Real-time modeling method and system suitable for loading and unloading area of unmanned loader

Abstract

The invention discloses a real-time modeling method and system suitable for a loading and unloading area of an unmanned loader, and the method comprises the steps: starting cooperative scanning based on a laser radar, a binocular camera and a millimeter wave radar carried by the unmanned loader, and obtaining the three-dimensional point cloud data of the loading and unloading area; dividing ground points and non-ground points based on the three-dimensional point cloud data, extracting cargo pile boundary sequence points, and determining a target loading position by fitting a boundary curve; when the bucket of the unmanned loading machine is lifted and shielded, simulating cargo displacement change based on pre-scanned key feature points before shielding, and adaptively fusing actual measurement points and simulation points based on credibility weights to generate a first fusion model; and when the unmanned loader is far away from the loading and unloading area and the bucket shielding is removed, re-scanning to obtain a sparse point cloud, carrying out secondary fusion on the sparse point cloud and the first fusion model, and outputting a final environment model. According to the method, the accuracy of loading position judgment is improved, and the continuity of environment modeling of the unmanned loader is ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned loading, and in particular relates to a real-time modeling method and system suitable for the loading and unloading area of ​​an unmanned loader. Background Art

[0002] The statements herein merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Against the backdrop of the rapid development of intelligent and automated technologies, unmanned driving technology for construction machinery not only significantly improves operational efficiency but also effectively reduces operator risk in hazardous environments. As a key piece of equipment in the construction machinery sector, loaders play a vital role in various industries, including construction, mining, and logistics. Therefore, research on unmanned loader technology is of great practical significance.

[0004] The inventors found that due to the fuzzy boundaries of bulk cargo piles (such as sand and gravel piles, coal piles, etc.), traditional point cloud clustering algorithms have large errors in identifying the boundaries of areas where the slope of the cargo pile changes, which easily leads to misjudgment of the loading position; when the bucket is raised after loading, it will cause a certain amount of obstruction to the laser radar, and at this time it is impossible to obtain real-time point cloud data of the current environment, resulting in a long interruption time for the traditional SLAM system, causing breakpoints in the operation process, destroying the continuity of the unmanned loader environment modeling, and thus affecting the loading efficiency. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies in the above-mentioned prior art and provide a real-time modeling method and system suitable for the loading and unloading area of ​​an unmanned loader. Dynamic environmental modeling is performed based on three-dimensional point cloud data collected by multiple sensors, effectively identifying the boundary between the cargo pile and the ground, and improving the accuracy of loading position judgment. The continuity of the unmanned loader environment modeling is ensured through the model fusion method of pre-scanning before occlusion, dynamic completion during occlusion, and verification after occlusion is released.

[0006] In order to achieve the above object, the present invention is implemented through the following technical solutions: On the one hand, the technical solution of the present invention provides a real-time modeling method applicable to the loading and unloading area of ​​an unmanned loader, comprising: When an unmanned loader enters the loading and unloading area, it initiates a collaborative scan using the laser radar, binocular camera, and millimeter-wave radar it carries to obtain 3D point cloud data of the loading and unloading area. Based on 3D point cloud data, ground points and non-ground points are segmented, and the boundary sequence points of the cargo pile are extracted. The target loading position is determined by fitting the boundary curve. When the unmanned loader bucket is raised to block the view, the displacement change of the cargo is simulated based on the key feature points pre-scanned before the blockage, and the measured points and the simulated points are adaptively fused based on the credibility weight to generate the first fusion model; When the unmanned loader is away from the loading and unloading area and the bucket obstruction is released, a sparse point cloud is obtained by rescanning, and the sparse point cloud is fused with the first fusion model for a second time to output the final environment model.

[0007] In at least one embodiment, ground points and non-ground points are segmented based on three-dimensional point cloud data, and a sequence of boundary points of the cargo pile is extracted, specifically comprising: robustly fitting the ground plane through the RANSAC algorithm based on the three-dimensional point cloud data, and segmenting the ground points and non-ground points using a height threshold; extracting a sequence of intersection points between the cargo pile and the ground through neighborhood search or edge detection, and filtering out scattered cargo and dust, rain and fog through DBSCAN clustering to obtain an ordered sequence of boundary points.

[0008] In at least one embodiment, the target loading position is determined by fitting a boundary curve, specifically including: projecting the ordered boundary points in the ordered boundary point sequence onto the ground plane, and using a spline curve fitting algorithm to generate a smooth two-dimensional boundary curve; calculating the curvature of the curve and setting a threshold to filter out noise disturbances, and selecting the local protruding point with the largest curvature as the target loading position of the unmanned loader.

[0009] In at least one embodiment, the measured points and the simulated points are adaptively fused based on the credibility weight to generate a first fusion model, specifically including: the measured points use distance weights, giving super-high weights to the near points, and using exponential decay weights when the distance increases; the simulated points use double suppression weights, based on the global attenuation coefficient, and combined with the distance penalty to the nearest measured point to determine the weight of the simulated point; the space of the first fusion model is a three-dimensional voxel grid, and the coordinates of the fusion point in each unit are fused by weighted average of the coordinates of all the simulated points and the measured points in the unit according to their weights.

[0010] In at least one embodiment, the weight of the measured point is specifically: Where, is the distance between the lidar and the cargo; is the reliable cargo modeling distance, determined by the quality of the lidar; , is the variance of the lidar at the reference distance; This is the law that the optical sensor error increases with distance. It is usually set to 2, which is consistent with the law that the error of most optical sensors increases with the square of the distance.

[0011] In at least one embodiment, the simulation point weights are specifically:

[0012] Where, is the global attenuation coefficient, usually 0.3-0.7; is the Euclidean distance between the simulation point and the nearest measured point; It is the distance penalty coefficient. The larger the value, the faster the weight of the simulation point decreases due to the Euclidean distance.

[0013] In at least one embodiment, the fusion formula of the first fusion model is specifically:

[0014] Where, are the center coordinates of the voxel; is the number of simulation points within a voxel, is the number of measured points within a voxel; is the three-dimensional coordinate of the simulation point, is the three-dimensional coordinate of the measured point; is the weight of the simulation point, is the weight of the measured point.

[0015] In at least one embodiment, the fusion formula for secondary fusion of the sparse point cloud and the first fusion model is specifically:

[0016] Where, is the sparse point cloud coordinate obtained by rescanning after the occlusion is removed, is the weight of the sparse point cloud obtained by rescanning after the occlusion is removed, are the center coordinates of the first fused model voxel.

[0017] In at least one embodiment, the weight of the sparse point cloud obtained by rescanning after the occlusion is removed is specifically:

[0018] Where, is the confidence, adjusted according to the quality of the lidar; is the Euclidean distance of the sparse point cloud obtained by rescanning after the LiDAR distance occlusion is removed, It is the distance penalty coefficient of the sparse point cloud obtained by rescanning after the occlusion is removed. The larger the value, the faster the weight of the simulation point decreases due to the Euclidean distance.

[0019] On the other hand, the technical solution of the present invention also provides a real-time modeling system suitable for the loading and unloading area of ​​an unmanned loader, comprising: The point cloud data acquisition module is configured to: when the unmanned loader enters the loading and unloading area, initiate collaborative scanning based on the laser radar, binocular camera, and millimeter wave radar carried by the unmanned loader to obtain three-dimensional point cloud data of the loading and unloading area; The target loading position determination module is configured to: segment ground points and non-ground points based on the three-dimensional point cloud data, extract the boundary sequence points of the cargo pile, and determine the target loading position by fitting the boundary curve; The first fusion model generation module is configured to: when the bucket of the unmanned loader is raised to block the load, simulate the displacement change of the cargo based on the key feature points pre-scanned before the blockage, and adaptively fuse the measured points and the simulated points based on the credibility weight to generate the first fusion model; The final environment model generation module is configured to: when the unmanned loader is away from the loading and unloading area and the bucket obstruction is released, re-scan to obtain a sparse point cloud, perform a secondary fusion of the sparse point cloud with the first fusion model, and output the final environment model.

[0020] The beneficial effects of the technical solution of the present invention are as follows: 1) The real-time modeling method for unmanned loader loading and unloading areas of the present invention performs dynamic environmental modeling based on three-dimensional point cloud data collected by multiple sensors, effectively identifying the boundary between the cargo pile and the ground, and improving the accuracy of loading position judgment. The continuity of the unmanned loader environment modeling is ensured through a model fusion method that pre-scans before occlusion, dynamically completes during occlusion, and verifies after occlusion is released.

[0021] 2) The real-time modeling method of the present invention, which is applicable to the loading and unloading area of ​​an unmanned loader, requires a comprehensive scan of the cargo to obtain the target loading position when the loader first enters the site. The loading position is then determined by combining the fusion model with the previously scanned unaffected model. After each loading is completed, the loading position is determined according to the fusion model, thereby saving scanning time and improving efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0023] Figure 1 This is a schematic diagram of the steps of the real-time modeling method for the loading and unloading area of ​​an unmanned loader disclosed in Example 1 of the present invention; Figure 2 It is a flow chart of the real-time modeling method applicable to the loading and unloading area of ​​an unmanned loader disclosed in Example 1 of the present invention. DETAILED DESCRIPTION

[0024] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0025] As introduced in the background technology, the purpose of the present invention is to overcome the shortcomings of the above-mentioned existing technologies and provide a real-time modeling method and system suitable for the loading and unloading areas of unmanned loaders. Dynamic environment modeling is performed based on three-dimensional point cloud data collected by multiple sensors, effectively identifying the boundary between the cargo pile and the ground, and improving the accuracy of loading position judgment; through the model fusion method of pre-scanning before occlusion, dynamic completion during occlusion, and verification after occlusion is released, the continuity of unmanned loader environment modeling is guaranteed.

[0026] Example 1 In a typical embodiment of the present invention, Figure 1 As shown, this embodiment discloses a real-time modeling method applicable to the loading and unloading area of ​​an unmanned loader, including the following steps: S100. The unmanned loader enters the loading and unloading area and initiates a collaborative scan using the unmanned loader's onboard laser radar, binocular camera, and millimeter-wave radar to acquire 3D point cloud data of the loading and unloading area. S200. Segmenting ground points and non-ground points based on the three-dimensional point cloud data, extracting the boundary sequence points of the cargo pile, and determining the target loading position by fitting the boundary curve; S300. When the unmanned loader bucket is raised to block the occlusion, the cargo displacement change is simulated based on the key feature points pre-scanned before the occlusion, and the measured points and the simulated points are adaptively fused based on the credibility weight to generate a first fusion model; S400. When the unmanned loader is away from the loading and unloading area and the bucket blocks contact, rescan to obtain a sparse point cloud, perform a secondary fusion of the sparse point cloud with the first fusion model, and output a final environment model.

[0027] The real-time modeling method applicable to the loading and unloading area of ​​an unmanned loader will be described in detail below in conjunction with specific implementation methods.

[0028] like Figure 2As shown, in step S100, the unmanned loader is equipped with a laser radar, a binocular camera, and a millimeter-wave radar. The laser radar is used to perform SLAM modeling on the scanned environment, while the monocular camera continuously takes photos to obtain 2D images of the working environment. The millimeter-wave radar supplements this function, possessing a certain degree of penetration and capable of providing a certain amount of point cloud data even in obstructed conditions. When the unmanned loader reaches the loading and unloading area, the laser radar, binocular camera, and millimeter-wave radar initiate a coordinated scan, acquiring 3D point cloud data of the loading and unloading area for SLAM modeling.

[0029] In step S200, three-dimensional point cloud data are collected collaboratively based on the lidar, binocular camera and millimeter-wave radar. The three-dimensional point cloud data collected by the lidar, binocular camera and millimeter-wave radar are fused and unified into the same coordinate system with the help of multi-sensor calibration technology. Then, the RANSAC algorithm is used to robustly fit the ground plane, and the "ground points" and "non-ground points" are segmented using the height threshold. Then, the intersection point sequence between the cargo pile and the ground is extracted through neighborhood search or edge detection. The DBSCAN clustering algorithm is used to filter out influencing factors such as scattered cargo and dust, rain and fog to obtain an ordered boundary point sequence. Specifically, three non-collinear points near 0 in the z-axis direction of the point cloud are selected to fit a candidate plane model. The RANSAC algorithm continuously iterates and then calculates the distance from all points in the entire point cloud to the plane. Points with distances less than the set threshold (inlier threshold d, such as 0.1 meters) are regarded as "inliers" supporting the model. Finally, the plane model with the most inliers support is selected as the best ground plane estimate. The maximum inlier set of the model is the preliminary reliable ground point cloud. After obtaining the ground plane, points that are a certain distance above the ground are set as "non-ground points". The distance above the ground is the height threshold. If it is too small, some point clouds on the ground may be regarded as non-ground points. If it is too large, some protrusions may be set as the ground. Therefore, it can be based on The value of the height threshold is adjusted according to actual conditions. Then, a sequence of intersection points between the cargo pile and the ground is extracted through neighborhood search or edge detection. The DBSCAN clustering algorithm is used to filter out influencing factors such as scattered cargo and dust, rain, and fog. First, all data points are scanned, and objects that have at least a minimum number of neighboring points within the neighborhood radius are marked as "core points". Then, based on density accessibility, each core point and its directly accessible boundary points (points within the neighborhood radius of the core point but which do not meet the core condition) are expanded and connected into clusters. Finally, isolated points not associated with any core point are judged as noise points, while clusters with less than the minimum number of points may be scattered cargo. Clusters with a point number of 1 and no other point clouds around them are dust, rain, and fog, and finally an ordered sequence of boundary points is obtained.

[0030] Based on the obtained ordered boundary point sequence, the ordered boundary points are projected onto the ground plane, and a spline curve fitting algorithm is used to generate a smooth two-dimensional boundary curve. The curve curvature is then calculated by analyzing the extreme points of the second-order derivative, and a threshold is set to filter out noise disturbances. Finally, the local convex point with the largest curvature is selected as the target loading position of the unmanned loader. Specifically, based on the obtained ordered boundary points, several adjacent points are fitted into a curve using cubic splines, and a complete, smooth two-dimensional boundary curve is obtained by segmented fitting. The curve curvature is then calculated by analyzing the extreme points of the second-order derivative. To prevent the sudden protrusion of the boundary point cloud caused by dust, rain, and fog, which leads to a sudden increase in curvature and affects the curve curvature calculation, a threshold is set to filter out noise disturbances. The local convex point with the largest curve curvature can represent the outermost position of the cargo in the cargo pile. Loading at this point is easy to cut in and prevent cargo collapse. Therefore, the local convex point with the largest curvature is selected as the target loading position of the unmanned loader.

[0031] Because the unmanned loader needs to raise its bucket when loading cargo, this can obstruct the LiDAR and interrupt the LiDAR scan, affecting the modeling of the work area after loading and, in turn, the continuity of the environmental modeling. Therefore, in this embodiment, steps S300 and S400 propose a modeling method that includes pre-scanning before obstruction, dynamic completion during obstruction, and verification after obstruction is removed.

[0032] Specifically, the pre-scan before occlusion in step S300 refers to a preliminary scan of the post-loading working area using a laser radar after the unmanned loader bucket is loaded with cargo and before it is raised. Since the laser radar is close to the cargo pile at this time, the scanning range is relatively limited. Therefore, the pre-scan before occlusion records some edge information of the ground and the cargo and some key feature points such as the scanned cargo information to construct a partial real model of the cargo pile. Dynamic completion during occlusion refers to the use of a physics engine such as PyBullet to simulate the cargo displacement changes based on the key feature points of the pre-scan before occlusion after the bucket is raised and blocks the laser radar, and to construct a simulation model of the cargo pile after loading. The simulation model of the cargo pile is then integrated and adjusted with the partial real model of the cargo pile constructed by the pre-scan before occlusion to obtain a preliminary estimated model. The following takes PyBullet as an example to explain dynamic completion during occlusion in detail.

[0033] First, the PyBullet-based cargo loading simulation system employs a discretized modeling approach. The Hertz-Mindlin contact model defines cargo characteristics, including density, particle size, and friction coefficient. The shape of the cargo pile is also defined, and the size of the voxelized grid is set. Dynamic influence domain calculation is used to detect effective contact particles within a range of three times the bucket's size around the bucket in real time. Combined with bucket trajectory planning and mass conservation verification, the simulation output includes a voxel displacement field. The voxel displacement field (resolution file) is then parsed using Python's json module. Displacement vectors are extracted, mapped to a 3D voxel grid, and normalized. A custom Dataset class is constructed based on PyTorch to load data, integrating augmentation strategies such as 3D random rotation. The 3D CNN architecture is designed using an encoder-decoder structure, with the input layer adapting to three-channel displacement field data (dx, dy, dz). Finally, multi-GPU parallel training is used to achieve end-to-end processing from raw simulation data to a high-precision prediction model.

[0034] To simulate cargo in different states, the LiDAR point cloud acquired by the loader is converted into a file recognizable by PyBullet, which is then used for simulation to improve efficiency. When converting LiDAR-scanned bulk material point clouds (such as sand, gravel, and coal) into a PyBullet particle system, the point cloud is first subjected to voxel downsampling and statistical outlier removal to reduce data density. The processed point cloud coordinates are then batch-converted into spherical collider particles (GEOM_SPHERE). Each particle is assigned a radius, mass, and friction coefficient, and grouped and aggregated to optimize computational efficiency. Finally, the particles are imported into PyBullet via the API, ultimately achieving efficient physical simulation of the cargo pile.

[0035] Then, the cargo pile simulation model is fused and adjusted with the real model of the cargo pile constructed by pre-scanning before occlusion. The fusion method is as follows: First, the real-time pose data of the unmanned loader is used to robustly align the simulation points of the cargo pile simulation model with the local high-precision measured point cloud obtained by close-range scanning, and the two are precisely aligned to the same coordinate system.

[0036] Subsequently, adaptive spatial fusion based on credibility weights is performed to fuse the measured points with the simulated points to generate the first fusion model. Specifically, the measured points are distance-weighted, and the near points are given an extremely high weight (node ​​1.0) according to the sensor distance model. As the distance increases, an exponential decay weight is used, which is specifically expressed as:

[0037] Where, is the distance between the lidar and the cargo; is the reliable cargo modeling distance, determined by the quality of the lidar; , is the variance of the lidar at the reference distance, where the reference distance is the distance range for which the lidar specification gives the accuracy; This is the law that the optical sensor error increases with distance. It is usually set to 2, which is consistent with the law that the error of most optical sensors increases with the square of the distance.

[0038] The simulation point applies a double suppression strategy through the global attenuation coefficient Limiting the maximum influence of the simulation point and combining it with the distance penalty from the nearest measured point ensures that the weight of the predicted point is always significantly lower than the measured point. The weight of the simulation point is specifically expressed as:

[0039] Where, is the global attenuation coefficient, usually 0.3-0.7; is the Euclidean distance between the simulation point and the nearest measured point; It is the distance penalty coefficient. The larger the value, the faster the weight of the simulation point decreases due to the Euclidean distance.

[0040] At the fusion execution layer, the loader's operating space is divided into a 3D voxel grid, and weighted coordinate synthesis is implemented within each cell: the coordinates of the fusion point are calculated by weighted average of the coordinates of all predicted points and measured points within the cell. The measured area is dominated by the high-weighted fusion result, and the unscanned areas are supplemented by the simulation model to generate the first fusion model. The fusion formula is as follows:

[0041] Where, are the center coordinates of the voxel; is the number of simulation points within a voxel, is the number of measured points within a voxel; is the three-dimensional coordinate of the simulation point, is the three-dimensional coordinate of the measured point; is the weight of the simulation point, is the weight of the measured point.

[0042] In step S400, the post-blocking verification means that when the unmanned loader is a certain distance away from the loading and unloading area, the bucket block will be weakened or even removed. However, at this time, the laser radar is far away from the cargo pile and the number of point clouds is relatively sparse. At this time, the laser radar rescans the cargo pile to obtain a sparse point cloud, and fuses the sparse point cloud with the first fusion model, sets the weights of the two, and thus obtains the final environment model. Specifically, in the first fusion model, The voxel centered at contains the sparse point cloud coordinates obtained by rescanning after the occlusion is lifted. These sparse point clouds are compared with Perform secondary fusion to obtain the final environment model. The formula for secondary fusion is as follows:

[0043] Where, is the sparse point cloud coordinate obtained by rescanning after the occlusion is removed, is the weight of the sparse point cloud obtained by rescanning after the occlusion is removed, are the center coordinates of the first fused model voxel.

[0044] The weight of the sparse point cloud obtained by rescanning after occlusion is removed , specifically:

[0045] Where, is the confidence, adjusted according to the quality of the lidar; is the Euclidean distance of the sparse point cloud obtained by rescanning after the LiDAR distance occlusion is removed, This is the distance penalty coefficient for the sparse point cloud obtained by rescanning after occlusion is removed. A larger value indicates a faster decrease in the weight of the simulated point due to Euclidean distance. After obtaining the final environment model, the next loading location can be directly determined by fitting the ground boundary curve.

[0046] Example 2 In a typical embodiment of the present invention, this embodiment discloses a real-time modeling system for the loading and unloading area of ​​an unmanned loader, characterized by comprising: The point cloud data acquisition module is configured to: when the unmanned loader enters the loading and unloading area, initiate collaborative scanning based on the laser radar, binocular camera, and millimeter wave radar carried by the unmanned loader to obtain three-dimensional point cloud data of the loading and unloading area; The target loading position determination module is configured to: segment ground points and non-ground points based on the three-dimensional point cloud data, extract the boundary sequence points of the cargo pile, and determine the target loading position by fitting the boundary curve; The first fusion model generation module is configured to: when the bucket of the unmanned loader is raised to block the load, simulate the displacement change of the cargo based on the key feature points pre-scanned before the blockage, and adaptively fuse the measured points and the simulated points based on the credibility weight to generate the first fusion model; The final environment model generation module is configured to: when the unmanned loader is away from the loading and unloading area and the bucket obstruction is released, re-scan to obtain a sparse point cloud, perform a secondary fusion of the sparse point cloud with the first fusion model, and output the final environment model.

[0047] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A real-time modeling method for the loading and unloading area of ​​an unmanned loader, characterized in that: include: When an unmanned loader enters the loading and unloading area, it initiates a collaborative scan using the laser radar, binocular camera, and millimeter-wave radar it carries to obtain 3D point cloud data of the loading and unloading area. Based on 3D point cloud data, ground points and non-ground points are segmented, and the boundary sequence points of the cargo pile are extracted. The target loading position is determined by fitting the boundary curve. When the unmanned loader bucket is raised to block the view, the displacement change of the cargo is simulated based on the key feature points pre-scanned before the blockage, and the measured points and the simulated points are adaptively fused based on the credibility weight to generate the first fusion model; When the unmanned loader is away from the loading and unloading area and the bucket obstruction is released, a sparse point cloud is obtained by rescanning, and the sparse point cloud is fused with the first fusion model for a second time to output the final environment model.

2. The real-time modeling method for the loading and unloading area of ​​an unmanned loader according to claim 1, characterized in that: Based on 3D point cloud data, ground points and non-ground points are segmented, and the boundary sequence points of the cargo pile are extracted. Specifically, based on the 3D point cloud data, the ground plane is robustly fitted using the RANSAC algorithm, and the ground points and non-ground points are segmented using the height threshold; the intersection point sequence between the cargo pile and the ground is extracted through neighborhood search or edge detection, and scattered cargo and dust, rain and fog are filtered out through DBSCAN clustering to obtain an ordered boundary point sequence.

3. The real-time modeling method for the loading and unloading area of ​​an unmanned loader according to claim 2, characterized in that: The target loading position is determined by fitting the boundary curve, specifically including: projecting the ordered boundary points in the ordered boundary point sequence onto the ground plane, and using the spline curve fitting algorithm to generate a smooth two-dimensional boundary curve; calculating the curve curvature and setting a threshold to filter out noise disturbances, and selecting the local convex point with the largest curvature as the target loading position of the unmanned loader.

4. The real-time modeling method for the loading and unloading area of ​​an unmanned loader according to claim 1, characterized in that: The measured points and simulation points are adaptively fused based on the credibility weight to generate the first fusion model. Specifically, the measured points use distance weights, giving super-high weights to the near points, and use exponential decay weights when the distance increases; the simulation points use double suppression weights, which are based on the global attenuation coefficient and combined with the distance penalty to the nearest measured point to determine the weight of the simulation point; the space of the first fusion model is a three-dimensional voxel grid, and the coordinates of the fusion point in each unit are the weighted average fusion of the coordinates of all simulation points and measured points in the unit according to their weights.

5. The real-time modeling method for the loading and unloading area of ​​an unmanned loader according to claim 4, characterized in that: The weights of the measured points are as follows: Where, is the distance between the lidar and the cargo; is the reliable cargo modeling distance, determined by the quality of the lidar; , is the variance of the lidar at the reference distance; This is the law that the optical sensor error increases with distance. It is usually set to 2, which is consistent with the law that the error of most optical sensors increases with the square of the distance.

6. The real-time modeling method for the loading and unloading area of ​​an unmanned loader according to claim 4, characterized in that: The specific weights of simulation points are: Where, is the global attenuation coefficient, usually 0.3-0.7; is the Euclidean distance between the simulation point and the nearest measured point; It is the distance penalty coefficient. The larger the value, the faster the weight of the simulation point decreases due to the Euclidean distance.

7. The real-time modeling method for the loading and unloading area of ​​an unmanned loader according to claim 4, characterized in that: The fusion formula of the first fusion model is as follows: Where, are the center coordinates of the voxel; is the number of simulation points within a voxel, is the number of measured points within a voxel; is the three-dimensional coordinate of the simulation point, is the three-dimensional coordinate of the measured point; is the weight of the simulation point, is the weight of the measured point.

8. The real-time modeling method for the loading and unloading area of ​​an unmanned loader according to claim 1, characterized in that: The fusion formula for secondary fusion of sparse point cloud and the first fusion model is as follows: Where, is the sparse point cloud coordinate obtained by rescanning after the occlusion is removed, is the weight of the sparse point cloud obtained by rescanning after the occlusion is removed, are the center coordinates of the first fused model voxel.

9. The real-time modeling method for the loading and unloading area of ​​an unmanned loader according to claim 8, characterized in that: The weights of the sparse point cloud obtained by rescanning after the occlusion is removed are: Where, is the confidence, adjusted according to the quality of the lidar; is the Euclidean distance of the sparse point cloud obtained by rescanning after the LiDAR distance occlusion is removed, It is the distance penalty coefficient of the sparse point cloud obtained by rescanning after the occlusion is removed. The larger the value, the faster the weight of the simulation point decreases due to the Euclidean distance.

10. A real-time modeling system for unmanned loader loading and unloading areas, characterized by: include: The point cloud data acquisition module is configured to: when the unmanned loader enters the loading and unloading area, initiate collaborative scanning based on the laser radar, binocular camera, and millimeter wave radar carried by the unmanned loader to obtain three-dimensional point cloud data of the loading and unloading area; The target loading position determination module is configured to: segment ground points and non-ground points based on the three-dimensional point cloud data, extract the boundary sequence points of the cargo pile, and determine the target loading position by fitting the boundary curve; The first fusion model generation module is configured to: when the bucket of the unmanned loader is raised to block the load, simulate the displacement change of the cargo based on the key feature points pre-scanned before the blockage, and adaptively fuse the measured points and the simulated points based on the credibility weight to generate the first fusion model; The final environment model generation module is configured to: when the unmanned loader is away from the loading and unloading area and the bucket obstruction is released, re-scan to obtain a sparse point cloud, perform a secondary fusion of the sparse point cloud with the first fusion model, and output the final environment model.

Images