Terrain mapping method and device based on uncertainty analysis, electronic equipment and storage medium

Abstract

The invention relates to a landform mapping method based on uncertainty analysis. The landform mapping method comprises the following steps: generating a landform candidate point set; obtaining an effective voxel set; obtaining a plane feature set; obtaining an octree hash map structure; and updating the structure of the octree hash map to complete topographic mapping. The invention further relates to a landform mapping device based on uncertainty analysis. The landform mapping device comprises an acquisition module, a first execution module, a second execution module, a processing module and an updating module. The invention further relates to electronic equipment and a storage medium. The invention provides a landform mapping method and device based on uncertainty analysis, electronic equipment and a storage medium. The problems that in a point cloud map, a plane representation map or a fixed-resolution voxel mapping technology, measurement noise and pose errors exist, construction and query expenses are large, stability is poor, sparse and dense areas are difficult to consider, balance is difficult to obtain in mapping, and the mapping efficiency is low can be solved. And the registration precision, the real-time performance and the map compactness of the three-dimensional environment mapping of the unmanned aerial vehicle are limited.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and specifically to a terrain mapping method, apparatus, electronic device, and storage medium based on uncertainty analysis. Background Technology

[0002] To meet the demands of autonomous navigation and surveying for real-time and high-precision 3D mapping, existing technologies typically employ point cloud maps, planar representation maps, or fixed-resolution voxel mapping. However, point cloud maps struggle to explicitly model map uncertainties, often treating local planes as deterministic quantities and ignoring measurement noise and pose errors. Furthermore, the construction and query overhead based on kd-trees increases with map size. Planar representation maps are highly sensitive to sparse / non-uniform point clouds, affecting stability. Fixed voxel mapping struggles to balance the representation of sparse and dense regions, finding it difficult to achieve a balance between coarse and fine detail when dealing with diverse structures such as ground, walls, and debris. These issues limit the registration accuracy, real-time performance, and map compactness of UAV 3D environment mapping.

[0003] Therefore, existing technologies have shortcomings and need to be improved and developed. Summary of the Invention

[0004] The present invention provides a terrain mapping method, apparatus, electronic device, and storage medium based on uncertainty analysis, which addresses the problems of existing technologies that typically employ point cloud maps, planar representation maps, or fixed-resolution voxel mapping. These problems include: point cloud maps struggle to explicitly model map uncertainties, often treating local planes as deterministic quantities and ignoring measurement noise and pose errors; the increasing overhead of kd-tree construction and querying as map size increases; the sensitivity of planar representation maps to sparse / non-uniform point clouds affecting stability; and the difficulty of achieving a balance between coarse and fine representation when dealing with diverse structures such as ground, walls, and debris. Furthermore, these problems limit the registration accuracy, real-time performance, and map compactness of UAV 3D environment mapping.

[0005] In a first aspect, the present invention provides a terrain mapping method based on uncertainty analysis, comprising: Raw point cloud data and initial pose estimation obtained by UAVs are collected to generate a set of candidate terrain points; Voxel partitioning is performed on the candidate terrain points to obtain an effective voxel set; Perform plane fitting on the effective voxel set to obtain a planar feature set; Process the planar feature set to obtain an octagonal hash map structure; New raw point cloud data acquired by UAVs and new initial pose estimates are collected to update the octagonal hash map structure and complete terrain mapping.

[0006] Furthermore, the process of collecting raw point cloud data and initial pose estimation from the UAV to generate a set of candidate terrain points includes: The raw point cloud data acquired by the radar mounted on the UAV and the initial pose estimate output by the state estimation module installed in the UAV are collected. Based on the initial pose estimation, the original point cloud data is transformed to the current global reference coordinate system to obtain transformed point cloud data; The converted point cloud data is sorted by height channel to filter terrain candidate points according to a preset minimum height percentage; All the selected terrain candidate points are combined to generate the terrain candidate point set.

[0007] Furthermore, the step of performing voxel partitioning on the terrain candidate point set to obtain an effective voxel set includes: The terrain candidate point set is divided into voxel blocks according to coarse resolution; Principal component analysis was performed on each of the voxel blocks to obtain the eigenvalues ​​of the covariance matrix; If the ratio of the smallest eigenvalue to the second largest eigenvalue in the covariance matrix does not decrease significantly, or the angle between the fitted plane obtained from the covariance matrix eigenvalues ​​and the horizontal plane is not less than a preset angle threshold, then the planarity is insufficient. The voxel block is then subdivided into smaller sub-voxel blocks, and principal component analysis and planarity checks are repeated for each sub-voxel block until the planarity is sufficient or the minimum voxel size is reached. If the ratio of the smallest eigenvalue to the second largest eigenvalue in the covariance matrix decreases significantly, and the angle between the fitted plane obtained from the covariance matrix eigenvalues ​​and the horizontal plane is less than a preset angle threshold, then the planarity is sufficient. All voxel blocks with sufficient planarity are retained to generate an effective voxel set.

[0008] Furthermore, performing plane fitting on the effective voxel set to obtain a planar feature set includes: Perform plane fitting on each point set in the effective voxel set to obtain the plane normal vector n, the plane center point c, and the plane covariance matrix Σ. Based on the plane normal vector n, the plane center point c, and the plane covariance matrix Σ, construct a triple (n, c, Σ). The planar feature set is obtained by combining the triples of each point set in the effective voxel set.

[0009] Furthermore, the process of processing the planar feature set to obtain the octagonal hash map structure includes: Each planar feature in the planar feature set is converted into an integer grid index according to its voxel space coordinates and a hash table Hash[(x, y, z)]=(n, q, Σ) is constructed, where x, y, and z are the x-coordinate, y-coordinate, and z-coordinate of the voxel space coordinates, respectively. An octree structure is used to manage the spatial hierarchy and obtain the octree hash map structure.

[0010] Furthermore, the acquisition of new raw point cloud data and new initial pose estimation obtained by the UAV to update the octagonal hash map structure and complete terrain mapping includes: The newly added original point cloud data is searched in the effective voxel set. If the newly added original point cloud data exists in the effective voxel set, the number of observations is accumulated and the newly added point cloud data is added to the effective voxel set. When the number of accumulated observations reaches the accumulated observation threshold, the original point cloud data existing in the effective voxel set is deleted. If the newly added original point cloud data does not exist in the effective voxel set, the newly added original point cloud data is subjected to plane fitting and then added to the plane feature set.

[0011] Furthermore, after completing the topographic mapping, it also includes: Collect current point cloud data acquired by the drone; Based on the current point cloud data, locate the voxel containing the current point cloud data in the octagonal hash map structure; Traverse the planar features contained within the voxel to be processed to match the current point cloud data and calculate the point-to-plane distance residual; An optimization equation is constructed by weighting the point-to-plane distance residuals according to their covariance. The optimization equation is solved iteratively to obtain the current pose estimate of the UAV under the current point cloud data; The current pose estimate is fed back to the UAV to optimize the UAV pose.

[0012] A second aspect of the present invention provides a terrain mapping apparatus based on uncertainty analysis, comprising: The acquisition module is used to acquire raw point cloud data and initial pose estimation obtained by the UAV in order to generate a set of candidate terrain points; The first execution module is used to perform voxel partitioning on the terrain candidate point set to obtain an effective voxel set; The second execution module is used to perform plane fitting on the effective voxel set to obtain a planar feature set; The processing module is used to process the planar feature set to obtain an octagonal hash map structure; The update module is used to collect new raw point cloud data acquired by the UAV and add initial pose estimation to update the octagonal hash map structure and complete terrain mapping.

[0013] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the aforementioned terrain mapping method based on uncertainty analysis.

[0014] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the terrain mapping method based on uncertainty analysis.

[0015] Beneficial effects: As can be seen from the above technical solutions, the present invention provides a terrain mapping method based on uncertainty analysis, which combines multi-source data processing from UAV-borne radar, voxelized spatial modeling, probabilistic plane feature representation, and an efficient indexing and online update mechanism, and has the following beneficial effects: 1. Introducing covariance modeling to quantify map uncertainty: In each effective voxel, while calculating the normal vector and center point through plane fitting, the plane parameter covariance matrix is ​​explicitly estimated. The covariance matrix comprehensively considers sources such as sensor ranging noise and attitude estimation error, giving each plane cell in the map a statistically meaningful confidence expression. This representation not only allows for dynamic adjustment of residual weights in registration optimization, reducing interference from low-confidence constraints, but also provides a reliability reference index for subsequent tasks (such as path planning and obstacle avoidance decisions).

[0016] 2. Adaptive Voxel Resolution for Enhanced Mapping Efficiency and Detail Retention: Principal Component Analysis (PCA) is employed for planarity detection combined with an angle threshold criterion. In flat areas with low information content, larger voxel sizes are maintained to reduce redundant computation. In structurally complex or detail-rich areas (such as building edges and around obstacles), the map is automatically subdivided into smaller voxels to capture key geometric information. This mechanism avoids the common problems of "over-refinement of sparse areas" and "loss of information in dense areas" in fixed-resolution voxel mapping, ensuring a balance between spatial resolution and computational cost.

[0017] 3. The "octree + hash table" combined structure enables efficient retrieval and management: the octree structure handles spatial hierarchical management and local subdivision control, while the hash table uses voxel coordinates as keys to achieve near constant-time neighborhood lookups and insertions. This dual-structure combination significantly improves retrieval efficiency when performing high-frequency operations such as point-to-area matching and map updates, reducing the computational overhead of traversal and comparisons compared to traditional kd-tree or pure octree solutions.

[0018] 4. Online incremental updates and redundancy removal maintain map compactness and timeliness: Each new frame of data only updates newly covered areas or voxels with significant changes in observation angle, avoiding the time consumption caused by full map reconstruction. For voxels where the number of observations reaches a threshold and the plane parameters converge, the original point data is deleted, and only the plane statistics are retained, achieving long-term compressed storage of map data. When significant changes in plane parameters are detected, local reconstruction is automatically triggered, effectively responding to dynamic environmental changes (such as the addition of obstacles or structural changes).

[0019] 5. Closed-loop registration optimization improves the accuracy of both localization and mapping: During the pose optimization stage, planar covariance is used to weight the point-to-surface residuals, reducing the impact of high uncertainty constraints and making the optimization results more stable. The optimized current pose is used as the initial pose input for the next frame of data processing, forming a "mapping-localization" closed loop, continuously improving the overall accuracy and robustness of the system. This closed-loop mechanism can maintain stable pose estimation and map updates even under adverse conditions such as multipath reflections, local occlusion, and sparse point clouds.

[0020] 6. Versatility and Availability Under Resource Constraints: The method can adaptively adjust voxel resolution and map update frequency in various scenarios, including low-altitude urban patrol, mountain mapping, and forest exploration. The map storage format retains only planar triplet parameters and covariance, significantly reducing storage space requirements compared to directly storing the original point cloud, making it suitable for UAV platforms with limited onboard computing resources. The efficient retrieval and update strategy enables real-time operation in embedded computing environments, without relying on high-performance ground computing stations.

[0021] 7. Providing directly usable uncertainty information for subsequent tasks: In addition to mapping and localization, the planar covariance matrix can be used as input to provide additional decision-making basis for path planning, obstacle avoidance strategies, and robustness assessment. The introduction of uncertainty information makes the map not only contain geometric information but also have a quantifiable reliability description, thereby improving the autonomy and safety of the entire UAV system in complex environments.

[0022] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below can be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other.

[0023] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0024] The accompanying drawings are not drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a terrain mapping method based on uncertainty analysis in an embodiment of this application.

[0025] Figure 2 This is a flowchart of step S102 of a terrain mapping method based on uncertainty analysis in an embodiment of this application.

[0026] Figure 3 This is a flowchart of step S104 of a terrain mapping method based on uncertainty analysis in an embodiment of this application.

[0027] Figure 4 This is a flowchart of step S106 of a terrain mapping method based on uncertainty analysis in an embodiment of this application.

[0028] Figure 5 This is a flowchart of step S108 of a terrain mapping method based on uncertainty analysis in an embodiment of this application.

[0029] Figure 6 This is a schematic diagram of terrain mapping of flat land using a terrain mapping method based on uncertainty analysis in an embodiment of this application.

[0030] Figure 7 This is a schematic diagram of terrain mapping of a slope using a terrain mapping method based on uncertainty analysis in an embodiment of this application. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art to which this invention pertains.

[0032] The terms "first," "second," and similar words used in the specification and claims of this patent application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, unless the context clearly indicates otherwise, the singular forms of "an," "a," or "the," etc., do not indicate a quantity limitation, but rather indicate the presence of at least one. Terms such as "comprising" or "including" indicate that the element or object preceding "comprising" encompasses the features, integrals, steps, operations, elements, and / or components listed following "comprising" or "including," and do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or sets thereof. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0033] Existing technologies typically employ point cloud maps, planar representation maps, or fixed-resolution voxel mapping. However, point cloud maps struggle to explicitly model map uncertainties, often treating local planes as deterministic quantities and ignoring measurement noise and pose errors. As map size increases, the construction and query overhead based on kd-trees rises. Planar representation maps are highly sensitive to sparse / non-uniform point clouds, affecting stability. Fixed voxel mapping struggles to balance the representation of sparse and dense regions, making it difficult to achieve a balance between coarse and fine detail when dealing with diverse structures such as ground, walls, and debris. These issues limit the registration accuracy, real-time performance, and map compactness of UAV 3D environment mapping.

[0034] In view of this, refer to Figure 1 This invention provides a terrain mapping method based on uncertainty analysis, comprising: Step S102: Collect raw point cloud data and initial pose estimation obtained by the UAV to generate a set of candidate terrain points.

[0035] Step S104: Perform voxel partitioning on the terrain candidate point set to obtain an effective voxel set.

[0036] Step S106: Perform plane fitting on the effective voxel set to obtain a planar feature set.

[0037] Step S108: Process the planar feature set to obtain the octagonal hash map structure.

[0038] Step S110: Collect newly added raw point cloud data and newly added initial pose estimation obtained by UAV to update the octagonal hash map structure and complete terrain mapping.

[0039] In some embodiments, reference is made to Figure 2The system collects raw point cloud data and initial pose estimations from UAVs to generate a set of candidate terrain points, including: Step S1021: Collect the raw point cloud data acquired by the radar mounted on the UAV and the initial pose estimate output by the state estimation module installed in the UAV.

[0040] When the UAV carrier acquires radar point clouds, the initial pose estimation is provided synchronously by the state estimation module. The state estimation module can be based on IMU / GNSS / visual odometry to output the initial pose.

[0041] Step S1022: Based on the initial pose estimation, transform the original point cloud data to the current global reference coordinate system. For example, use homogeneous transformation to transform the original point cloud data from the radar coordinate system to the current global system to obtain transformed point cloud data.

[0042] Step S1023: Sort the converted point cloud data by height channel to filter out terrain candidate points according to a preset minimum height percentage. For example, the preset minimum height percentage can be set to 5% to filter out terrain candidate points with a height percentage of not less than 5%.

[0043] In addition to filtering possible surface point sets based on the lowest elevation percentage, possible surface point sets can also be filtered based on elevation thresholds to remove non-topographic points that are significantly higher than the ground. Filtering based on elevation thresholds or lowest elevation percentages reduces computational load and provides a reference for subsequent planar positioning extraction.

[0044] Step S1024: Collect all the selected terrain candidate points to generate a terrain candidate point set.

[0045] Through steps S1021 to S1024, height percentile screening is completed, non-terrain points are eliminated as early as possible, reducing the burden on the voxelization and fitting stages; coordinate unification is completed in advance, eliminating subsequent cumulative errors caused by inconsistencies in multi-source coordinates; and a more uniform set of candidate points is generated, which is beneficial to the stability of subsequent planarity criteria.

[0046] In some embodiments, reference is made to Figure 3 Voxel partitioning is performed on the candidate terrain points to obtain an effective voxel set, including: Step S1041: Divide the terrain candidate point set into voxel blocks at a coarse resolution.

[0047] To adapt to the diversity of environmental structure and point cloud distribution, the area to be mapped is first divided into grids at a coarse resolution. For example, the three-dimensional space of the terrain is divided into voxel blocks at a coarse resolution, with each voxel block having a volume of 2 cubic meters.

[0048] Step S1042: Perform principal component analysis on each voxel block to obtain the eigenvalues ​​of the covariance matrix.

[0049] Within each voxel, the covariance matrix of the point set is calculated using principal component analysis (PCA), and its eigenvalues ​​are taken.

[0050] Step S1043: Determine the flatness. If the ratio of the smallest eigenvalue to the second largest eigenvalue in the covariance matrix does not decrease significantly, or the angle between the fitted plane obtained through the covariance matrix eigenvalues ​​and the horizontal plane is not less than a preset angle threshold (e.g., the preset angle threshold is 15°), then the flatness is insufficient. The voxel block is then subdivided into smaller sub-voxel blocks, and principal component analysis and flatness determination are repeated for each sub-voxel block until the flatness is sufficient or the minimum voxel size is reached. If the ratio of the smallest eigenvalue to the second largest eigenvalue in the covariance matrix decreases significantly, and the angle between the fitted plane obtained through the covariance matrix eigenvalues ​​and the horizontal plane is less than a preset angle threshold, then the flatness is sufficient. The method of judging flatness by whether it is significantly reduced or not means comparing the ratio of the smallest eigenvalue to the second largest eigenvalue with a predetermined threshold for judging flatness. If the ratio is less than the threshold, the ratio is not significantly reduced, and the flatness is significant. If the ratio is not less than the threshold, the ratio is significantly reduced, and the flatness is insufficient. Furthermore, the threshold for judging flatness depends on the sensor noise and the scene, and is generally set between 0.01 and 0.1. The smaller the value, the higher the required flatness. For example, a value of 0.01 indicates an extremely smooth plane, while a value of 0.1 allows for a certain degree of roughness.

[0051] If the ratio of the smallest eigenvalue to the second largest eigenvalue does not decrease significantly, or if the angle between the fitted plane and the horizontal plane exceeds a preset threshold (e.g., within 15° is considered near-horizontal terrain), then the point cloud within the voxel is considered not to meet the requirements of a large planar structure. In this case, the voxel is split into 1 / 8 octrees, recursively divided into 8 sub-voxels, and the above checks and subdivision operations are repeated for each sub-voxel until the planarity of the point cloud within the voxel is sufficient or the minimum voxel size (e.g., 0.2 m³) is reached.

[0052] Step S1044: Retain all voxel blocks with sufficient planarity to generate a valid voxel set. Finally, only voxel units that satisfy the planar characteristics are retained for terrain map modeling, thereby eliminating non-terrain data that could not be identified in step S102, where non-terrain data refers to planes that are not smaller than a preset included angle threshold (e.g., 15°).

[0053] Ultimately, only voxel units that satisfy planar features are retained for map modeling, thus avoiding over-refinement of low-information areas and improving mapping efficiency. The decision to perform octree recursive partitioning is based on the planarity of the point set, preventing over-partitioning of low-information areas and further improving mapping efficiency. Combining hash-based fast localization and octree spatial recursive partitioning improves query and update speed. Point cloud data is no longer retained on the convergence plane; only statistics are maintained, achieving compressed mapping. Tightly coupled pose optimization is performed using point-to-surface distance and covariance weighting.

[0054] Steps S1041 to S1044 complete the following steps: retaining large voxels in information-poor regions to reduce computation and storage; automatically subdividing in detailed regions to ensure local geometric representation; establishing the criterion for "whether to subdivide" on both statistical and geometric indicators to reduce misjudgments caused by uneven sampling; and the output "effective voxel set" provides a controllable input set for subsequent probability plane fitting.

[0055] In some embodiments, reference is made to Figure 4 Perform plane fitting on the effective voxel set to obtain a planar feature set, including: Step S1061: Perform plane fitting on each point set in the effective voxel set to obtain the plane normal vector n, the plane center point c, and the plane covariance matrix Σ.

[0056] Step S1062: Construct a triple (n, c, Σ) based on the plane normal vector n, the plane center point c, and the plane covariance matrix Σ.

[0057] Step S1063: Collect the triples of each point set in the effective voxel set to obtain the planar feature set.

[0058] For each valid voxel, a plane fitting is performed on the point set to obtain the plane normal vector n, the plane center point c, and the plane covariance matrix Σ. This is used to quantify the uncertainty of mapping, taking into account ranging error, attitude error, etc. Each plane is described by only a triple (n, c, Σ), and there is no need to save the original points.

[0059] For each valid set of voxels that passes the planeness test, a plane fitting is performed, such as using least squares fitting or RANSAC to suppress outliers, to obtain the normal vector and point cloud centroid, and the plane is described by combining this with the plane covariance matrix Σ. Unlike traditional deterministic representations, this invention calculates the uncertainty of the plane parameters, i.e., the plane covariance matrix Σ, by measuring the noise model and the current pose estimation error. For example, based on the propagation formula of laser ranging error and attitude error, the plane normal vector n, the plane center point c, and the plane covariance matrix Σ are derived. In this way, each plane feature is explicitly represented by triples, which can quantitatively characterize the uncertainty factors in the mapping process.

[0060] Steps S1061 to S1063 use the statistically quantified planar covariance matrix Σ as the weighting basis for subsequent registration, which can better express reliability than the "deterministic plane"; by discarding the original points and keeping only the triples, it is possible to form a compressed map storage; the planar uncertainty provides a threshold reference for subsequent change detection.

[0061] In some embodiments, reference is made to Figure 5 Processing planar feature sets to obtain an octagonal hash map structure, including: Step S1081: Convert each planar feature in the planar feature set into an integer grid index according to its voxel space coordinates and construct a hash table Hash[(x, y, z)]=(n, q, Σ), where x, y, and z are the x-coordinate, y-coordinate, and z-coordinate of the voxel space coordinates, respectively.

[0062] Step S1082: Use an octree structure to manage the spatial hierarchy and obtain an octree hash map structure.

[0063] All planar structures are organized into hash tables, with voxel coordinates as keys and planar triplet parameter descriptions as values; octree structures are used in conjunction with hash tables for local lookups and split control.

[0064] All constructed planar features are organized into a hash table according to their voxel spatial coordinates, while an octree structure is used to manage the spatial hierarchy. Specifically, the voxel's grid coordinates or octree node number are used as hash keys, and the corresponding planar parameters (planar normal vector n, planar center point c, and planar covariance matrix Σ) are stored as values. This hash index combined with an octree approach allows for querying, inserting, and updating neighborhood voxels in approximately constant time. Compared to traditional kd-trees, the octree hash structure is more efficient for continuous queries and dynamic insertions.

[0065] Steps S1081 to S1082 utilize a hash table in conjunction with an octree structure to achieve near constant-time location and neighborhood query, reducing the retrieval cost during the registration and update stages; hierarchical management facilitates local reconstruction and visualization, and provides a unified index for upper-level algorithms to be invoked.

[0066] In some embodiments, acquiring new raw point cloud data and new initial pose estimates from the UAV to update the octahedral hash map structure and complete terrain mapping includes: Step S1101: Search for newly added original point cloud data in the effective voxel set. If the newly added original point cloud data exists in the effective voxel set, accumulate the observation count and add the newly added point cloud data to the effective voxel set. When the accumulated observation count reaches the accumulated observation count threshold, delete the original point cloud data that exists in the effective voxel set. If the newly added original point cloud data does not exist in the effective voxel set, perform plane fitting on the newly added original point cloud data and add it to the plane feature set.

[0067] The radar scan frequency is 10 scans per second. After each radar scan, only voxels covering new areas or with significantly different observation angles are incrementally updated. When the cumulative number of observations of a certain planar feature reaches a threshold (e.g., 50 times), the original point is no longer retained to prevent map inflation.

[0068] For each new point cloud scan frame, the point cloud is first registered in the map coordinate system using the currently estimated UAV pose. For points falling into a new area without a corresponding voxel, a new voxel is created and an initial plane fit is performed. For points falling into an existing voxel, they are added to the voxel set, and the parameters and covariance of the plane within that voxel are updated in real time. As the number of observations increases, the uncertainty of the plane parameters converges rapidly, stabilizing after the cumulative number of observations reaches a threshold. Once the plane uncertainty converges, the historical original points within that voxel are deleted, retaining only the converged plane normal vector n, plane center point c, and plane covariance matrix Σ. This significantly reduces storage overhead and prevents the map from expanding over time. If newly added points in subsequent scans cause a large shift in the plane parameters, the structure of the area can be considered to have changed, triggering the re-division of the voxel and reconstruction of the plane features. Through incremental updates, the map remains up-to-date and compact, while eliminating redundant data, achieving real-time online maintenance.

[0069] Step S1101 controls map size and memory usage to reduce storage pressure during long-term operation; event-oriented updates driven by statistical convergence and change detection reduce invalid recalculations and maintain map freshness and availability in dynamic environments.

[0070] In some embodiments, after completing the terrain mapping, the method further includes: Collect current point cloud data acquired by the drone.

[0071] Based on the current point cloud data, locate the voxel containing the current point cloud data in the octagonal hash map structure.

[0072] Traverse the planar features contained within the voxel to be processed to match the current point cloud data and calculate the point-to-plane distance residual.

[0073] The point-to-plane distance residuals are weighted according to covariance to construct an optimization equation.

[0074] The optimization equation is solved iteratively to obtain the current pose estimate of the UAV under the current point cloud data.

[0075] Feedback of the current pose estimate to the UAV is provided to optimize the UAV pose.

[0076] By leveraging map information to optimize pose estimation for the current frame, the residual between the current frame point cloud and the plane in the map is calculated. Based on the plane covariance, a residual weight matrix is ​​constructed, explicitly considering mapping errors, and the state estimation module is given pose optimization updates to achieve high-precision positioning.

[0077] After a coarse positional transformation, the current frame's point cloud is finely registered with the constructed map's planar features. For each new point from the radar, the voxel containing that point is first quickly located in a hash octree, and then the planar features contained within that voxel are traversed for matching. During matching, the point-to-plane distance residual is calculated. Since the plane normal vector n and the plane center point c both have estimation uncertainties, the variance of the distance residual can be calculated. During optimization, the point-to-plane residual of each effective match is weighted according to its covariance, for example, using weighted least squares or Kalman filtering, to construct an optimization equation. By solving this optimization equation, the UAV pose and local planar parameters can be updated simultaneously, explicitly considering measurement noise and mapping errors, achieving high-precision registration results.

[0078] The above steps directly map the uncertainty of the map into optimization weights, reduce the impact of unreliable constraints, form a closed loop of "mapping and localization", improve the consistency of attitude estimation and reduce cumulative drift, and provide more stable pose input for subsequent planning and control modules.

[0079] Figure 6 and Figure 7 The following are schematic diagrams illustrating terrain mapping of flat land and sloping land using a terrain mapping method based on uncertainty analysis according to embodiments of this application: Figure 6 The image shows the mapping effect in a standardized farmland. The color of the voxel represents the z-axis height information of that location. The bottom is the farmland, and the top is the power line in the field. Figure 7 The image shows the mapping effect on a slope. The voxel color represents the z-axis height information of that location. The image shows the bow-shaped flight path of the UAV, which is an example of scanning and mapping directly on the slope.

[0080] Another embodiment of the present invention also provides a terrain mapping device based on uncertainty analysis, comprising: The acquisition module is used to acquire raw point cloud data and initial pose estimation obtained by the UAV in order to generate a set of candidate terrain points; The first execution module is used to perform voxel partitioning on the terrain candidate point set to obtain an effective voxel set; The second execution module is used to perform plane fitting on the effective voxel set to obtain a planar feature set; The processing module is used to process the planar feature set to obtain the octagonal hash map structure; The update module is used to collect new raw point cloud data acquired by UAVs and add initial pose estimates to update the octahedral map structure and complete terrain mapping.

[0081] Another embodiment of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements a terrain mapping method based on uncertainty analysis.

[0082] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the gateway's data processing device, connecting various parts of the gateway's data processing device through various interfaces and lines.

[0083] Another embodiment of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a terrain mapping method based on uncertainty analysis.

[0084] The memory may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created by the processor, etc. Furthermore, the memory is preferably, but not limited to, high-speed random access memory (RAM). For example, it may also be non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may also optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks (LANs), mobile communication networks, and combinations thereof.

[0085] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be accomplished by a computer program instructing related hardware, and can be stored in a computer-readable storage medium. When executed, the program can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium can also include combinations of the above types of memory.

[0086] In summary, this invention provides a terrain mapping method, apparatus, electronic device, and storage medium based on uncertainty analysis. It combines multi-source data processing from UAV-borne radar, voxelized spatial modeling, probabilistic plane feature representation, and an efficient indexing and online update mechanism. This allows for real-time positioning and navigation in UAV tasks such as low-altitude 3D environment mapping, path planning, obstacle avoidance, logistics delivery, agricultural flight, and urban perception. Therefore, the terrain mapping method, apparatus, electronic device, and storage medium based on uncertainty analysis provided by this invention have the following beneficial effects: Introducing covariance modeling to achieve quantitative expression of map uncertainty: In each effective voxel, while calculating the normal vector and center point through plane fitting, the plane parameter covariance matrix is ​​explicitly estimated. The covariance matrix comprehensively considers sources such as sensor ranging noise and attitude estimation error, giving each plane unit in the map a statistically meaningful confidence expression. This representation not only dynamically adjusts residual weights in registration optimization, reducing interference from low-confidence constraints, but also provides a reliability reference index for subsequent tasks (such as path planning and obstacle avoidance decisions). Adaptive voxel resolution enhances both mapping efficiency and detail: Principal Component Analysis (PCA) planeness detection combined with angle thresholding maintains larger voxel sizes in flat areas with low information content, reducing redundant computation. In structurally complex or detail-rich areas (such as building edges and around obstacles), it automatically subdivides into smaller voxels to capture key geometric information. This mechanism avoids the common problems of "over-refinement of sparse areas" and "information loss in dense areas" in fixed-resolution voxel mapping, ensuring a balance between spatial resolution and computational overhead. An octree + hash table combination enables efficient retrieval and management: The octree structure handles spatial hierarchical management and local subdivision control, while the hash table uses voxel coordinates as keys to achieve near-constant-time neighborhood lookups and insertions. This dual-structure combination significantly improves retrieval efficiency during high-frequency operations such as point-to-area matching and map updates, reducing the computational overhead of traversal and comparisons compared to traditional kd-tree or pure octree schemes. Online incremental updates and redundancy removal maintain map compactness and timeliness: Each new frame of data only updates newly covered areas or voxels with significant changes in observation angle, avoiding the time consumption of full map reconstruction. For voxels where the number of observations reaches a threshold and the plane parameters converge, the original point data is deleted, and only plane statistics are retained, achieving long-term compressed storage of map data. When significant changes in plane parameters are detected, local reconstruction is automatically triggered, effectively responding to dynamic environmental changes (such as the addition of obstacles or structural changes). Closed-loop registration optimization improves the bidirectional accuracy of localization and mapping: In the pose optimization stage, the plane covariance is used to weight the point-to-surface residuals, reducing the impact of high uncertainty constraints and making the optimization results more stable. The optimized current pose is used as the initial pose input for the next frame of data processing, forming a "mapping-localization" closed loop, continuously improving the overall accuracy and robustness of the system.This closed-loop mechanism maintains stable attitude estimation and map updates even under adverse conditions such as multipath reflection, local occlusion, and sparse point clouds. It is adaptable to various scenarios and usable under resource constraints: it can adaptively adjust voxel resolution and map update frequency in different scenarios such as low-altitude urban inspection, mountain mapping, and forest exploration. The map storage format only retains planar triplet parameters and covariance, significantly reducing storage space requirements compared to directly storing the original point cloud, making it suitable for UAV platforms with limited onboard computing resources. The efficient retrieval and update strategy enables this method to run in real-time in embedded computing environments without relying on high-performance ground computing stations. It provides directly usable uncertainty information for subsequent tasks: in addition to mapping and localization, the planar covariance matrix can be used as input, providing additional decision-making basis for path planning, obstacle avoidance strategies, and robustness assessment. The introduction of uncertainty information means that the map not only contains geometric information but also has a quantifiable reliability description, thereby improving the autonomy and safety of the entire UAV system in complex environments.

[0087] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method of topographic mapping based on uncertainty analysis, characterized by, The method comprises the following steps: Collecting original point cloud data and initial pose estimation obtained by a UAV to generate a terrain candidate point set; Performing voxel division on the terrain candidate point set to obtain an effective voxel set; Performing plane fitting on the effective voxel set to obtain a plane feature set; Processing the plane feature set to obtain an octree hash map structure; Collecting new original point cloud data and new initial pose estimation obtained by the UAV to update the octree hash map structure and complete terrain mapping.

2. The method of claim 1, wherein, The collecting of the original point cloud data and the initial pose estimation obtained by the UAV to generate the terrain candidate point set comprises the following steps: Collecting the original point cloud data obtained by a radar mounted on the UAV and the initial pose estimation output by a state estimation module installed in the UAV; Converting the original point cloud data to a current global reference coordinate system according to the initial pose estimation to obtain converted point cloud data; Sorting the converted point cloud data according to the height channel to filter out terrain candidate points according to a preset minimum height percentage; Collecting all the filtered terrain candidate points to generate the terrain candidate point set.

3. The method of claim 1, wherein, The performing of the voxel division on the terrain candidate point set to obtain the effective voxel set comprises the following steps: Dividing the terrain candidate point set into voxel blocks according to a coarse resolution; Performing principal component analysis on each voxel block to obtain eigenvalues of a covariance matrix; Determining the planarity, if the ratio of the minimum eigenvalue to the second largest eigenvalue in the covariance matrix eigenvalues is not significantly reduced, or the angle between the fitted plane obtained from the covariance matrix eigenvalues and the horizontal plane is not less than a preset angle threshold, the planarity is insufficient, and the voxel block is subdivided into smaller sub-voxel blocks, and the principal component analysis and the planarity determination are repeatedly performed on each sub-voxel block until the planarity is sufficient or the minimum voxel size is reached; if the ratio of the minimum eigenvalue to the second largest eigenvalue in the covariance matrix eigenvalues is significantly reduced, and the angle between the fitted plane obtained from the covariance matrix eigenvalues and the horizontal plane is less than the preset angle threshold, the planarity is sufficient; Retaining all the voxel blocks with sufficient planarity to generate the effective voxel set.

4. The method of claim 1, wherein, The performing of the plane fitting on the effective voxel set to obtain the plane feature set comprises the following steps: Performing plane fitting on each point set in the effective voxel set to obtain a plane normal vector n, a plane center point c and a plane covariance matrix Σ; Constructing a triple (n, c, Σ) according to the plane normal vector n, the plane center point c and the plane covariance matrix Σ; Collecting the triple of each point set in the effective voxel set to obtain the plane feature set.

5. The method of mapping terrain based on uncertainty analysis according to claim 4, wherein, The processing of the plane feature set to obtain the octree hash map structure comprises the following steps: Converting each plane feature in the plane feature set to an integer grid index according to the voxel space coordinates thereof and constructing a hash table Hash[(x, y, z)]=(n, q, Σ), wherein x, y and z are respectively the horizontal coordinate, the vertical coordinate and the vertical coordinate of the voxel space coordinates; Managing the spatial hierarchy by using an octree structure and obtaining the octree hash map structure.

6. The method of claim 1, wherein, The collection of the new original point cloud data and the new initial pose estimation obtained by the unmanned aerial vehicle is used to update the octree hash map structure, and the terrain mapping is completed, including: The new original point cloud data is searched in the effective voxel set, if the new original point cloud data exists in the effective voxel set, the cumulative observation number is accumulated and the new point cloud data is added in the effective voxel set, when the cumulative observation number reaches the cumulative observation number threshold, the original point cloud data existing in the effective voxel set is deleted, if the new original point cloud data does not exist in the effective voxel set, the plane fitting is performed on the new original point cloud data and is added in the plane feature set.

7. The method of claim 1, wherein, After the terrain mapping is completed, further including: The current point cloud data obtained by the unmanned aerial vehicle is collected; The to-be-processed voxel containing the current point cloud data is located in the octree hash map structure according to the current point cloud data; The plane feature contained in the to-be-processed voxel is traversed to match the current point cloud data and calculate the point-to-plane distance residual; The point-to-plane distance residual is weighted according to the covariance to construct an optimization equation; The optimization equation is iteratively solved to obtain the current pose estimation of the unmanned aerial vehicle under the current point cloud data; The current pose estimation is fed back to the unmanned aerial vehicle to optimize the unmanned aerial vehicle pose.

8. A terrain mapping device based on uncertainty analysis, characterized by, Including: A collection module is used to collect the original point cloud data and the initial pose estimation obtained by the unmanned aerial vehicle to generate a terrain candidate point set; A first execution module is used to perform voxel division on the terrain candidate point set to obtain an effective voxel set; A second execution module is used to perform plane fitting on the effective voxel set to obtain a plane feature set; A processing module is used to process the plane feature set to obtain an octree hash map structure; An update module is used to collect the new original point cloud data and the new initial pose estimation obtained by the unmanned aerial vehicle to update the octree hash map structure and complete the terrain mapping.

9. An electronic device, comprising: The computer program is executed by the processor to implement the terrain mapping method based on uncertainty analysis according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to implement the terrain mapping method based on uncertainty analysis according to any one of claims 1-7.

Images