Method and device for local updating of a three-dimensional point cloud map based on laser projection, and medium
By employing a laser-projection-based method for local updating of 3D point cloud maps, old obstacles are removed using voxel probability maps and ray projection models, and ground point clouds are extracted using the RANSAC algorithm. This solves the problems of low update efficiency and sparse ground point clouds in 3D point cloud maps, and achieves high-precision environmentally adaptive updates.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHENGDU SIWI POWER ELECTRONICS TECH
- Filing Date
- 2026-02-25
- Publication Date
- 2026-06-02
AI Technical Summary
Existing technologies have low efficiency in updating 3D point cloud maps after environmental changes, cannot effectively remove the afterimages of dynamic obstacles, and the updated map ground point cloud is sparse, leading to failure in localization and path planning.
A local update method for 3D point cloud maps based on laser projection is adopted. By constructing a voxel probabilistic map and a ray projection model, old obstacles are removed and new obstacles are added. The RANSAC algorithm is used to extract high-precision ground point clouds and save them independently to generate an updated point cloud map.
It achieves efficient and accurate local updates of 3D point cloud maps, maintains the accuracy of ground point clouds, ensures map consistency with the environment and compatibility with lane drawing software, and improves the robustness and reliability of generated maps.
Smart Images
Figure CN122134956A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of robot autonomous navigation and 3D vision technology, and more specifically, to a method, apparatus and medium for local updating of 3D point cloud maps based on laser projection. Background Technology
[0002] In the fields of mobile robotics and autonomous driving, high-precision 3D point cloud maps are fundamental for localization, navigation, and path planning. However, real-world application environments are often dynamically changing; for example, cars parked on the roadside may be driven away, and old houses may be demolished. Traditional mapping methods typically assume a static environment. Once the environment changes, the pre-built map will deviate from the actual environment, leading to robot localization loss or path planning failure. Furthermore, maps updated using traditional methods have sparse ground point clouds, making them unsuitable for lane mapping software such as Vector Map Builder.
[0003] Current solutions mainly include re-scanning the entire map or local patching based on keyframes. Full map reconstruction is time-consuming, resource-intensive, and computationally expensive, making it unsuitable for applications with high real-time requirements. Traditional simple point cloud stitching methods cannot effectively remove old obstacle point cloud information, resulting in a large number of false obstacles in the map. Therefore, a method is needed that can efficiently and accurately update 3D point cloud maps locally while preserving the ground point cloud. Summary of the Invention
[0004] The technical problem to be solved by the present invention is to provide a method, device and medium for local updating of three-dimensional point cloud maps based on laser projection, which can effectively solve the problems of low map updating efficiency after environmental changes, inability to effectively remove dynamic obstacle shadows and sparse ground point cloud of updated map in the prior art.
[0005] The solution adopted by this invention to solve the technical problem is: A method for local updating of a 3D point cloud map based on laser projection specifically includes the following steps: Step S1: Read the initial 3D point cloud map, perform preprocessing, and obtain the processed point cloud; Step S2: Map the processed point cloud to voxel space to construct a voxel probability map; Step S3: Extract and save the ground point cloud; Step S4: Acquire new observation data, including lidar location and environmental point cloud data; Step S5: Based on the ray projection model, virtual rays are emitted from the lidar position to each observation point, the negative probabilities of voxels along the ray path are updated, and the positive probabilities of the voxels at the observation points are updated simultaneously to obtain the updated voxel probability map. Step S6: Based on the updated voxel probability map, restore the locally updated point cloud to form an updated non-ground point cloud set; Step S7: Combine the non-ground point cloud set with the saved ground point cloud to generate an updated 3D point cloud map.
[0006] In some possible implementations, step S1 specifically refers to: Step S11: Define a three-dimensional voxel mesh, wherein the resolution parameters of the three-dimensional voxel mesh in each direction are as follows: ; The initial 3D point cloud map read through the traversal Calculate the maximum and minimum values of all points in the 3D point cloud map along the X, Y, and Z axes, and determine the 3D bounding box that can enclose the entire point cloud.
[0007] Step S12: Based on resolution parameters Divide the 3D bounding box into Individual factors; Calculate each point in the point cloud Voxel Index The calculation formula is as follows: ; in, This indicates rounding down. These are the minimum coordinates of the bounding box.
[0008] Step S13: Calculate the geometric centroid of all points within each non-empty voxel, and use it as the representative point of that non-empty voxel. Assuming that non-empty voxels contain A point, center of gravity The calculation formula is: ; Step S14: Use the set of centroids calculated from all non-empty voxels as the output point cloud. The data sparsity processing is completed to generate a filtered point cloud. Step S15: Use the RANSAC algorithm to perform ground fitting; The mathematical model of a three-dimensional plane is defined as follows: ; in, Let be the normal vector of the plane, and satisfy . , The distance from the plane to the origin; Step S16: Iteratively select the optimal ground model; Set the maximum number of iterations and distance threshold In each iteration, the process begins with the point cloud. Three non-collinear points are randomly selected from the given data, and the plane parameters determined by these points are calculated. Then calculate the perpendicular distance from all other points in the point cloud to the plane. Finally, the distance was calculated. The number of points; among them, the plane parameter with the most points is selected as the optimal ground model; Step S17: Obtain the normal vector of the optimal ground model ; Calculate the normal vector and the unit vector of the Z-axis in the world coordinate system. Rotation matrix between ; Multiply all points in the point cloud map by the rotation matrix. This makes the ground normal vector parallel to the Z-axis, thus obtaining the calibrated point cloud; Step S18: Based on the calibrated point cloud and the optimal ground model, traverse all points again; This will satisfy the distance condition. The points are marked as ground point cloud sets. The remaining points that do not meet the conditions are marked as a non-ground point cloud set. .
[0009] In some possible implementations, step S2 specifically includes the following steps: Step S21: Determine the memory space for the voxel probability map and resolution ,in ; Calculate non-terrestrial point cloud sets The three-dimensional boundary is defined, and a preset safety boundary is extended outward from this boundary to determine the physical coverage of the voxel probability map. ; Based on physical coverage and resolution Memory space required to compute the voxel probability map Finally, allocate memory. A hash table of a certain size is used to store the probability state of each voxel; ; Step S22: Traverse the non-terrestrial point cloud set Each point in Using the map coordinate origin and resolution Calculate the unique index of this point in the voxel probability map. : ; The calculated unique index The position is marked as "occupied" and the position not marked as "unknown". Step S23: Establish a conversion model between probability values and stored values, using logarithmic probability representation: ; Initialize the logarithmic probability values of all "unknown state" voxels in the map to . =0, the corresponding probability is This indicates that there is no observation information for this area; Initialize the logarithmic probability value of the "occupancy state" voxel to the preset prior occupancy value. .
[0010] In some possible implementations, step S4 specifically includes the following steps: Step S41: Acquire current real-time scanning data from the lidar, the data including a series of three-dimensional points in the lidar coordinate system. ; Step S42: Synchronously acquire the pose information of the LiDAR in the global map coordinate system corresponding to the real-time scanning data, represented as a 4×4 transformation matrix. ,matrix Includes rotation matrix Translation vector ; Step S43: Use the pose matrix to transform the scan data to the global map coordinate system to obtain the global observation point set used to update the map. and the global center coordinates of the sensor at the current moment. ;in, .
[0011] In some possible implementations, step S5 specifically includes the following steps: Step S51: For the global observation point set Each observation point in Construct a line from the lidar center Shot towards line segments; Step S52: Calculate the sequence of all voxel indices that the line segment passes through in the 3D mesh; Step S53: Update the logarithmic probability values of all voxels corresponding to the line segment to obtain the updated voxel probability map.
[0012] In some possible implementations, step S53 specifically refers to: For all voxels that are passed through but not hit on the line segment path, they are considered to be in an "idle state"; their logarithmic probability values are then reduced and updated. ; in, The log-probability value before updating for all voxels that have passed through but not been hit along the line segment path; A negative value indicates an increased probability that the region is empty; Step S54: For a voxel hit at the end of a line segment, treat it as "occupied"; cumulatively update its logarithmic probability value: ; in, This is the log-probability value of the voxel hit at the end of the line segment before the update; If after each update, if Then set to ,like Then set to ; in, The maximum threshold; The minimum threshold.
[0013] In some possible implementations, step S6 specifically includes the following steps: Step S61: Traverse the updated voxel probability map; Step S62: Set a judgment threshold ,in Read the log-probability value of each voxel in the updated voxel probability map; like > The voxel was determined to be a valid obstacle. like ≤ The voxel is determined to be an empty region and discarded.
[0014] Step S63: For voxels determined to be valid obstacles, based on the resolution parameters... Calculate its corresponding three-dimensional geometric coordinates The calculation formula is as follows: ; in Indicates the origin of the map. For voxel indexing; All recovered points are combined into an updated non-terrestrial point cloud set. .
[0015] In some possible implementations, step S7 specifically includes the following steps: Step S71: Read ground point cloud data from step S18 ; Step S72: Aggregate non-terrestrial point clouds ground point clouds Merge the points to obtain the merged point cloud. : .
[0016] Step S73: Process the merged point cloud Perform final data encapsulation, add header information, save, and generate the updated map.
[0017] A local update device for a 3D point cloud map based on laser projection, and a local update method for a 3D point cloud map based on laser projection as described above, comprising an initialization module, an update module, and a reconstruction module; The initialization module is used to execute steps S1-S3; The update module is used to execute steps S4-S5; The reconstruction module is used to perform steps S6-S7.
[0018] A computer terminal storage medium stores computer terminal executable instructions, which are used to execute the laser projection-based three-dimensional point cloud map local update method as described above.
[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention introduces a light projection model and utilizes the penetration characteristics of light to update the voxels along the path of light with negative probability. Once the lidar detects that an original obstacle area is now being traversed by light, the system will automatically reduce the occupancy probability of that area, thereby clearing the disappeared obstacles in the point cloud map. This ensures that the point cloud map can reflect the latest state of the environment in real time and significantly improves the consistency between the map and the real environment. This invention employs a strategy of separating and independently storing ground point clouds. During the probability update stage, it only processes obstacle point clouds while retaining the original high-precision ground data. This results in a final map that possesses both updated environmental features and an accurate ground layer, effectively ensuring compatibility with lane drawing software such as "Vector Map Builder". This invention introduces a logarithmic probability statistical model, allowing the final state of a voxel to be determined by the cumulative results of multiple observations. This effectively filters out the instantaneous measurement noise of the lidar, and the voxel will only appear on the map when a certain location is repeatedly identified as an obstacle, thereby improving the robustness and reliability of the generated map. Attached Figure Description
[0020] Figure 1 This is a flowchart of the process of the present invention; Detailed Implementation In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. The terms "first," "second," and similar terms used in this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, "a" or "one," etc., do not indicate a quantity limitation, but rather indicate the existence of at least one. In the implementation of this application, "and / or" describes the association relationship of related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. In the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more. For example, multiple positioning posts refer to two or more positioning posts. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.
[0021] The present invention will now be described in detail.
[0022] like Figure 1 As shown, a method for local updating of a 3D point cloud map based on laser projection mainly includes three stages: initialization, updating, and reconstruction. The initialization module performs the initialization phase by first reading the original point cloud map and downsampling it using voxel grid filtering. Then, the RANSAC algorithm is used to extract high-precision ground point clouds, which are saved independently. The remaining non-ground point clouds are mapped onto a voxel grid to initialize the voxel probability map.
[0023] The update module performs the update phase, reading in new LiDAR scan data and transforming it to the map coordinate system. A ray casting algorithm is used to simulate the process of light traversing space. For each ray, voxels along its path are marked as free, and voxels at its point of impact are marked as occupied. This mechanism automatically clears obstacles in the old map that have been traversed by the light rays and generates newly detected objects.
[0024] The reconstruction module is used to perform the reconstruction phase. Based on the updated voxel probability map, high-probability voxels are restored to point clouds and then fused with the previously saved ground point clouds to finally generate the updated map.
[0025] The method for local updating of a 3D point cloud map includes the following steps: S1: Read the initial 3D point cloud map, perform preprocessing, and obtain the processed point cloud; S2: Map the processed point cloud to voxel space to construct a voxel probability map; S3: Extract and save the ground point cloud; S4: Acquire new observation data, including lidar location information and environmental point cloud data; S5: Based on the ray projection model, virtual rays are emitted from the lidar position to each observation point, updating the negative probabilities of voxels along the ray path and the positive probabilities of the voxels where the observation point is located. S6: Restore the locally updated point cloud based on the updated voxel probability map; S7: Combine the non-ground point cloud set with the saved ground point cloud to generate an updated 3D point cloud map; This embodiment ensures the effective removal of obstacles that have disappeared from the 3D point cloud map, while also avoiding the loss of accuracy caused by voxelized meshes on flat ground.
[0026] In this embodiment, step S1 includes the following sub-steps: Step S11: Set the anisotropic resolution parameters of a three-dimensional voxel mesh. ; The initial 3D point cloud map read through the traversal Calculate the maximum and minimum values of all points in the 3D point cloud map along the X, Y, and Z axes to determine the 3D bounding box that can enclose the entire point cloud. Step S12: Based on resolution parameters The bounding box is divided into Individual element. Calculate each point in the point cloud. Voxel Index The calculation formula is as follows: ; in, This indicates rounding down. These are the minimum coordinates of the bounding box.
[0027] Step S13: Within each non-empty voxel, instead of directly selecting the voxel's geometric center, calculate the geometric centroid of all points within that voxel, and use this as the representative point of that voxel; assuming a voxel contains A point, center of gravity The calculation formula is: ; Step S14: Use the set of centroids calculated from all non-empty voxels as the downsampled output point cloud. Complete the data sparsity processing; Step S15: Use the RANSAC algorithm to fit the ground. Define the mathematical model of the three-dimensional spatial plane as follows: ; in, Let be the normal vector of the plane, and satisfy . , This is the distance from the plane to the origin.
[0028] Step S16: Set the maximum number of iterations and distance threshold In each iteration, the downsampled point cloud is first... Three non-collinear points are randomly selected from the given data, and the plane parameters determined by these points are calculated. Then calculate the perpendicular distance from all other points in the point cloud to the plane. Finally, the distance was calculated. The number of points. During this iteration, the plane parameter with the most points is selected as the optimal ground model.
[0029] Step S17: Obtain the normal vector of the optimal ground model Calculate the normal vector and the unit vector of the Z-axis in the world coordinate system. Rotation matrix between Multiply all points in the point cloud map by a rotation matrix. This ensures that the ground's normal vector is parallel to the Z-axis, thereby correcting the map's tilt and ensuring the ground is aligned correctly in subsequent processing. On the horizontal plane, the calibrated point cloud is obtained.
[0030] Step S18: Based on the calibrated point cloud and the optimal ground model, traverse all points again. Points that meet the distance condition will be... The points are marked as ground point cloud sets. Separate and save these points from the current processing stream. Mark the remaining points that do not meet the conditions as a non-ground point cloud set. This is used for subsequent construction of voxel probability maps.
[0031] This embodiment uses a voxel grid filtering method to divide the initial point cloud space into tiny grids and calculate the geometric centroid of points within the grids, achieving downsampling and data sparsification of the original 3D point cloud map. Its effect is to significantly reduce the amount of map data, filter out high-frequency noise from the original scan, and greatly reduce the computational load of subsequent algorithms while preserving the main geometric features of the environment.
[0032] In this embodiment, step S2 includes the following sub-steps: Step S21: Determine the memory space for the voxel probability map and resolution
[0033] First, calculate the non-ground point cloud set. The three-dimensional boundary is defined, and a preset safety boundary is extended outward from this boundary to determine the physical coverage of the voxel probability map. The memory space required to calculate the voxel probability map based on the physical coverage and resolution. Finally, allocate memory. A hash table of size is used to store the probability state of each voxel.
[0034] ; Step S22: Traverse Each point in Using the map coordinate origin and resolution Calculate the unique index of this point in the voxel probability map. : ; Mark all calculated index locations as "occupied" and mark locations on the map not marked by this step as "unknown".
[0035] Step S23: Establish a conversion model between probability values and stored values, using logarithmic probability representation: ; Initialize the logarithmic probability values of all "unknown state" voxels in the map to . =0, therefore the corresponding probability is This indicates that there is no observation information for this area; the logarithmic probability value of the "occupied state" voxel is initialized to the preset prior occupancy value. This step completes the transformation from "geometric point cloud" to "probability field," providing a mathematical basis for subsequent dynamic updates using laser projection.
[0036] This embodiment uses the RANSAC algorithm to fit a planar model and filters the point cloud based on a distance threshold. This achieves the function of separating and independently saving high-precision ground point clouds from non-ground obstacle point clouds, preventing thickness errors in the ground due to discretization during subsequent voxelization, and ensuring the flatness of the ground area in the final generated map.
[0037] In this embodiment, step S3 is to aggregate the separated ground point clouds. Stored in a separate memory container, it does not participate in the ray projection update in subsequent step S5, in order to maintain the original geometric accuracy of the ground.
[0038] In this embodiment, step S4 includes the following sub-steps: Step S41: Acquire the current real-time scanning data from the lidar. The data consists of a series of three-dimensional points in the lidar coordinate system. composition.
[0039] Step S42: Synchronously acquire the pose information of the LiDAR in the global map coordinate system corresponding to the real-time scanning data, represented as a 4×4 transformation matrix. ,matrix Includes rotation matrix Translation vector .
[0040] Step S43: Use the pose matrix to transform the scan data to the global map coordinate system to obtain the global observation point set used to update the map. and the global center coordinates of the sensor at the current moment. ;in, .
[0041] This embodiment achieves the function of uniformly transforming local observation points to the global map coordinate system by reading real-time LiDAR scanning data and combining it with the LiDAR pose matrix for coordinate transformation. This provides an instant input that is aligned with the original map space for map updates, ensuring that subsequent ray projection calculations are performed under a unified spatial reference.
[0042] In this embodiment, step S5 includes the following sub-steps: Step S51: For the global observation point set Each observation point in Construct a line from the lidar center Shot towards The line segment.
[0043] Step S52: Calculate the sequence of all voxel indices that the line segment passes through in the 3D mesh.
[0044] Step S53: For all voxels that have been passed through but not hit on the line segment path, treat them as "idle". Calculate the logarithmic probability value for these voxels. Perform deduction update: ; in, A negative value indicates an increased probability that the area is empty. This step effectively removes existing but currently penetrated dynamic obstacles from the map.
[0045] For voxels hit at the end of a line segment, they are considered to be in an "occupied state"; the logarithmic probability value of that voxel is cumulatively updated. ; To prevent the system from becoming stagnant due to excessively high or low probability values, a maximum threshold is set. and minimum threshold After each update, if Then set to ,like Then set to .
[0046] This embodiment uses a ray projection model to simulate the process of a laser beam traveling from the center of a lidar to an observation point. It performs negative probability updates on the voxels that the ray passes through and positive probability updates on the voxels that are hit at the end, thus realizing the function of dynamically clearing old obstacles that have disappeared and adding new obstacles at the probability level.
[0047] In this embodiment, step S6 includes the following sub-steps: Step S61: Traverse the updated voxel probability map; Step S62: Set a judgment threshold. ,in Read the log-probability value of each voxel in the updated voxel probability map; like The voxel was determined to be a valid obstacle. like The voxel is determined to be an empty region and discarded.
[0048] Step S63: For voxels determined to be valid obstacles, based on resolution... Calculate its corresponding three-dimensional geometric coordinates The calculation formula is as follows: ; in Indicates the origin of the map. For voxel indexing; All recovered points are combined into an updated non-terrestrial point cloud set. .
[0049] In this embodiment, by traversing the voxel probability map and setting a probability threshold for filtering, voxels with an occupation probability higher than the threshold are extracted and their geometric center coordinates are calculated. This realizes the function of recovering the updated obstacle 3D point cloud from the abstract probability statistical grid, and converts the updated statistical results back into intuitive geometric point cloud data, filtering out low-probability measurement noise and transient interference.
[0050] In this embodiment, step S7 includes the following sub-steps: Step S71: Read ground point cloud data from step S18 .
[0051] Step S72: Aggregate non-terrestrial point clouds ground point clouds Merge:
[0052] Because the ground point cloud retains its original accuracy, while the non-ground point cloud is dynamically updated, this stitching method ensures both the flatness of the ground and the real-time reflection of environmental changes.
[0053] Step S73: Process the merged point cloud Perform final data encapsulation, add header information, and save as a standard .PCD point cloud file format.
[0054] In this embodiment, the dynamic obstacle point cloud recovered in step S6 is merged with the high-precision ground point cloud saved in step S3 by the point cloud set stitching method to generate the final updated three-dimensional point cloud map. This map contains the latest dynamic change information of the environment and completely preserves the original geometric accuracy of the ground, which solves the problem of not being able to draw lanes due to the sparse ground point cloud in the traditional update method.
[0055] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.
Claims
1. A method for local updating of a 3D point cloud map based on laser projection, characterized in that, Specifically, the following steps are included: Step S1: Read the initial 3D point cloud map, perform preprocessing, and obtain the processed point cloud; Step S2: Map the processed point cloud to voxel space to construct a voxel probability map; Step S3: Extract and save the ground point cloud; Step S4: Acquire new observation data, including lidar location and environmental point cloud data; Step S5: Based on the ray projection model, virtual rays are emitted from the lidar position to each observation point, the negative probabilities of voxels along the ray path are updated, and the positive probabilities of the voxels at the observation points are updated simultaneously to obtain the updated voxel probability map. Step S6: Based on the updated voxel probability map, restore the locally updated point cloud to form an updated non-ground point cloud set; Step S7: Combine the non-ground point cloud set with the saved ground point cloud to generate an updated 3D point cloud map.
2. The method for local updating of a 3D point cloud map based on laser projection according to claim 1, characterized in that, Step S1 specifically refers to: Step S11: Define a three-dimensional voxel mesh, wherein the resolution parameters of the three-dimensional voxel mesh in each direction are as follows: ; The initial 3D point cloud map read through the traversal Calculate the maximum and minimum values of all points in the 3D point cloud map along the X, Y, and Z axes, and determine the 3D bounding box that can enclose the entire point cloud. Step S12: Based on resolution parameters Divide the 3D bounding box into Individual factors; Calculate each point in the point cloud Voxel Index The calculation formula is as follows: ； in, This indicates rounding down. These are the minimum coordinates of the bounding box; Step S13: Calculate the geometric centroid of all points within each non-empty voxel, and use it as the representative point of that non-empty voxel. Assuming that non-empty voxels contain A point, center of gravity The calculation formula is: ； Step S14: Use the set of centroids calculated from all non-empty voxels as the output point cloud. The data sparsity processing is completed to generate a filtered point cloud. Step S15: Use the RANSAC algorithm to perform ground fitting; The mathematical model of a three-dimensional plane is defined as follows: ； in, Let be the normal vector of the plane, and satisfy . , The distance from the plane to the origin; Step S16: Iteratively select the optimal ground model; Set the maximum number of iterations and distance threshold In each iteration, the process begins with the point cloud. Three non-collinear points are randomly selected from the given data, and the plane parameters determined by these points are calculated. Then calculate the perpendicular distance from all other points in the point cloud to the plane. Finally, the distance was calculated. The number of points; among them, the plane parameter with the most points is selected as the optimal ground model; Step S17: Obtain the normal vector of the optimal ground model ; Calculate the normal vector and the unit vector of the Z-axis in the world coordinate system. Rotation matrix between ; Multiply all points in the point cloud map by the rotation matrix. This makes the ground normal vector parallel to the Z-axis, thus obtaining the calibrated point cloud; Step S18: Based on the calibrated point cloud and the optimal ground model, traverse all points again; This will satisfy the distance condition. The points are marked as ground point cloud sets. The remaining points that do not meet the conditions are marked as a non-ground point cloud set. .
3. The method for local updating of a 3D point cloud map based on laser projection according to claim 2, characterized in that, Step S2 specifically includes the following steps: Step S21: Determine the memory space for the voxel probability map and resolution ,in ; Calculate non-terrestrial point cloud sets The three-dimensional boundary is defined, and a preset safety boundary is extended outward from this boundary to determine the physical coverage of the voxel probability map. ; Based on physical coverage and resolution Memory space required to compute the voxel probability map Finally, allocate memory. A hash table of a certain size is used to store the probability state of each voxel; ； Step S22: Traverse the non-terrestrial point cloud set Each point in Using the map coordinate origin and resolution Calculate the unique index of this point in the voxel probability map. : ； The calculated unique index The position is marked as "occupied" and the position not marked as "unknown". Step S23: Establish a conversion model between probability values and stored values, using logarithmic probability representation: ； Initialize the logarithmic probability values of all "unknown state" voxels in the map to . =0, the corresponding probability is This indicates that there is no observation information for this area; Initialize the logarithmic probability value of the "occupancy state" voxel to the preset prior occupancy value. .
4. The method for local updating of a 3D point cloud map based on laser projection according to claim 3, characterized in that, Step S4 specifically includes the following steps: Step S41: Acquire current real-time scanning data from the lidar, the data including a series of three-dimensional points in the lidar coordinate system. ; Step S42: Synchronously acquire the pose information of the LiDAR in the global map coordinate system corresponding to the real-time scanning data, represented as a 4×4 transformation matrix. ,matrix Includes rotation matrix Translation vector ; Step S43: Use the pose matrix to transform the scan data to the global map coordinate system to obtain the global observation point set used to update the map. and the global center coordinates of the sensor at the current moment. ;in, .
5. The method for local updating of a 3D point cloud map based on laser projection according to claim 4, characterized in that, Step S5 specifically includes the following steps: Step S51: For the global observation point set Each observation point in Construct a line from the lidar center Shot towards line segments; Step S52: Calculate the sequence of all voxel indices that the line segment passes through in the 3D mesh; Step S53: Update the logarithmic probability values of all voxels corresponding to the line segment to obtain the updated voxel probability map.
6. The method for local updating of a 3D point cloud map based on laser projection according to claim 5, characterized in that, Step S53 specifically refers to: All voxels that are passed through but not hit on the line segment path are considered "idle"; their logarithmic probability values are updated by deduction. ； in, The log-probability value before updating for all voxels that have passed through but not been hit along the line segment path; A negative value indicates an increased probability that the region is empty; Step S54: For a voxel hit at the end of a line segment, treat it as "occupied"; cumulatively update its logarithmic probability value: ； in, This is the log-probability value of the voxel hit at the end of the line segment before the update; If after each update, if Then set to ,like Then set to ; in, The maximum threshold; The minimum threshold.
7. The method for local updating of a 3D point cloud map based on laser projection according to claim 5, characterized in that, Step S6 specifically includes the following steps: Step S61: Traverse the updated voxel probability map; Step S62: Set a judgment threshold Read the log-probability value of each voxel in the updated voxel probability map. ; like The voxel was determined to be a valid obstacle. like The voxel is determined to be an empty region and discarded; Step S63: For voxels determined to be valid obstacles, based on the resolution parameters... Calculate its corresponding three-dimensional geometric coordinates The calculation formula is as follows: ； in Indicates the origin of the map. For voxel indexing; All recovered points are combined into an updated non-terrestrial point cloud set. .
8. The method for local updating of a 3D point cloud map based on laser projection according to claim 7, characterized in that, Step S7 specifically includes the following steps: Step S71: Read ground point cloud data from step S18 ; Step S72: Aggregate non-terrestrial point clouds ground point clouds Merge the points to obtain the merged point cloud. : ； Step S73: Process the merged point cloud Perform final data encapsulation and save the data to generate an updated map.
9. A device for local updating of a three-dimensional point cloud map based on laser projection, characterized in that, A method for local updating of a three-dimensional point cloud map based on laser projection according to any one of claims 1-8 includes an initialization module, an update module, and a reconstruction module; The initialization module is used to execute steps S1-S3; The update module is used to execute steps S4-S5; The reconstruction module is used to perform steps S6-S7.
10. A computer terminal storage medium, characterized in that, The device stores computer terminal executable instructions, characterized in that the computer terminal executable instructions are used to execute the laser projection-based three-dimensional point cloud map local update method as described in any one of claims 1-8.