A method for air-ground collaborative mapping and point cloud fusion in GPS-denied environments

Abstract

This application discloses a method, apparatus, medium, and device for air-ground collaborative mapping and point cloud fusion in GPS-denied environments. By controlling air and ground nodes to generate local point cloud maps respectively, after statistical denoising and voxel downsampling, the optimal transformation matrix is ​​obtained by using viewpoint feature compensation coarse registration and singular value decomposition fine registration, thereby achieving seamless stitching and global fusion of heterogeneous air and ground maps.

Description

Technical Field

[0001] This application relates to the field of multi-robot collaborative mapping and environmental perception technology, and in particular to a method, apparatus, medium and device for air-ground collaborative mapping point cloud fusion in GPS denied environments. Background Technology

[0002] In applications such as autonomous driving, complex indoor mapping, mine exploration, and disaster search and rescue, constructing high-precision 3D environmental maps is crucial for achieving autonomous robot navigation. Traditional mapping methods heavily rely on global positioning information provided by the Global Positioning System (GPS). However, in typical GPS-denied scenarios such as urban canyons, indoor environments, and underground parking garages, satellite signals are blocked or interfered with by physical obstacles. This causes traditional SLAM algorithms based on single robots to generate significant cumulative drift, making it difficult to construct a consistent global map.

[0003] Existing research mostly employs single mobile nodes for mapping. Aerial robots (such as drones) offer a wide field of view but have limited endurance and insufficient perception of obstructed areas from their overhead perspective; ground robots (such as unmanned vehicles) have large payloads and generate dense, high-precision radar data, but their field of view is limited and their mobility is poor. How to combine the advantages of both perspectives to quickly and accurately perform high-precision spatial alignment and fusion of sparse point clouds collected by heterogeneous aerial nodes and dense point clouds collected by ground nodes, without global positioning signals, is a pressing technical challenge in this field.

[0004] Traditional point cloud registration methods often face challenges when processing heterogeneous air-to-ground data, such as large initial pose differences, inconsistent point cloud density, and interference from non-overlapping regions, leading to slow convergence or even failure. Therefore, a robust two-stage registration framework and feature enhancement mechanism are needed to achieve rapid air-to-ground collaborative mapping under GPS-denied conditions. Summary of the Invention

[0005] The main objective of this application is to provide a method, apparatus, medium, and device for air-ground collaborative mapping point cloud fusion in GPS-denied environments, aiming to solve the technical problems of positioning drift caused by GPS signal loss and the difficulty of fusion due to large differences in perspective between air and ground heterogeneous data.

[0006] To achieve the above objectives, this application provides a method for air-to-ground collaborative mapping and point cloud fusion in GPS-denied environments, comprising: In a GPS denied environment, control the airborne mobile node and the ground mobile node to start the local SLAM process respectively to generate airborne local point cloud maps and ground local point cloud maps. Based on the characteristic that the aerial mobile nodes are prone to motion artifacts, statistical noise reduction is performed on the aerial local point cloud map; based on the dense characteristic of the ground mobile node data, voxel downsampling is performed on the ground local point cloud map to obtain local point cloud data after consistency characterization processing, so as to balance the spatial distribution density of heterogeneous air and ground data. Based on the point cloud data after the consistency characterization process, coarse registration based on view feature compensation is used to lock the initial spatial mapping, and sub-pixel level fine registration based on singular value decomposition is used to obtain the optimal rotation and translation matrix. The optimal rotation and translation matrix is ​​used to map the local point cloud map in the air to the ground reference coordinate system, and the seamless stitching and global fusion of the heterogeneous local maps in the air and on the ground are realized through an incremental update mechanism.

[0007] Optionally, in a GPS denied environment, control the airborne mobile node and the ground mobile node to start local SLAM processes respectively to generate airborne local point cloud maps and ground local point cloud maps; Consistency characterization processing steps: Based on the characteristic that the aerial mobile nodes are prone to motion artifacts, statistical noise reduction is performed on the aerial local point cloud map; based on the dense characteristic of the ground mobile node data, voxel downsampling is performed on the ground local point cloud map to obtain local point cloud data after consistency characterization processing, so as to balance the spatial distribution density of heterogeneous air and ground data. Based on the point cloud data after the consistency characterization process, coarse registration based on view feature compensation is used to lock the initial spatial mapping, and sub-pixel level fine registration based on singular value decomposition is used to obtain the optimal rotation and translation matrix. The optimal rotation and translation matrix is ​​used to map the local point cloud map in the air to the ground reference coordinate system, and the seamless stitching and global fusion of the heterogeneous local maps in the air and on the ground are realized through an incremental update mechanism.

[0008] Optionally, the statistical denoising of the aerial local point cloud map includes: Calculate the average distance from each point in the local point cloud map to its k-neighbor points, assuming that the average distance follows a Gaussian distribution; Points that meet the outlier criteria are removed to eliminate dynamic ghost noise caused by the flight jitter of aerial nodes, providing a clean geometric topology for subsequent feature extraction; wherein, the outlier criteria are: the average distance from a point to its k-neighbors is greater than the sum of the products of the mean average distance and the standard deviation multiple threshold.

[0009] Optionally, the voxel downsampling of the local ground point cloud map includes: Divide the local point cloud map of the ground into a voxel grid of a preset size; Calculate the centroid coordinates of the point set within each voxel grid, and use the centroid coordinates as the representative point of that voxel grid to generate a density-equalized local point cloud map of the ground. By replacing the original dense point set with the centroid coordinates, the weight of ground data in the error function of subsequent registration algorithms is reduced.

[0010] Optionally, the step of using coarse registration based on viewpoint feature compensation to lock the initial spatial mapping includes: Fast feature histogram (FPFH) descriptors for extracting local point cloud maps from the air and ground; The Random Sample Consensus (RANSAC) algorithm is used to find the optimal initial transformation matrix between two sets of point clouds; The objective of solving the optimal initial transformation matrix is ​​to minimize the feature descriptor residuals of the sampling point pairs, and to satisfy that the number of corresponding point pairs is greater than a preset threshold. The rotation invariance of the FPFH descriptor is used to compensate for the huge deviation of the air-ground perspective, and the RANSAC algorithm is used to lock the initial spatial mapping in the presence of a large number of non-overlapping regions.

[0011] Optionally, obtaining the optimal rotation and translation matrix by combining sub-pixel-level fine registration based on singular value decomposition includes: Within the overlapping region determined by coarse registration, construct an objective function that minimizes the sum of squared Euclidean residuals; Calculate the cross-covariance matrix of the corresponding point set, and perform singular value decomposition on the cross-covariance matrix to obtain the rotation matrix and translation vector; The optimal rotation matrix is ​​determined by the product of the right singular vector and the left singular vector obtained from the singular value decomposition. Based on the initial pose provided by coarse registration, the optimal rotation matrix is ​​directly solved by SVD analytical solution, avoiding the local minima problem in the iteration process and achieving pose transformation with sub-pixel accuracy.

[0012] Optionally, the step of achieving seamless stitching and global fusion of heterogeneous local maps via incremental update mechanism includes: The optimal rotation and translation matrix is ​​used to perform coordinate transformation on a local point cloud map in the air. Perform a Boolean union operation between the transformed aerial local point cloud map and the ground local point cloud map, and update the global 3D map. When the airborne or ground-based mobile node continues to collect new local point cloud data, the steps of processing the local point cloud data after consistency characterization and obtaining the optimal rotation and translation matrix based on sub-pixel-level fine registration using singular value decomposition are repeated to stitch the new local point cloud map incrementally into the global 3D map. By continuously updating incrementally, the topological consistency of the global map is maintained in GPS-denied environments, eliminating perception blind spots from a single robot's perspective.

[0013] To achieve the above objectives, a second aspect of this application also provides an air-to-ground collaborative mapping point cloud fusion device for GPS-denied environments, comprising: The parallel construction module is used to control the airborne mobile node and the ground mobile node to start the local SLAM process respectively in the GPS denied environment, and generate the airborne local point cloud map and the ground local point cloud map. The consistency characterization module is used for the consistency characterization processing steps: based on the characteristic that the aerial mobile nodes are prone to motion artifacts, statistical noise reduction is performed on the aerial local point cloud map; based on the dense characteristic of the ground mobile node data, voxel downsampling is performed on the ground local point cloud map to obtain the local point cloud data after consistency characterization processing, so as to balance the spatial distribution density of heterogeneous air and ground data. The secondary registration module is used to lock the initial spatial mapping based on the point cloud data after the consistency representation processing by adopting coarse registration based on view feature compensation, and obtain the optimal rotation and translation matrix by combining sub-pixel level fine registration based on singular value decomposition. The incremental update module is used to map the local point cloud map in the air to the ground reference coordinate system using the optimal rotation and translation matrix, and to achieve seamless stitching and global fusion of heterogeneous local maps between the air and the ground through the incremental update mechanism.

[0014] To achieve the above objectives, a third aspect of this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the program, it implements the air-ground collaborative mapping point cloud fusion method for GPS denied environments as described in the first aspect.

[0015] To achieve the above objectives, the fourth aspect of this application also provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when executed by a processor, the program implements the air-ground collaborative mapping point cloud fusion method for GPS-denied environments as described in the first aspect.

[0016] This application discloses a method for air-ground collaborative mapping and point cloud fusion in GPS-denied environments. The method includes: in a GPS-denied environment, controlling an airborne mobile node and a ground mobile node to initiate local SLAM processes respectively, generating an airborne local point cloud map and a ground local point cloud map; a consistency characterization processing step: based on the characteristic that the airborne mobile node is prone to motion artifacts, statistical noise reduction is performed on the airborne local point cloud map; based on the dense data of the ground mobile node, voxel downsampling is performed on the ground local point cloud map to obtain consistent characterization processed local point cloud data, thus balancing the spatial distribution density of heterogeneous air-ground data; based on the consistent characterization processed point cloud data, coarse registration based on viewpoint feature compensation is used to lock the initial spatial mapping, combined with sub-pixel level fine registration based on singular value decomposition to obtain the optimal rotation and translation matrix; using the optimal rotation and translation matrix, the airborne local point cloud map is mapped to the ground reference coordinate system, and seamless stitching and global fusion of heterogeneous air-ground local maps are achieved through an incremental update mechanism. This application effectively eliminates the perception blind spots of a single robot's viewpoint and significantly improves the mapping completeness and real-time performance in GPS-denied environments. Attached Figure Description

[0017] Figure 1 This application provides an embodiment of an air-ground collaborative rapid mapping system architecture for GPS-denied environments. Figure 2 This is a schematic diagram of a local sparse point cloud generated by an aerial node in an embodiment of this application; Figure 3 This is a schematic diagram of a locally dense point cloud generated by ground nodes in an embodiment of this application; Figure 4 This is the final globally consistent 3D map with viewpoint compensation capability generated in the embodiments of this application; Figure 5 A flowchart illustrating the method provided in this application embodiment.

[0018] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit this application.

[0020] Aerial local point cloud map: refers to a set of three-dimensional point cloud data generated by an aerial mobile node through a local SLAM process in a GPS-denied environment.

[0021] Local point cloud map: refers to a set of three-dimensional point cloud data generated by a ground mobile node through a local SLAM process in a GPS-denied environment.

[0022] Consistent characterization processing: refers to the process of performing statistical noise reduction and voxel downsampling on the characteristics of heterogeneous data in the air and on the ground to balance the spatial distribution density.

[0023] Optimal rotation and translation matrix: refers to the transformation matrix obtained through coarse and fine registration, used to map a local point cloud map in the air to a ground reference coordinate system.

[0024] Incremental update mechanism: refers to the update method that repeats the processing steps and stitches the new local point cloud map into the global 3D map when the node continues to collect new data.

[0025] Existing technologies for GPS-denied environments often employ single mobile node mapping. While aerial robots offer a wide field of view, they suffer from blind spots and are prone to motion artifacts. Ground robots, on the other hand, have dense data but limited perspectives. Traditional point cloud registration methods face challenges when processing heterogeneous air-ground data, including significant initial pose differences, inconsistent point cloud density, and interference from non-overlapping regions, leading to slow convergence or even failure.

[0026] Example 1 refer to Figure 1 The first embodiment of this application provides a point cloud fusion method for air-ground collaborative mapping in GPS-denied environments, to solve the technical problems of positioning drift caused by GPS signal loss and the large differences in perspective between air and ground heterogeneous data, making fusion difficult, as mentioned in the background art. This method can be executed by a processor, which can be located in a terminal or a server. The execution process of the method is as follows: Step S101: In a GPS denied environment, control the airborne mobile node and the ground mobile node to start the local SLAM process respectively to generate an airborne local point cloud map and a ground local point cloud map.

[0027] Step S102: Based on the characteristic that aerial mobile nodes are prone to motion artifacts, statistical noise reduction is performed on the aerial local point cloud map; based on the dense characteristic of ground mobile node data, voxel downsampling is performed on the ground local point cloud map to obtain local point cloud data after consistent characterization processing, so as to balance the spatial distribution density of heterogeneous air and ground data.

[0028] Step S103: Based on the point cloud data after consistency characterization processing, coarse registration based on view feature compensation is used to lock the initial spatial mapping, and sub-pixel level fine registration based on singular value decomposition is used to obtain the optimal rotation and translation matrix.

[0029] Step S104: The aerial local point cloud map is mapped to the ground reference coordinate system using the optimal rotation and translation matrix, and the seamless stitching and global fusion of the heterogeneous aerial and ground local maps are achieved through an incremental update mechanism.

[0030] In one embodiment of this application, step S102, which involves statistical denoising of the aerial local point cloud map, includes: Calculate the average distance from each point in the local point cloud map to its k-neighbors, assuming that the average distance follows a Gaussian distribution; Points that meet the outlier criteria are removed to eliminate dynamic ghost noise caused by the flight jitter of aerial nodes, providing a clean geometric topology for subsequent feature extraction. The outlier criteria are: the average distance from a point to its k-neighbors is greater than the sum of the products of the mean average distance and the standard deviation multiple threshold.

[0031] Further calculation Each point in the middle to its The average distance of each neighboring point Assuming it follows a Gaussian distribution, remove outliers that satisfy the following formula:

[0032] in, The average distance is the mean. Standard deviation, This is the threshold value for multiples of the standard deviation.

[0033] In one embodiment of this application, step S102, which involves voxel downsampling of the local ground point cloud map, includes: Divide the local point cloud map of the ground into a voxel grid of a preset size; Calculate the centroid coordinates of the point set within each voxel grid, use the centroid coordinates as the representative point of that voxel grid, and generate a density-equalized local point cloud map of the ground. By replacing the original dense point set with centroid coordinates, the weight of ground data in the error function of subsequent registration algorithms is reduced.

[0034] Furthermore, for dense point clouds on the ground... Perform voxel downsampling, assuming the voxel grid size is... The set of points within the grid is Then the centroid of the grid The calculation formula is:

[0035] In one embodiment of this application, step S103 employs coarse registration based on viewpoint feature compensation to lock the initial spatial mapping, including: Fast feature histogram (FPFH) descriptors for extracting local point cloud maps from the air and ground; The Random Sample Consensus (RANSAC) algorithm is used to find the optimal initial transformation matrix between two sets of point clouds; The objective of solving the optimal initial transformation matrix is ​​to minimize the feature descriptor residuals of the sampling point pairs, while satisfying that the number of corresponding point pairs is greater than a preset threshold. The rotation invariance of the FPFH descriptor is used to compensate for the huge deviation of the air-ground perspective, and the RANSAC algorithm is used to lock the initial spatial mapping in the presence of a large number of non-overlapping regions.

[0036] Furthermore, the Random Sample Consensus (RANSAC) algorithm is employed. and Find the optimal initial transformation matrix between The objective function is:

[0037] in, The number of sample point pairs, and the feature descriptor residuals of the corresponding point pairs.

[0038] Minimum.

[0039] Using the Random Sample Consensus (RANSAC) algorithm, in and Find the optimal initial transformation matrix between The objective function is:

[0040] in, The number of sample point pairs, and the feature descriptor residuals of the corresponding point pairs. Minimum.

[0041] In one embodiment of this application, step S103, which combines sub-pixel-level fine registration based on singular value decomposition to obtain the optimal rotation and translation matrix, includes: Within the overlapping region determined by coarse registration, construct an objective function that minimizes the sum of squared Euclidean residuals; Calculate the cross-covariance matrix of the corresponding point set, and perform singular value decomposition on the cross-covariance matrix to obtain the rotation matrix and translation vector; The optimal rotation matrix is ​​determined by the product of the right singular vector and the left singular vector obtained from the singular value decomposition. Based on the initial pose provided by coarse registration, the optimal rotation matrix is ​​directly solved by SVD analytical solution, avoiding the local minima problem in the iteration process and achieving pose transformation with sub-pixel accuracy.

[0042] Furthermore, within the overlapping region determined by coarse registration, fine alignment is performed using the Iterative Closest Point (ICP) algorithm. This is achieved by minimizing the sum of squared Euclidean residuals:

[0043] in, This represents the total number of point pairs in the overlapping area.

[0044] The optimal rotation is solved using singular value decomposition (SVD). First, the cross-covariance matrix is ​​constructed. :

[0045] in, and These are the centroid coordinates of the corresponding point set. SVD decomposition yields Then the optimal rotation matrix and translation vector They are respectively:

[0046]

[0047] In one embodiment of this application, step S104 achieves seamless stitching and global fusion of heterogeneous local maps between air and ground through an incremental update mechanism, including: The optimal rotation and translation matrix is ​​used to perform coordinate transformation on a local point cloud map in the air. Perform a Boolean union operation between the transformed aerial local point cloud map and the ground local point cloud map, and update the global 3D map. When the airborne or ground-based mobile nodes continue to collect new local point cloud data, the steps of processing the local point cloud data after consistency characterization and obtaining the optimal rotation and translation matrix by sub-pixel-level fine registration based on singular value decomposition are repeated to stitch the new local point cloud map incrementally into the global 3D map. By continuously updating incrementally, the topological consistency of the global map is maintained in GPS-denied environments, eliminating perception blind spots from a single robot's perspective.

[0048] Furthermore, using the obtained and The aerial local map is mapped to the ground reference coordinate system, and the global map is updated through a Boolean union operation:

[0049] For example, in an indoor garage environment where GPS is denied, an aerial robot (drone) and a ground robot (unmanned vehicle) work collaboratively. The drone is equipped with an RGBD camera and achieves fast SLAM through loop closure detection based on a bag-of-words model. Due to the limited computing power of the drone, the point cloud map it generates is relatively sparse and contains noise caused by flight jitter.

[0050] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and mathematical logic.

[0051] In an indoor garage environment where GPS is denied, an aerial robot (drone) and a ground robot (unmanned vehicle) work collaboratively. The drone, equipped with an RGBD camera, achieves fast SLAM through loop closure detection based on a bag-of-words model. Due to the limited computing power of the drone, the generated point cloud map is relatively sparse and contains noise caused by flight jitter.

[0052] According to step 2 of the present invention, the system first performs statistical filtering on the UAV point cloud as shown in formula (1), where the parameters are... The value is set to 2.0, effectively eliminating dynamic ghost noise during flight. Subsequently, the system performs voxel downsampling on the millions of point clouds generated by the unmanned vehicle using formula (2), reducing the grid side length. Setting it to 0.05m reduces the computational complexity of subsequent registration while preserving key features such as wall corners and columns.

[0053] In the coarse registration stage of step 3, the system uses the FPFH descriptor to establish local geometric associations. Since the UAV has a top-down view and the unmanned vehicle has a level view, the overlapping area is concentrated on the ground and some pillars. The RANSAC algorithm iteratively selects overlapping feature points and calculates the initial rotation matrix, reducing the relative rotation angle of the air-ground map to within 5 degrees.

[0054] In the fine registration stage of step 4, the system initiates ICP iteration based on the coarse registration. By calculating the covariance matrix of formula (5) and performing singular value decomposition of formula (6), the system can solve for the translation parameters with sub-pixel accuracy within milliseconds. Experiments show that after the fusion processing of the method of this invention, the point cloud stitching overlap rate is improved by more than 15% compared with the traditional method.

[0055] Finally, the two types of local mappings are merged into a global mapping using formula (8). For example... Figure 5 As shown, the final generated map includes both the top-down panoramic view from the drone's perspective (eliminating blind spots for the unmanned vehicle) and the high-precision elevation information from the unmanned vehicle's perspective, achieving a dual improvement in map completeness and construction speed in GPS-denied environments.

[0056] Example 2 refer to Figure 3 The second embodiment of this application provides an air-ground collaborative mapping point cloud fusion device for GPS-denied environments to solve the same technical problem as the method embodiment. The device 1000 may include the following modules: Parallel construction module 1001 is used to control the airborne mobile node and the ground mobile node to start local SLAM processes respectively in a GPS denied environment, and generate airborne local point cloud maps and ground local point cloud maps. The consistency characterization module 1002 is used for the consistency characterization processing steps: based on the characteristic that aerial mobile nodes are prone to motion artifacts, statistical noise reduction is performed on the aerial local point cloud map; based on the characteristic of dense ground mobile node data, voxel downsampling is performed on the ground local point cloud map to obtain the local point cloud data after consistency characterization processing, so as to balance the spatial distribution density of heterogeneous air and ground data. The secondary registration module 1003 is used to lock the initial spatial mapping based on the point cloud data after consistency characterization processing. It adopts coarse registration based on view feature compensation and obtains the optimal rotation and translation matrix by combining sub-pixel level fine registration based on singular value decomposition. The incremental update module 1004 is used to map the local point cloud map in the air to the ground reference coordinate system using the optimal rotation and translation matrix, and to achieve seamless stitching and global fusion of heterogeneous local maps between the air and the ground through the incremental update mechanism.

[0057] Among them, the parallel construction module 1001 is used to control the airborne mobile node and the ground mobile node to start the local SLAM process respectively in the GPS denied environment, and generate the airborne local point cloud map and the ground local point cloud map.

[0058] The consistency characterization module 1002 is used to perform statistical noise reduction on the local point cloud map in the air based on the characteristic that the air mobile node is prone to motion artifacts; and to perform voxel downsampling on the local point cloud map in the ground based on the characteristic that the ground mobile node data is dense, so as to obtain the local point cloud data after consistency characterization processing, so as to balance the spatial distribution density of heterogeneous air and ground data.

[0059] The secondary registration module 1003 is used for point cloud data after consistency characterization processing. It uses coarse registration based on view feature compensation to lock the initial spatial mapping, and combines it with sub-pixel level fine registration based on singular value decomposition to obtain the optimal rotation and translation matrix.

[0060] The incremental update module 1004 is used to map the local point cloud map in the air to the ground reference coordinate system using the optimal rotation and translation matrix, and to achieve seamless stitching and global fusion of heterogeneous local maps between air and ground through the incremental update mechanism.

[0061] It is understood that this embodiment is a device claim corresponding to the method claim, and therefore this embodiment has all the technical effects of the method claim, which will not be repeated here.

[0062] Example 3 The third embodiment of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The feature is that when the processor executes the program, it implements the air-ground collaborative mapping point cloud fusion method for GPS denied environments provided in the previous embodiments.

[0063] The fourth embodiment of this application provides a computer-readable storage medium storing a computer program, characterized in that, when executed by a processor, the program implements the air-to-ground collaborative mapping point cloud fusion method for GPS-denied environments provided in the preceding embodiments. The above description is merely a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for air-ground collaborative mapping and point cloud fusion in GPS-denied environments, characterized in that, include: In a GPS denied environment, control the airborne mobile node and the ground mobile node to start the local SLAM process respectively to generate airborne local point cloud maps and ground local point cloud maps. Based on the characteristic that the aerial mobile nodes are prone to motion artifacts, statistical noise reduction is performed on the aerial local point cloud map; based on the dense characteristic of the ground mobile node data, voxel downsampling is performed on the ground local point cloud map to obtain local point cloud data after consistency characterization processing, so as to balance the spatial distribution density of heterogeneous air and ground data. Based on the point cloud data after the consistency characterization process, coarse registration based on view feature compensation is used to lock the initial spatial mapping, and sub-pixel level fine registration based on singular value decomposition is used to obtain the optimal rotation and translation matrix. The optimal rotation and translation matrix is ​​used to map the local point cloud map in the air to the ground reference coordinate system, and the seamless stitching and global fusion of the heterogeneous local maps in the air and on the ground are realized through an incremental update mechanism.

2. The method for air-to-ground collaborative mapping and point cloud fusion in GPS-denied environments according to claim 1, characterized in that, The statistical denoising of the aerial local point cloud map includes: Calculate the average distance from each point in the local point cloud map to its k-neighbor points, assuming that the average distance follows a Gaussian distribution; Points that meet the outlier criteria are removed to eliminate dynamic ghost noise caused by the flight jitter of aerial nodes, providing a clean geometric topology for subsequent feature extraction; wherein, the outlier criteria are: the average distance from a point to its k-neighbors is greater than the sum of the products of the mean average distance and the standard deviation multiple threshold.

3. The method for air-to-ground collaborative mapping and point cloud fusion in GPS-denied environments according to claim 1, characterized in that, The voxel downsampling of the local point cloud map on the ground includes: Divide the local point cloud map of the ground into a voxel grid of a preset size; Calculate the centroid coordinates of the point set within each voxel grid, and use the centroid coordinates as the representative point of that voxel grid to generate a density-equalized local point cloud map of the ground. By replacing the original dense point set with the centroid coordinates, the weight of ground data in the error function of subsequent registration algorithms is reduced.

4. The method for air-to-ground collaborative mapping and point cloud fusion in GPS-denied environments according to claim 1, characterized in that, The method of using coarse registration to lock the initial spatial mapping based on viewpoint feature compensation includes: Fast feature histogram (FPFH) descriptors for extracting local point cloud maps from the air and ground; The Random Sample Consensus (RANSAC) algorithm is used to find the optimal initial transformation matrix between two sets of point clouds. The objective of solving the optimal initial transformation matrix is ​​to minimize the feature descriptor residuals of the sampling point pairs, and to satisfy that the number of corresponding point pairs is greater than a preset threshold. The rotation invariance of the FPFH descriptor is used to compensate for the huge deviation of the air-ground perspective, and the RANSAC algorithm is used to lock the initial spatial mapping in the presence of a large number of non-overlapping regions.

5. The method for air-to-ground collaborative mapping and point cloud fusion in GPS-denied environments according to claim 1, characterized in that, The method of obtaining the optimal rotation and translation matrix by combining sub-pixel-level fine registration based on singular value decomposition includes: Within the overlapping region determined by coarse registration, construct an objective function that minimizes the sum of squared Euclidean residuals; Calculate the cross-covariance matrix of the corresponding point set, and perform singular value decomposition on the cross-covariance matrix to obtain the rotation matrix and translation vector; The optimal rotation matrix is ​​determined by the product of the right singular vector and the left singular vector obtained from the singular value decomposition. Based on the initial pose provided by coarse registration, the optimal rotation matrix is ​​directly solved by SVD analytical solution, avoiding the local minima problem in the iteration process and achieving pose transformation with sub-pixel accuracy.

6. The method for air-to-ground collaborative mapping and point cloud fusion in GPS-denied environments according to claim 1, characterized in that, The method of achieving seamless stitching and global fusion of heterogeneous local maps between air and ground through incremental update mechanism includes: The optimal rotation and translation matrix is ​​used to perform coordinate transformation on a local point cloud map in the air. Perform a Boolean union operation between the transformed aerial local point cloud map and the ground local point cloud map, and update the global 3D map. When the airborne or ground-based mobile node continues to collect new local point cloud data, the steps of processing the local point cloud data after consistency characterization and obtaining the optimal rotation and translation matrix based on sub-pixel-level fine registration using singular value decomposition are repeated to stitch the new local point cloud map incrementally into the global 3D map. By continuously updating incrementally, the topological consistency of the global map is maintained in GPS-denied environments, eliminating perception blind spots from a single robot's perspective.

7. A point cloud fusion device for air-to-ground collaborative mapping in GPS-denied environments, characterized in that, include: The parallel construction module is used to control the airborne mobile node and the ground mobile node to start the local SLAM process respectively in the GPS denied environment, and generate the airborne local point cloud map and the ground local point cloud map. The consistency characterization module is used for the consistency characterization processing steps: based on the characteristic that the aerial mobile nodes are prone to motion artifacts, statistical noise reduction is performed on the aerial local point cloud map; based on the dense characteristic of the ground mobile node data, voxel downsampling is performed on the ground local point cloud map to obtain the local point cloud data after consistency characterization processing, so as to balance the spatial distribution density of heterogeneous air and ground data. The secondary registration module is used to lock the initial spatial mapping based on the point cloud data after the consistency representation processing by adopting coarse registration based on view feature compensation, and obtain the optimal rotation and translation matrix by combining sub-pixel level fine registration based on singular value decomposition. The incremental update module is used to map the local point cloud map in the air to the ground reference coordinate system using the optimal rotation and translation matrix, and to achieve seamless stitching and global fusion of heterogeneous local maps between the air and the ground through the incremental update mechanism.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the air-ground collaborative mapping point cloud fusion method for GPS denied environments as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the air-ground collaborative mapping point cloud fusion method for GPS denied environments as described in any one of claims 1 to 6.

Images