the characteristics of point cloud data in the loading area of ​​the mining area. These characteristics include dense ground points and discrete equipment points. The core parameters are the number of neighboring points K and the standard deviation factor α. The first point cloud data collected in step S100 is read. For each point cloud data point, the K-nearest neighbor (KNN) algorithm is used to search for its K nearest neighbors. The Euclidean distance from this point to its K nearest neighbors is calculated, resulting in a distance set D = {d1, d2, ..., d...}. K}; Calculate the mean μ and standard deviation σ of the distance set, and set the outlier detection threshold T = μ + 2.0 × σ. Traverse all point cloud data points. If the mean distance μ of a point is greater than T, mark it as an outlier and remove it. Keep the point cloud that meets the threshold requirement and generate a statistically filtered point cloud.

[0035] For radius filtering, the radius search range r and the minimum neighbor number threshold min_neighbors are set based on the point cloud density of the mining area terrain. After loading the statistically filtered point cloud, for each point cloud data point, a search sphere is constructed with that point as the center and r as the radius. The number of neighbor points N within the sphere is counted. If N < min_neighbors, it is identified as an isolated noise point (such as dust floating in the air or sensor false trigger points) and removed. If N ≥ min_neighbors, the point is retained, and the denoised point cloud is finally generated.

[0036] S202, based on timestamps, aligns the first GNSS data, the first IMU data, and each frame of laser point cloud data in time. Then, it transforms the latitude, longitude, and height coordinates of the GNSS into a local Cartesian coordinate system with the fixed base point of the mining area as the origin through UTM or Gauss-Kriging projection algorithm.

[0037] The system uses a hardware clock or a precise time protocol to assign a unified timestamp to all sensor data. During processing, the first GNSS data, the first IMU data, and each frame of laser point cloud data are precisely aligned based on the timestamps to ensure that sensor observations at the same moment can be correlated and used.

[0038] The latitude, longitude, and altitude coordinates (usually in the WGS84 coordinate system) acquired by the GNSS receiver are transformed into a local Cartesian coordinate system with a fixed reference point in the mining area as the origin, using UTM projection or Gauss-Kriging projection algorithms. This fixed reference point can be a control point A. This step unifies the observations from all sensors into a common coordinate system that is easy to measure and calculate.

[0039] S203, integrate the IMU angular velocity and acceleration between two adjacent laser point cloud moments to obtain the relative pose change between these two moments, which is used as motion prediction; match the current frame point cloud with the previous frame point cloud using the iterative nearest point algorithm or the normal distribution transformation algorithm to calculate the inter-frame pose transformation matrix based on the laser data.

[0040] The purpose of this step is to make a preliminary estimate of the vehicle's continuous pose, including position and attitude, through a tightly coupled approach.

[0041] First, IMU pre-integration is performed, that is, the laser point cloud time t of two adjacent frames is integrated. k and t k+1 The angular velocity and acceleration data from the IMU are integrated. Using the IMU's kinematic equations, the relative pose change of the vehicle during this short time interval is derived, including rotation increment ΔR, velocity increment Δv, and translation increment Δp, as a motion prediction for the system.

[0042] Then, inter-frame matching of the lidar is performed, and the current time t is... k+1 The point cloud and the previous moment t k The point clouds are matched using either the iterative nearest point algorithm or the normal distribution transformation algorithm. This algorithm calculates an optimal rigid body transformation matrix T_lidar through iterative optimization, minimizing the difference between the point clouds of two frames. T_lidar is the inter-frame pose transformation calculated based on the laser data.

[0043] S204 fuses the relative pose change with the inter-frame pose transformation matrix within the prediction-update framework of the extended Kalman filter or the error state Kalman filter, and outputs a preliminary fused pose sequence.

[0044] Specifically, the relative pose change [ΔR, Δv, Δp] obtained from IMU pre-integration and the pose transformation matrix T_lidar obtained from LiDAR matching are jointly input into an extended Kalman filter or an error-state Kalman filter. In the filter's prediction-update framework, the IMU provides high-frequency motion priors (prediction step), while the LiDAR point cloud matching results serve as absolute observations to correct the accumulated error of the IMU prediction (update step). Through this process, a smooth preliminary fused pose sequence {P_0, P_1, ..., P_n} is output.

[0045] S205, the vehicle state at each time step is added as a node to the pose graph. The vehicle state includes position and attitude. Based on the preliminary fused pose sequence, pose relative transformation edges are added between nodes at adjacent time steps. At the time when GNSS observations are available, absolute position constraint edges are added between the node and the absolute coordinates after GNSS conversion, resulting in a pose graph model with a series of nodes, pose relative transformation edges, and absolute position constraint edges.

[0046] To obtain a globally consistent map, backend optimization is required, followed by the construction of a pose graph model.

[0047] First, add nodes, that is, add each pose P_t (including position and orientation) in the preliminary fused pose sequence output by S204 as a node to the graph.

[0048] Then, edges, or constraints, are added, including odometry edges and GNSS absolute position constraint edges. Adding odometry edges refers to adding a relative pose transformation edge between nodes at adjacent time points based on the initial fused pose sequence. This edge represents the observed relative motion between the two nodes, and its constraint comes from the fusion result of S204. Adding GNSS absolute position constraint edges refers to adding an absolute position constraint edge between the corresponding node and the GNSS absolute coordinates (X_gnss, Y_gnss, Z_gnss) transformed by S202 at the time when high-precision GNSS observations are available. This edge pulls the local trajectory towards global absolute coordinates, effectively suppressing accumulated errors.

[0049] S206 calls the graph optimization library and uses the Levenberg-Marquardt algorithm to minimize the sum of errors of all edges in the pose graph and solve for the optimal pose of each node. Using the optimized optimal pose, each frame of point cloud is transformed from its own sensor coordinate system to the world coordinate system. All point clouds transformed to the world coordinate system are superimposed and downsampled using a voxelized mesh filter, and finally the initial point cloud map in the world coordinate system is output.

[0050] The first step involves graph optimization, which utilizes an open-source graph optimization library and employs the Leveenberg-Marquardt algorithm to solve the nonlinear least squares problem. This algorithm iteratively adjusts the poses of all nodes in the graph to minimize the overall error of all odometer edges and GNSS constraint edges, thus obtaining the optimal vehicle pose P_optimized(t) at each time step. In essence, solving the nonlinear least squares problem involves iteratively fine-tuning the pose of each node using optimization algorithms such as the Leveenberg-Marquardt algorithm, ultimately finding the optimal pose P_optimized(t) that minimizes the sum of the squared errors of all odometer and GNSS constraints.

[0051] The second step involves point cloud stitching and map generation. Using the optimized optimal pose P_optimized(t), each frame of the laser point cloud is transformed from its own sensor coordinate system to the world coordinate system. All transformed point cloud frames are then superimposed to form a dense point cloud map.

[0052] Specifically, for each time t, the corresponding optimal pose P_optimized(t) is processed. This pose is usually represented by a 4x4 homogeneous transformation matrix T_world_sensor(t), which defines the transformation from the sensor coordinate system to the world coordinate system, including rotation and translation.

[0053] Then, for the original single-frame point cloud Cloud_sensor(t) in the sensor coordinate system at time t, each three-dimensional point P_sensor=(x,y,z,1)^T is transformed to the world coordinate system through matrix multiplication:

[0054] After performing this operation on all points, we obtain the representation of the point cloud in the world coordinate system, Cloud_world(t), and then obtain the transformed point cloud frames {Cloud_world(t0), Cloud_world(t1), ..., Cloud_world(t2)} at all time points. n )}.

[0055] An empty point cloud data structure is created in memory as a global point cloud map, GlobalMap. Each frame of Cloud_world(t) is traversed in chronological or arbitrary order, and all points in it are added to GlobalMap. This allows point clouds collected at different times and from different perspectives to be uniformly registered in the same world coordinate system based on their optimized poses, thus achieving data fusion.

[0056] The third step involves downsampling, using a voxelized mesh filter to downsample the stitched point cloud. This filter divides the 3D space into a uniform voxel mesh and replaces all points within each voxel with its centroid (or center point), reducing the amount of point cloud data, eliminating overlapping points, and ultimately outputting a lightweight initial point cloud map in world coordinates.

[0057] Specifically, the GlobalMap formed after the second step of overlay may have significant overlap, therefore downsampling is performed. First, the entire 3D space of the GlobalMap is divided into a uniform voxel grid with sides of length `voxel_size`. For each voxel, all points falling within that voxel are identified, and the 3D centroid of these points is calculated. This centroid is then used to replace all the original points within that voxel. Finally, a final initial point cloud map with uniform point density and a significantly reduced number of points is obtained.

[0058] S300, Export the initial raster map.

[0059] This step is based on a map editing platform. It loads the initial point cloud map from step S200, draws the initial map boundary based on the boundary features of the point cloud map of the mining area loading area, and exports the initial raster map. Specifically, it includes the following steps.

[0060] S301, traverse the X and Y coordinates of all points in the initial point cloud map, find their minimum and maximum values, and obtain the span of the point cloud on the X and Y axes X_min, X_max, Y_min, Y_max.

[0061] Specifically, the initial point cloud map in the world coordinate system is read, the X and Y coordinates of all points are traversed, and the minimum and maximum values ​​of the X coordinates, X_min and X_max, and the minimum and maximum values ​​of the Y coordinates, Y_min and Y_max, are calculated and found. Thus, the geographical span of the point cloud on the horizontal plane is determined as [X_min, X_max] and [Y_min, Y_max].

[0062] S302, Calculate the actual physical extent of the raster map, including physical length, based on the preset expansion scale γ. Physical width .

[0063] To avoid the map boundary being too close to the point cloud data, the system expands the physical range according to a preset expansion ratio γ, which can be 1.1.

[0064] S303, select the raster resolution R from the preset parameters based on the terrain features of the loading area.

[0065] Based on the actual terrain features of the loading area, select the resolution R of the raster map from the preset parameter library. For loading areas with a large area and relatively flat terrain, choose a lower resolution, such as R=1.0 meters; for areas with complex terrain or requiring detailed representation, choose a higher resolution, such as R=0.5 meters.

[0066] S304. Calculate the size of the two-dimensional matrix of the raster map based on the physical range and resolution, including the number of rows Rows = ceil(W / R) and the number of columns Cols = ceil(L / R), where ceil represents rounding up; then obtain the raster map parameters, including the origin coordinates (X_origin, Y_origin), resolution R, and dimensions (Rows, Cols).

[0067] Define the map origin coordinates (X_origin, Y_origin), which is usually set to the bottom left vertex of the expanded map area, i.e.: X_origin=X_min-(L-(X_max-X_min)) / 2 Y_origin=Y_min-(W-(Y_max-Y_min)) / 2 Finally, we obtain the complete set of raster map parameters: origin coordinates (X_origin, Y_origin), resolution R, and dimensions (Rows, Cols).

[0068] S305 converts the (X,Y) coordinates of each point in the point cloud into the row and column indices of the raster matrix, represented as col_index=floor((X-X_origin) / R), row_index=floor((Y-Y_origin) / R), where floor represents rounding down. Then, each point cloud is associated with its corresponding raster cell.

[0069] It should be noted that the elevation value of a point is associated with the raster cell determined by its row and column indices.

[0070] S306. For each raster cell, count the elevation values ​​of all the point cloud data associated with it, and calculate the average elevation and standard deviation of the raster.

[0071] S307. Each grid is classified based on preset rules, which include rule 1 "if the standard deviation of elevation within the grid is lower than a first preset threshold, it is determined to be flat terrain" and rule 2 "if the average elevation within the grid differs from the elevation of surrounding grids from a second preset threshold, it is determined to be without elevation abrupt change". Grids that satisfy both rule 1 and rule 2 are classified as drivable areas and marked as 0 in the matrix; otherwise, they are classified as non-drivable areas and marked as 1, thus obtaining a preliminary grid matrix containing 0 and 1 for two-dimensional drivability.

[0072] It is understandable that rule 1 is a flatness rule, and rule 2 is a height abrupt change rule.

[0073] S308: Perform morphological opening operation on the initial raster matrix to obtain the smoothed final raster matrix; perform an edge detection algorithm on the processed raster matrix to find all adjacent raster boundaries of drivable and non-drivable areas, and then convert the extracted pixel-level boundaries into vector polygon boundaries composed of a series of vertex coordinates, and record the vector polygon boundaries as the original boundary polygons.

[0074] This step involves post-processing the initial raster matrix to improve map quality and extract vector boundaries.

[0075] Morphological optimization involves performing a morphological opening operation (erosion followed by dilation) on Grid_initial to eliminate small, isolated noise regions and smooth the boundaries of drivable regions, resulting in the optimized final grid matrix Grid_final.

[0076] Regarding boundary extraction and vectorization, an edge detection algorithm is performed on Grid_final to identify all grids located at the boundary between "drivable area (0)" and "non-drivable area (1)", obtaining pixel-level boundaries. These pixel boundary points are then connected and simplified, and converted into a vector polygon boundary composed of a series of ordered vertex coordinates using a polygon approximation algorithm. This vector polygon is defined as the original boundary polygon Polygon_original, serving as the benchmark for subsequent map updates and comparisons.

[0077] S309 serializes the final raster matrix and raster map parameters into a target file format and exports the initial raster map file.

[0078] The final raster map data is serialized and output. Specifically, this involves writing the Grid_final matrix, map origin (X_origin, Y_origin), resolution R, and other parameters into a standardized map file, thus completing the export of the initial raster map file. For example, it can be in PNG image format, where a pixel value of 0 represents passable and 255 represents impassable; or in a custom binary format containing metadata.

[0079] S400, grid map updated in real time.

[0080] This embodiment establishes a real-time update mechanism based on an initial grid map. This mechanism is configured to: continuously collect dynamic operation data of the target area of ​​the loading point through multi-source sensors during the operation, extract the current loading boundary features based on the dynamic operation data, and then compare the current loading boundary features with the boundary on which the initial grid map is based to determine whether the loading area has changed. If the change exceeds a set threshold, an incremental map update process is triggered to update the initial grid map. Specifically, it includes the following steps.

[0081] S401 collects dynamic operation data.

[0082] During operation, dynamic operational data of the target area at the loading point is continuously collected through multi-source sensors. After the initial grid map is built and put into use, when the unmanned mining truck receives a scheduling instruction and enters the loading area to prepare for the loading task, the system automatically starts the dynamic data acquisition mode. Unlike the comprehensive and static scanning performed when building the initial map, the data collection at this stage needs to be high-frequency, continuous, and target-oriented.

[0083] Throughout the entire process of the mining truck queuing, positioning, or cooperating with excavating equipment near the loading point, the onboard multi-source sensor system remains continuously operational, regardless of whether the mining truck is in motion. Dynamic operational data is collected synchronously by the following sensors.

[0084] The second point cloud data is generated by lidar scanning: the lidar continuously scans the environment around the loading point at a set operating frequency. This second point cloud data records in real time the three-dimensional geometric information of the working face, the mining truck itself, the excavator boom, and the new terrain created by the excavation.

[0085] Second GNSS data providing absolute vehicle location information: The GNSS receiver continuously records the second GNSS data of the mining truck at a high frequency.

[0086] Second IMU data providing vehicle attitude information: The IMU also records the mining truck's second IMU data at a high frequency (e.g., 100 Hz). In operational scenarios where mining trucks may frequently start and stop and fine-tune their positions, IMU data is crucial for maintaining positioning continuity when GNSS signals are briefly blocked (such as when near large equipment), ensuring the accuracy of point cloud data stitching.

[0087] S402, extract loading boundary features.

[0088] Extracting the current loading boundary features based on the dynamic operation data collected in step S401 includes the following steps.

[0089] S4021 utilizes the synchronously acquired second GNSS data and second IMU data to register multiple frames of second point cloud data to the world coordinate system through coordinate transformation, forming a current dynamic point cloud snapshot of the target area of ​​the loading point.

[0090] First, based on a unified timestamp, the second point cloud data of each frame is aligned with the corresponding second GNSS position data and second IMU attitude data.

[0091] IMU data was used to correct point cloud distortion caused by vehicle movement during a single-frame scan of the lidar. Then, combining the absolute position provided by GNSS and the attitude provided by the IMU, a coordinate transformation matrix was used to accurately register multiple consecutive frames of the second point cloud data to the initial point cloud in the world coordinate system. Figure 1 In the global coordinate system.

[0092] Specifically, since a single-frame scan by the lidar requires a certain amount of time, the movement of the mining truck itself during this period causes distortion within the point cloud of that frame, which is known as motion distortion. To eliminate this effect, an IMU-based correction method is employed. During the period from the start time t_start to the end time t_end of the lidar's single-frame scan, the angular velocity and acceleration measured by the IMU are recorded synchronously at a high frequency. By integrating the angular velocity and acceleration data from the IMU and combining it with the initial state, the relative pose change of the lidar during each minute time interval from t_start to t_end is calculated, thereby reconstructing the continuous motion trajectory of the radar sensor during that frame scan.

[0093] For each point P_i in the point cloud of this frame, its corresponding precise scan time t_i is known. Based on the reconstructed motion trajectory, the radar coordinate system transformation T_(start→i) from the scan center time to time t_i is calculated. Then, this transformation is used to uniformly correct the coordinates of point P_i to the radar coordinate system at the scan start time t_start, with the formula: P_i_corrected=T_(start→i)*P_i. After performing this operation on all points, a geometrically consistent point cloud frame with motion distortion eliminated within the current frame is obtained.

[0094] By overlaying all registered point cloud frames from a short past period, a current dynamic point cloud snapshot representing the latest terrain state of the target area is formed. This snapshot integrates multi-view observations and can more completely reflect the current boundary morphology.

[0095] S4022, with the loading point of the current job as the center, extracts the point cloud block to be processed from the dynamic point cloud snapshot according to a preset fixed size.

[0096] To improve processing efficiency and focus on the core changing areas, a pre-defined two-dimensional bounding box of fixed size is defined, using the loading point of the current operation determined by the GNSS data in the preceding steps as the geometric center. The system then crops all point cloud data within this bounding box from the current dynamic point cloud snapshot, forming a point cloud block to be processed. This effectively filters out distant background point clouds that are irrelevant to the current loading operation.

[0097] S4023, the point cloud block to be processed is input into the pre-trained PointNet++ model. The model uses the farthest point sampling algorithm for downsampling, selects key points, and uses the ball query algorithm or K-nearest neighbor algorithm to group the local neighborhood of each key point. Then, the local geometric features within each group are extracted and aggregated layer by layer through a multilayer perceptron. Finally, the probability of each point in the point cloud belonging to the loading boundary is output. According to the probability output by the model, points with probability values ​​higher than a set threshold are classified into candidate boundary point sets.

[0098] The point cloud blocks to be processed are input into a PointNet++ deep learning model pre-trained with a large amount of mining scene data for point-by-point boundary semantic segmentation. The model performs the following operations in sequence.

[0099] First, feature encoding is performed. A set of key points representing the overall geometric structure is iteratively selected from a massive point cloud using a farthest-point sampling algorithm. For each key point, a ball query algorithm is used to group its neighboring points within its spatial neighborhood. Subsequently, a multilayer perceptron is used to learn features from the point set within this group, aggregating local geometric information such as planarity and curvature layer by layer.

[0100] Then, feature decoding and classification are performed. Through upsampling and feature propagation, the learned high-level semantic features are passed back to each point in the original point cloud. Finally, the model outputs a probability value between 0 and 1 for each point in the point cloud block to be processed, indicating that it belongs to the loading boundary.

[0101] Set a probability threshold, filter out all points with probability values ​​higher than this threshold, and classify them into a candidate boundary point set. This set contains all points in the current snapshot that could potentially constitute the boundary of the loading area.

[0102] The PointNet++ model takes as input a point cloud patch extracted from a dynamic point cloud snapshot. This patch consists of N 3D points, represented as an N×3 matrix, where each row (x, y, z) represents the 3D coordinates of a point in the world coordinate system. N is a variable value that depends on the size of the extracted region and the point cloud density.

[0103] The PointNet++ model employs a hierarchical encoder-decoder structure, which can effectively learn multi-scale geometric features from disordered and sparse 3D point clouds.

[0104] The encoder consists of multiple cascaded ensemble abstraction layers for feature extraction and downsampling. At each layer, a farthest-point sampling algorithm selects M keypoints from the input point set as new, sparser point set centers. For each keypoint, a spherical neighborhood is constructed with that keypoint as the center and a preset radius *r*, grouping all points within that neighborhood into a single group. Points within each group (containing their coordinates and features from the previous layer) are fed into a miniature PointNet network composed of multilayer perceptrons for local feature extraction and aggregation. The output of the MLP is a new feature vector for that keypoint, representing abstract information about its local region. By stacking multiple ensemble abstraction layers, the model's receptive field increases layer by layer, enabling it to capture multi-scale features from local details to global semantics.

[0105] The decoder consists of multiple feature propagation layers, used to upsample the high-level features learned by the encoder and propagate them back to the original point set. Upsampling is achieved by interpolating the features of sparse keypoints onto a denser point set using methods such as inverse distance weighted interpolation. The interpolated features are then skip-connected with corresponding local features of the same resolution in the encoder's intermediate layers to recover details lost during encoding. The concatenated features are then refined using an MLP to output the final feature vector for each point.

[0106] The decoder's output is finally connected to a weighted MLP, followed by a Softmax activation function to generate a classification score for each point in the original input point cloud. The model's output is an N×2 matrix, where each row corresponds to a point in the input point cloud and contains two values, representing the probability that the point belongs to the "non-boundary" and "boundary" categories, respectively.

[0107] The PointNet++ model is implemented through the following supervised learning process: a) Training dataset construction: Collect a large amount of point cloud data covering different terrains and different operational stages of the mining loading area. Experts will perform fine annotation on each frame of point cloud data, labeling each point as a "boundary point" or "non-boundary point" to form ground truth data; b) Loss Function: During training, the cross-entropy loss function is used to quantify the difference between the probability distribution predicted by the model and the true label. For the prediction of each point in the point cloud, its loss is calculated, and the average or weighted average of the losses of all points is taken as the total loss of the model; c) Optimizer: Adaptive optimization algorithms, such as the Adam optimizer, are used to calculate the gradient of the loss function with respect to all trainable parameters of the model through backpropagation and iteratively update the parameters to minimize the total loss; d) Training strategy: During training, learning rate decay and data augmentation strategies are typically used to improve the model's generalization ability and robustness. Training continues until the model's performance on independent validation sets no longer shows significant improvement, thus obtaining the final pre-trained weights.

[0108] S4024: Perform the DBSCAN clustering algorithm on the candidate boundary point set to cluster the points according to their spatial density to obtain a pure boundary point set. Then, organize the pure boundary point set in spatial order and use the B-spline interpolation algorithm to perform curve fitting to generate a two-dimensional vector boundary curve. Perform morphological closing operation on the two-dimensional vector boundary curve to output the two-dimensional vector boundary curve, which is the currently loaded boundary feature.

[0109] The coarser set of candidate boundary points generated in the previous step is refined to generate smooth, continuous vector boundaries.

[0110] Specifically, the DBSCAN clustering algorithm is applied to the candidate boundary point set. This algorithm clusters points based on their spatial density, effectively identifying and removing outliers that do not belong to any major dense region, while grouping spatially connected boundary points into the same cluster, thus obtaining several pure boundary point sets.

[0111] The largest or most important set of pure boundary points is selected, and these points are sorted according to their spatial distribution to form an ordered point sequence. B-spline interpolation is used to curve fit this ordered point sequence, generating a smooth, continuous two-dimensional vector boundary curve. This two-dimensional vector boundary curve is converted into a rasterized profile, and a morphological closing operation (dilation followed by erosion) is performed on it. This operation fills any tiny breaks or holes that may exist in the vector boundary and smooths the edges of the boundary profile, ultimately outputting a high-quality, geometrically comparable current-load boundary feature.

[0112] S403, Loading area change judgment.

[0113] The current loading boundary features are compared with the boundary on which the initial raster map is based to determine whether the loading area has changed. This includes the following steps.

[0114] S4031, record the current loading boundary feature as the current loading boundary polygon, and unify the original boundary polygon and the current loading boundary polygon into the same world coordinate system.

[0115] The system reads the original boundary polygon, Polygon_original, generated during the initial map building phase (S308), from the storage unit. This polygon defines the initial drivable area of ​​the loading zone. Simultaneously, it receives the current loading boundary features output from the dynamic feature extraction process (S4024) and defines them as the current loading boundary polygon, Polygon_current, which reflects the latest boundary state of the loading zone. To ensure the accuracy of the geometric calculations, the system first confirms that both polygons are based on the same world coordinate system. If a coordinate system difference exists, the system performs the necessary coordinate transformation to unify Polygon_current to the coordinate system of Polygon_original.

[0116] S4032, calculates the union polygon between the original boundary polygon and the currently loaded boundary polygon.

[0117] The geometry calculation engine is invoked to perform a geometric union Boolean operation on Polygon_original and Polygon_current. This operation generates a new polygon, denoted as the union polygon Polygon_union, whose area is the sum of all spaces occupied by Polygon_original and Polygon_current. That is: Polygon_union = Polygon_original ∪ Polygon_current. This polygon represents the maximum extent of the loading region from the initial state to the current state.

[0118] S4033 uses geometric difference operations to subtract the original boundary polygon from the union polygon to obtain the polygon of the changed region.

[0119] By using geometric difference Boolean operations, newly added changed areas are accurately identified. The system calculates the difference between the union polygon Polygon_union and the original boundary polygon Polygon_original to obtain the changed area polygon Polygon_change. That is: Polygon_change = Polygon_union - Polygon_original. The geographical area covered by this Polygon_change polygon is the area that was originally outside the drivable area and has been newly generated since the initial map was built due to excavation operations and other reasons.

[0120] S4034 Calculate the area of ​​the polygon of the changed region, and compare the area with a preset area change threshold. If the area is greater than the area change threshold, it is determined that the loading area has changed; otherwise, it is determined that the loading area has not changed.

[0121] The identified areas of change are quantitatively evaluated and decisions are made. First, a geometry library function is called to calculate the area (Area_change) of the polygon (Polygon_change) in the changed area. The calculated Area_change is compared with a preset area change threshold (Threshold_area), which is an empirical value set based on operational accuracy and update sensitivity requirements. If Area_change > Threshold_area, the loading area is determined to have changed; otherwise, it is determined that the loading area has not changed significantly. Here, "change" can be understood as whether a significant change has occurred. The judgment result serves as a system control signal. If the judgment is "yes," the subsequent incremental map update process is automatically triggered; if the judgment is "no," the system continues to collect data and enters the next monitoring cycle.

[0122] S404, map updated.

[0123] If the changes exceed the set threshold, an incremental map update process is triggered to update the initial raster map, which includes the following steps.

[0124] S4041, take the union of the original boundary polygon and the currently loaded boundary polygon to obtain the final target boundary polygon.

[0125] The system reads the original boundary polygon `Polygon_original` and the current loading boundary polygon `Polygon_current` obtained in the change judgment step. A final target boundary polygon (`Polygon_target`) is generated by calculating their geometric union. This polygon defines the complete and latest boundary of the drivable area of ​​the loading zone after this update. That is: `Polygon_target = Polygon_original ∪ Polygon_current`.

[0126] S4042, based on the geographical range covered by the final target boundary polygon, cut out all point cloud data within that range from the fused point cloud data to obtain the updated region point cloud block.

[0127] Based on the geographic area defined by the final target boundary polygon Polygon_target, spatial queries and clipping are performed from the fused point cloud database. This database contains initial static point clouds and subsequently acquired dynamic point clouds registered to the world coordinate system. The system extracts all point cloud data completely within the Polygon_target area to form updated region point cloud blocks, ensuring that the data used to reconstruct the map of this region is the most comprehensive and up-to-date.

[0128] S4043, for the grid range covered by the final target boundary polygon, use the updated area point cloud blocks to perform the same point cloud projection, grid cell statistical analysis and accessibility determination rules as the initial grid map construction, and generate a new local grid matrix.

[0129] Perform a small-scale reconstruction process that is completely consistent with the initial raster map construction algorithm (see steps S305 to S307 in claim 3) only within the raster area covered by Polygon_target: a) Point cloud projection and association: Each point in the updated area point cloud block is projected onto the corresponding raster cell according to the map origin (X_origin, Y_origin) and resolution R; b) Terrain statistical analysis: For each raster cell within the Polygon_target range, recalculate the elevation values ​​of all point clouds within it, and calculate the new elevation mean and standard deviation; c) Accessibility reclassification: Based on the same preset rules, accessibility is determined for each grid cell, generating a brand new local grid matrix Grid_local that is only for the target area.

[0130] S4044: The original data region in the initial grid matrix corresponding to the final target boundary polygon range is covered with the corresponding data in the new local grid matrix to obtain the updated global grid matrix.

[0131] The new local raster matrix Grid_local is merged with the initial raster matrix Grid_global stored in memory or a file. First, the row and column index range corresponding to the boundary of Polygon_target in the global raster matrix is ​​calculated. Then, the original raster data in the above index range in Grid_global is directly replaced with the corresponding new raster data in Grid_local.

[0132] S4045 serializes the global raster matrix and replaces the original initial raster map file to obtain the final updated raster map file.

[0133] The updated global raster matrix Grid_updated, along with its map parameters, is serialized into the specified target file format. Subsequently, the system atomically replaces the original initial raster map file in the file system with this newly generated map file. This ensures that downstream modules such as path planning can obtain the latest version the next time they read the map, thus achieving online incremental updates of the mining area raster map and providing real-time and accurate environmental information for unmanned mining trucks.

[0134] The foregoing has described in detail an embodiment of a method for generating a grid map of a mining area for unmanned mining trucks. Based on the above embodiment of the method for generating a grid map of a mining area for unmanned mining trucks, this embodiment of the invention also provides a corresponding system for generating a grid map of a mining area for unmanned mining trucks.

[0135] Figure 3 This is a schematic block diagram of a mining area grid map generation system for unmanned mining trucks provided in an embodiment of the present invention. In this embodiment, the mining area grid map generation system 300 for unmanned mining trucks can be divided into multiple functional modules according to the functions it performs, such as... Figure 2 As shown. The module referred to in this invention is a series of computer program segments that can be executed by at least one processor and perform a fixed function, and which are stored in memory.

[0136] The static environment data acquisition module 310 is used to collect static environment data of the loading area before equipment operation by using multi-source sensors deployed on unmanned vehicles in the mining area; the static environment data includes at least first GNSS data, first IMU data, and first point cloud data generated by lidar scanning.

[0137] The initial point cloud map generation module 320 is used to denoise the first point cloud data, then fuse the first GNSS data and the first IMU data, and then use a graph-optimized SLAM mapping algorithm to generate an initial point cloud map in the world coordinate system.

[0138] The initial raster map generation module 330 is used to load the initial point cloud map based on the map editing platform, draw the initial map boundary according to the boundary features of the point cloud map of the mining area loading area, and export the initial raster map.

[0139] The grid map update module 340 is used to establish a real-time update mechanism based on the initial grid map. This mechanism is configured to: continuously collect dynamic operation data of the target area of ​​the loading point through multi-source sensors during the operation, extract the current loading boundary features based on the dynamic operation data, and then compare the current loading boundary features with the boundary on which the initial grid map is based to determine whether the loading area has changed. If the change exceeds a set threshold, an incremental map update process is triggered to update the initial grid map.

[0140] The mining area grid map generation system for unmanned mining trucks in this embodiment is used to implement the aforementioned mining area grid map generation method for unmanned mining trucks. Therefore, the specific implementation of this system can be found in the embodiment section of the mining area grid map generation method for unmanned mining trucks above. Thus, the specific implementation can be referred to the description of the corresponding embodiments, and will not be elaborated here.

[0141] Furthermore, since the mining area grid map generation system for unmanned mining trucks in this embodiment is used to implement the aforementioned mining area grid map generation method for unmanned mining trucks, its function corresponds to the function of the above method, and will not be repeated here.

[0142] Figure 4 This is a schematic diagram of a terminal 400 provided in an embodiment of the present invention, including: a processor 410, a memory 420, and a communication unit 430. The processor 410 is used to implement the process steps of the above-described embodiment of the method for generating a mining area grid map for unmanned mining trucks when implementing the mining area grid map generation program for unmanned mining trucks stored in the memory 420.

[0143] This invention also provides a computer storage medium, which may be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. The computer storage medium stores a mining area grid map generation program for unmanned mining trucks. When the processor executes the mining area grid map generation program for unmanned mining trucks, it implements the process steps of the above-described embodiment of the mining area grid map generation method for unmanned mining trucks.

[0144] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for generating a raster map of a mining area for unmanned mining trucks, characterized in that, Includes the following steps: By deploying multi-source sensors on unmanned vehicles in the mining area, static environmental data of the loading area is collected before equipment operation; the static environmental data includes first GNSS data, first IMU data, and first point cloud data generated by lidar scanning; The first point cloud data is denoised, then the first GNSS data and the first IMU data are fused together, and then a graph-optimized SLAM mapping algorithm is used to generate an initial point cloud map in the world coordinate system. Based on the map editing platform, the initial point cloud map is loaded, the initial map boundary is drawn according to the boundary features of the point cloud map of the mining area loading area, and the initial raster map is exported. Based on the initial grid map, a real-time update mechanism is established. This mechanism is configured to: continuously collect dynamic operation data of the target area of ​​the loading point through multi-source sensors during the operation, extract the current loading boundary features based on the dynamic operation data, and then compare the current loading boundary features with the boundary on which the initial grid map is based to determine whether the loading area has changed. If the change exceeds a set threshold, an incremental map update process is triggered to update the initial grid map.

2. The method for generating a grid map of a mining area for unmanned mining trucks according to claim 1, characterized in that, The first point cloud data is denoised and GNSS and IMU data from the static environment data are fused together. Then, a graph-optimized SLAM mapping algorithm is used to generate an initial point cloud map in the world coordinate system, specifically including: A statistical filtering algorithm is used to calculate the average distance and standard deviation between each point and its neighbors in the first point cloud data, and remove all outliers whose distance exceeds the standard deviation range. Then, a radius filtering algorithm is used to traverse each point in the current point cloud, count the number of its neighbors within a certain radius, and remove noise points whose number of neighbors is lower than a preset threshold, thereby obtaining a denoised point cloud. Based on the timestamp, the first GNSS data, the first IMU data and each frame of laser point cloud data are time-aligned. Then, the latitude, longitude and height coordinates of the GNSS are transformed into a local plane rectangular coordinate system with the fixed base point of the mining area as the origin through UTM or Gauss-Kriging projection algorithm. Integrate the IMU angular velocity and acceleration between two adjacent laser point cloud moments to obtain the relative pose change between these two moments, which is used for motion prediction; match the current frame point cloud with the previous frame point cloud using the iterative nearest point algorithm or the normal distribution transformation algorithm to calculate the inter-frame pose transformation matrix based on the laser data. The relative pose change and the inter-frame pose transformation matrix are fused within the prediction-update framework of the extended Kalman filter or the error state Kalman filter to output a preliminary fused pose sequence. The vehicle state at each time step is added as a node to the pose graph. The vehicle state includes position and attitude. Based on the preliminary fused pose sequence, relative pose transformation edges are added between nodes at adjacent time steps. At the time when GNSS observations are available, absolute position constraint edges are added between the node and the absolute coordinates after GNSS transformation. This results in a pose graph model consisting of a series of nodes, relative pose transformation edges, and absolute position constraint edges. The graph optimization library is called, and the Levenberg-Marquardt algorithm is used to minimize the sum of errors of all edges in the pose graph to solve for the optimal pose of each node. Using the optimized optimal pose, the point cloud of each frame is transformed from its own sensor coordinate system to the world coordinate system. All point clouds transformed to the world coordinate system are superimposed and downsampled using a voxelized mesh filter to finally output the initial point cloud map in the world coordinate system.

3. The method for generating a grid map of a mining area for unmanned mining trucks according to claim 1, characterized in that, Load the initial point cloud map, draw the initial map boundary based on the boundary features of the point cloud map of the mining area loading zone, and export the initial raster map, specifically including: Iterate through the X and Y coordinates of all points in the initial point cloud map, find their minimum and maximum values, and obtain the span of the point cloud on the X and Y axes: X_min, X_max, Y_min, Y_max. Based on the preset scaling factor γ, calculate the actual physical extent of the raster map, including its physical length. Physical width ; Based on the terrain features of the loading area, select the raster resolution R from the preset parameters; Based on the physical extent and resolution, calculate the size of the two-dimensional matrix of the raster map, including the number of rows Rows=ceil(W / R) and the number of columns Cols=ceil(L / R), where ceil represents rounding up; then obtain the raster map parameters, including the origin coordinates (X_origin, Y_origin), resolution R, and dimensions (Rows, Cols). The (X,Y) coordinates of each point in the point cloud are converted into the row and column indices of the raster matrix, represented as col_index=floor((X-X_origin) / R), row_index=floor((Y-Y_origin) / R), where floor represents rounding down. Then each point cloud is associated with its corresponding raster cell. For each raster cell, the elevation values ​​of all its associated point cloud data are counted, and the average and standard deviation of the elevation of the raster are calculated. Each grid is classified based on preset rules, including rule 1 "if the standard deviation of elevation within a grid is lower than a first preset threshold, it is determined to be flat terrain" and rule 2 "if the average elevation within a grid differs from the elevation of surrounding grids from a second preset threshold, it is determined to be without elevation abrupt change". Grids that satisfy both rule 1 and rule 2 are classified as drivable areas and marked as 0 in the matrix; otherwise, they are classified as non-drivable areas and marked as 1, thus obtaining a preliminary grid matrix containing 0 and 1 for two-dimensional drivability. Morphological opening is performed on the initial raster matrix to obtain the smoothed final raster matrix. An edge detection algorithm is then performed on the processed raster matrix to find all adjacent raster boundaries between drivable and non-drivable areas. The extracted pixel-level boundaries are then converted into vector polygon boundaries composed of a series of vertex coordinates, and these vector polygon boundaries are recorded as the original boundary polygons. The final raster matrix and raster map parameters are serialized into a target file format, and the initial raster map file is exported.

4. The method for generating a grid map of a mining area for unmanned mining trucks according to claim 3, characterized in that, The dynamic operation data includes: second point cloud data generated by LiDAR scanning at the operation frequency, second GNSS data providing absolute position information of the vehicle, and second IMU data providing attitude information of the vehicle.

5. The method for generating a grid map of a mining area for unmanned mining trucks according to claim 4, characterized in that, Extract the current loading boundary features based on dynamic operation data: Using the synchronously acquired second GNSS data and second IMU data, multiple frames of second point cloud data are registered to the world coordinate system through coordinate transformation to form a current dynamic point cloud snapshot of the target area of ​​the loading point; Centered on the loading point of the current operation, a point cloud block to be processed is extracted from the dynamic point cloud snapshot according to a preset fixed size; The point cloud block to be processed is input into the pre-trained PointNet++ model. The model uses the farthest point sampling algorithm for downsampling, selects key points, and uses the ball query algorithm or K-nearest neighbor algorithm to group the local neighborhood of each key point. Then, the local geometric features within each group are extracted and aggregated layer by layer through a multilayer perceptron. Finally, the probability of each point in the point cloud belonging to the loading boundary is output. According to the probability output by the model, points with probability values ​​higher than a set threshold are classified into candidate boundary point sets. The DBSCAN clustering algorithm is applied to the candidate boundary point set to cluster points based on their spatial density, resulting in a pure boundary point set. The pure boundary point set is then organized in spatial order, and a B-spline interpolation algorithm is used for curve fitting to generate a two-dimensional vector boundary curve. Morphological closing operations are then performed on this two-dimensional vector boundary curve to output the two-dimensional vector boundary curve, which is the currently loaded boundary feature.

6. The method for generating a grid map of a mining area for unmanned mining trucks according to claim 5, characterized in that, The current loading boundary features are compared with the boundary upon which the initial raster map was based to determine whether the loading area has changed. This includes: Record the current loading boundary feature as the current loading boundary polygon, and unify the original boundary polygon and the current loading boundary polygon into the same world coordinate system; Calculate the union polygon between the original boundary polygon and the currently loaded boundary polygon; By subtracting the original boundary polygon from the union polygon using the geometric difference operation, the polygon of the changed region is obtained. Calculate the area of ​​the polygon in the changed region and compare it with a preset area change threshold. If the area is greater than the area change threshold, it is determined that the loading area has changed; otherwise, it is determined that the loading area has not changed.

7. The method for generating a grid map of a mining area for unmanned mining trucks according to claim 6, characterized in that, Trigger the incremental map update process to update the initial raster map, specifically including: The final target boundary polygon is obtained by taking the union of the original boundary polygon and the currently loaded boundary polygon. Based on the geographical range covered by the final target boundary polygon, all point cloud data within that range are cropped from the fused point cloud data to obtain the updated region point cloud block; For the grid range covered by the final target boundary polygon, the same point cloud projection, grid cell statistical analysis and accessibility determination rules as those used in the initial grid map construction are performed using updated regional point cloud blocks to generate a new local grid matrix. The original data region in the initial grid matrix that corresponds to the polygon range of the final target boundary is covered with the corresponding data in the new local grid matrix to obtain the updated global grid matrix. The global raster matrix is ​​serialized and replaced with the original initial raster map file to obtain the final updated raster map file.

8. A system for generating a raster map of a mining area for unmanned mining trucks, characterized in that, include: The static environment data acquisition module is used to collect static environment data of the loading area before equipment operation by using multi-source sensors deployed on unmanned vehicles in the mining area. The static environment data includes at least the first GNSS data, the first IMU data, and the first point cloud data generated by lidar scanning; The initial point cloud map generation module is used to denoise the first point cloud data, then fuse the first GNSS data and the first IMU data, and then use a graph-optimized SLAM mapping algorithm to generate an initial point cloud map in the world coordinate system. The initial raster map generation module is used to load the initial point cloud map based on the map editing platform, draw the initial map boundary according to the boundary features of the point cloud map of the mining area loading area, and export the initial raster map. The grid map update module is used to establish a real-time update mechanism based on the initial grid map. This mechanism is configured to: continuously collect dynamic operation data of the target area of ​​the loading point through multi-source sensors during the operation, extract the current loading boundary features based on the dynamic operation data, and then compare the current loading boundary features with the boundary on which the initial grid map is based to determine whether the loading area has changed. If the change exceeds a set threshold, an incremental map update process is triggered to update the initial grid map.

9. A terminal, characterized in that, include: Memory for storing the mining area grid map generation program for unmanned mining trucks; A processor is configured to implement the steps of the method for generating a mining area grid map for an unmanned mining truck as described in any one of claims 1 to 7 when executing the mining area grid map generation program for unmanned mining trucks.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a mining area grid map generation program for unmanned mining trucks. When the mining area grid map generation program for unmanned mining trucks is executed by a processor, it implements the steps of the mining area grid map generation method for unmanned mining trucks as described in any one of claims 1 to 7.

Images