A high-precision unmanned aerial vehicle surveying and mapping data processing method and system

Abstract

This application relates to a high-precision UAV mapping data processing method and system. The data processing method includes: acquiring raw monitoring data of the UAV during its flight in the survey area and performing spatiotemporal synchronization processing to generate an enhanced dataset; calculating terrain complexity parameters and environmental dynamic parameters, and fusing them to generate an environmental index; dynamically partitioning the survey area and performing point cloud preprocessing to obtain a preprocessed point cloud set; calculating dynamic voxel parameters and optimizing the point cloud spatial distribution of the preprocessed point cloud set to generate an adaptive point cloud model and reconstructing its topology to generate an enhanced point cloud model; fusing the enhanced point cloud model with image data to perform semantically constrained 3D reconstruction and outputting a 3D model; extracting point cloud curvature features and generating an encrypted verification point cloud; when the elevation deviation between the encrypted verification point cloud and the 3D model exceeds a deviation threshold, feedback is provided to adjust the dynamic voxel parameters and re-optimize the point cloud spatial distribution. This application improves the reliability of the mapping results.

Description

Technical Field

[0001] This application relates to the field of UAV remote sensing and mapping technology, and in particular to a high-precision UAV mapping data processing method and system. Background Technology

[0002] With the rapid development of UAV platforms and sensor technologies, UAV mapping has become an important means of acquiring high spatiotemporal resolution surface information. Currently, utilizing UAVs equipped with optical cameras and LiDAR for collaborative operations, and generating realistic 3D models through oblique photogrammetry and laser scanning, has become the mainstream technical approach in the industry. A typical processing flow usually includes steps such as simultaneous acquisition of multi-source data, joint aerial triangulation adjustment of point clouds and images, dense matching to generate 3D point clouds, point cloud filtering and classification, construction of triangulation networks (TINs), texture mapping, and model refinement. This technical system greatly improves the efficiency and automation level of 3D data acquisition.

[0003] However, in practical, large-scale, and complex environments, traditional methods often neglect the assessment of the impact of external dynamic disturbances (such as wind speed changes, temperature fluctuations, and light intensity) on the measurement process. They lack effective adaptive adjustment mechanisms to address modeling quality issues arising from differences in mapping difficulty across different regions. Furthermore, the lack of perception and adaptive adjustment capabilities for dynamic environmental disturbances and scene heterogeneity, as well as the lack of effective integration of multi-source heterogeneous data, leads to irreversible damage to structural information in certain sensitive areas. This affects the authenticity and completeness of the final mapping results, making it difficult to meet the needs of in-depth geographic information mining. Summary of the Invention

[0004] To improve the automation level of the surveying process and the controllability of the output quality, this application provides a high-precision UAV surveying data processing method and system.

[0005] Firstly, this application provides a high-precision UAV mapping data processing method, which adopts the following technical solution:

[0006] A high-precision UAV mapping data processing method, the data processing method comprising:

[0007] Acquire raw monitoring data in real time during the flight of the UAV in the survey area, including image data, laser point cloud data, attitude data, positioning data and environmental monitoring data;

[0008] The original monitoring data is subjected to spatiotemporal synchronization processing to generate a spatiotemporally aligned enhanced dataset;

[0009] Based on the enhanced dataset, terrain complexity parameters and environmental dynamic parameters are calculated and fused to generate an environmental index;

[0010] The survey area is dynamically partitioned according to the environmental index, and partition point cloud preprocessing is performed to obtain a preprocessed point cloud set with partition annotations.

[0011] Dynamic voxel parameters are calculated by combining the environmental index with the environmental monitoring data;

[0012] The point cloud spatial distribution of the preprocessed point cloud set is optimized by applying the dynamic voxel parameters and the pre-configured airflow disturbance compensation model to generate an adaptive point cloud model.

[0013] The adaptive point cloud model is reconstructed in topology to generate an enhanced point cloud model with topology.

[0014] The enhanced point cloud model and the image data are fused together to perform semantically constrained 3D reconstruction, and a 3D model with semantic annotation is output.

[0015] Based on the adaptive point cloud model, point cloud curvature features are extracted, and encrypted verification point cloud is generated. The elevation deviation between the encrypted verification point cloud and the three-dimensional model is calculated.

[0016] When the elevation deviation exceeds the preset deviation threshold, feedback is provided to adjust the dynamic voxel parameters and re-optimize the point cloud spatial distribution.

[0017] By adopting the above technical solution, the inevitable trend of modern remote sensing and mapping technology towards automation and intelligence is fully demonstrated. This solution not only effectively solves the problem of integrating multi-source heterogeneous data, but also incorporates various advanced algorithm tools and engineering practice experience, successfully connecting every key node from bottom-level sensing and acquisition to top-level visualization. In particular, by introducing an environment-aware, driven adaptive control mechanism and a multi-level feedback optimization strategy, the overall performance indicators of the final results in terms of spatial resolution, geometric fidelity, and semantic richness are improved. This effectively addresses complex terrain and environmental interference, reduces the cost of manual intervention, and enhances the reliability and practicality of surveying and mapping results.

[0018] Secondly, this application provides a high-precision UAV mapping data processing system, which adopts the following technical solution:

[0019] A high-precision UAV mapping data processing system, the data processing system comprising:

[0020] The multi-source data acquisition module is used to acquire raw monitoring data collected in real time during the flight of the UAV in the test area, including image data, laser point cloud data, attitude data, positioning data and environmental monitoring data;

[0021] The spatiotemporal synchronization processing module is used to perform spatiotemporal synchronization processing on the original monitoring data to generate a spatiotemporally aligned enhanced dataset;

[0022] The environmental index calculation module is used to calculate terrain complexity parameters and environmental dynamic parameters based on the enhanced dataset, and then fuse them to generate an environmental index.

[0023] The dynamic partitioning preprocessing module is used to dynamically partition the test area according to the environmental index and perform partition point cloud preprocessing to obtain a preprocessed point cloud set with partition annotations.

[0024] The dynamic voxel parameter calculation module calculates dynamic voxel parameters by combining the environmental index with the environmental monitoring data.

[0025] The point cloud spatial optimization module is used to optimize the point cloud spatial distribution of the preprocessed point cloud set by applying the dynamic voxel parameters and the pre-configured airflow disturbance compensation model, and generate an adaptive point cloud model.

[0026] The topology reconstruction module is used to reconstruct the topology of the adaptive point cloud model to generate an enhanced point cloud model with topology.

[0027] The semantic 3D reconstruction module is used to fuse the enhanced point cloud model with the image data to perform semantically constrained 3D reconstruction and output a 3D model with semantic annotations.

[0028] The accuracy verification and evaluation module is used to extract point cloud curvature features based on the adaptive point cloud model, generate encrypted verification point cloud, and calculate the elevation deviation between the encrypted verification point cloud and the three-dimensional model.

[0029] The parameter feedback optimization module is used to adjust the dynamic voxel parameters and re-optimize the point cloud spatial distribution when the elevation deviation exceeds a preset deviation threshold.

[0030] In summary, this application includes at least one of the following beneficial technical effects: It quantifies the real-time disturbances and terrain complexity of the mapping scene through environmental indices, and drives dynamic partitioning preprocessing and spatial distribution optimization accordingly, achieving refined and differentiated adaptation of processing strategies; it completes high-fidelity 3D reconstruction under semantic constraints, and forms a feedback optimization closed loop from output results to front-end key parameters (dynamic voxel parameters) through deviation analysis of encrypted verification point clouds and the final model based on curvature features. This application effectively overcomes the limitations of single-parameter processing in complex environments, improves the overall quality of 3D models in terms of geometric accuracy, semantic integrity, and texture realism, while reducing the cost of manual intervention and enhancing the system's robustness and automation level for different operating scenarios. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the first process of a high-precision UAV mapping data processing method according to one embodiment of this application.

[0032] Figure 2 This is a schematic diagram of the second process of a high-precision UAV mapping data processing method according to one embodiment of this application.

[0033] Figure 3 This is a schematic diagram of the third process of a high-precision UAV mapping data processing method according to one embodiment of this application.

[0034] Figure 4 This is a schematic diagram of the fourth process of a high-precision UAV mapping data processing method according to one embodiment of this application.

[0035] Figure 5 This is a schematic diagram of the fifth step of a high-precision UAV mapping data processing method according to one embodiment of this application.

[0036] Figure 6 This is a schematic diagram of the sixth process of a high-precision UAV mapping data processing method according to one embodiment of this application.

[0037] Figure 7 This is a schematic diagram of the seventh process of a high-precision UAV mapping data processing method according to one embodiment of this application. Detailed Implementation

[0038] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-7 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0039] This application discloses a high-precision UAV mapping data processing method.

[0040] Reference Figure 1 A high-precision UAV mapping data processing method, specifically including:

[0041] Step S101: Obtain raw monitoring data collected in real time during the flight of the UAV in the test area, including image data, laser point cloud data, attitude data, positioning data and environmental monitoring data;

[0042] Image data typically comes from optical cameras or RGB-D devices, providing color and texture information of ground surfaces; laser point cloud data comes from LiDAR (Light Detection and Ranging) devices, which can accurately measure the distance to the target object and its three-dimensional geometry; attitude data records the aircraft's attitude angle changes (such as pitch, roll, and yaw), which is the basis for subsequent coordinate transformations; positioning data is generally obtained from GNSS / INS integrated navigation systems, providing position information in the geographic coordinate reference system corresponding to each frame of data; and environmental monitoring data may include meteorological elements such as temperature, humidity, and wind speed. Although they do not directly participate in 3D modeling, they can be used as auxiliary variables affecting point cloud quality and imaging stability in subsequent stages.

[0043] Step S102: Perform spatiotemporal synchronization processing on the original monitoring data to generate a spatiotemporally aligned enhanced dataset;

[0044] Specifically, multiple sensors can be triggered to start acquiring data at the same time using the IEEE 1588 precision clock protocol to ensure consistent start time. For high-frequency acquired attitude data, interpolation is used to extend it to a time resolution that matches other low-frequency signals, such as interpolating at 200Hz to obtain a continuous attitude sequence. In addition, the influence of external environmental factors must be considered, such as temperature fluctuations that may cause changes in hardware response delay. Therefore, a timestamp correction matrix with temperature compensation function is introduced to correct potential errors.

[0045] After this series of processing steps, the data from all sources are unified under a shared time frame and have the same spatial reference standard, forming a so-called "augmented dataset," which is a set of data that has been calibrated in both time and space dimensions, making it easier to conduct cross-modal association analysis later.

[0046] Step S103: Calculate terrain complexity parameters and environmental dynamic parameters based on the enhanced dataset, and fuse them to generate an environmental index;

[0047] The main purpose of this step is to quantify the overall surveying difficulty level of the current work area. The terrain complexity parameter reflects the degree of topographic relief and can be calculated by measuring the elevation gradient within a local area. The magnitude of the amplitude is used to estimate the data. The larger the value, the steeper and more rugged the terrain. The corresponding data collection is more susceptible to obstruction and interference, resulting in a decrease in effective coverage. The environmental dynamic parameters mainly include the intensity of wind disturbance and the interference of moving objects. The former is reflected in the swaying effect of the aircraft caused by airflow, while the latter refers to the noise pollution risk caused by pedestrians, vehicles and other moving targets in the survey area.

[0048] In this embodiment of the application, in order to comprehensively evaluate the influence weights of the above two types of uncertain factors, a weighted summation form of the environmental index formula is proposed:

[0049] Where the coefficients α, β, and γ represent the importance ratios of terrain factor, wind speed factor, and moving target factor, respectively, and can be flexibly set according to the actual application scenario, σ wind Let V be the wind speed variance. obj This index estimates the speed of a moving object. A higher index indicates more severe external conditions, requiring a more conservative and robust data processing strategy.

[0050] Step S104: Dynamically partition the survey area according to the environmental index, and perform partition point cloud preprocessing to obtain the preprocessed point cloud set of partition annotation;

[0051] The system automatically divides the environment into several relatively independent sub-regions based on the previously obtained environmental indices, and formulates differentiated processing plans for each type of region. This approach avoids the risk of damaging key details in sensitive areas through a one-size-fits-all approach of large-scale filtering or noise reduction.

[0052] For example, in locations with rugged terrain, dense vegetation, or high population density, the denoising intensity should be appropriately reduced to avoid accidentally deleting real feature edge information; while in flat, open areas with predominantly static backgrounds, the compression ratio can be increased to improve computational efficiency. Based on this, further routine point cloud cleaning tasks such as denoising, filtering, and outlier removal are performed, ultimately producing a set of pre-processed point clouds labeled by category for use in the next stage of parameter tuning.

[0053] Step S105: Calculate dynamic voxel parameters by combining environmental indices and environmental monitoring data;

[0054] Among them, traditional fixed-size voxel grids are difficult to take into account the visual scale distortion caused by the difference in depth of field, which can easily lead to the problem of excessive sparseness in the near-field area and excessive crowding in the far-field area.

[0055] In this embodiment of the application, a function is designed. To express how the side length of a voxel changes with height H and slope. The parameters are adjusted accordingly based on the changing trends of three key variables: instantaneous wind speed W. This parameter setting allows for a higher density near the ground to capture fine structural outlines, while allowing for a more relaxed spacing at higher altitudes to conserve memory resources. Furthermore, considering that aerodynamic characteristics indirectly affect the stability of the laser beam propagation path, a wind speed term is added to correct for this when determining the optimal granularity, thereby mitigating scanning drift caused by turbulence.

[0056] Step S106: Apply dynamic voxel parameters and a pre-configured airflow disturbance compensation model to optimize the spatial distribution of the preprocessed point cloud set and generate an adaptive point cloud model.

[0057] The airflow disturbance compensation model is a set of mathematical mapping relationships derived from physical simulation. It can predict the possible deflection angle and displacement magnitude based on the current measured wind field state, and adjust the estimated value of the point cloud coordinates accordingly to make them approach the ideal state. Combined with the previously calculated dynamic voxel parameters, it can adjust the node density distribution within individuals while maintaining the overall topology consistency, so that the originally disordered discrete sampling points gradually tend to be arranged in a regular and orderly manner, thereby greatly improving the quality performance of subsequent mesh generation.

[0058] It should be noted that the optimization here is not a simple and crude resampling, but a soft normalization achieved by applying nonlinear deformation to the existing observation samples, which preserves the authenticity of the original observations and enhances the robustness of the model.

[0059] Step S107: Reconstruct the topology of the adaptive point cloud model to generate an enhanced point cloud model with topology.

[0060] Specifically, the topology reconstruction steps include: constructing a triangular network using ICP registration in the high-frequency region; using feature descriptor matching in the low-frequency region; and deleting invalid edges with a side length exceeding 3 times the voxel size.

[0061] This step aims to restore the proper adjacency network between point clouds. Although the preceding steps have completed the initial sorting, blind spots and occlusion breaks are unavoidable in the original acquisition process, causing some adjacent points to lose their correct connection paths.

[0062] In this embodiment, the above two complementary methods can be used for repair: on the one hand, in areas with drastic terrain undulations and dense point clouds, the ICP (Iterative Closest Point) algorithm is used to establish a local triangulation network; on the other hand, in relatively flat areas with few features, the SIFT / SURF invariant feature descriptor matching method is used to find potential correspondences. Finally, by setting a reasonable maximum allowable side length threshold (such as three times the average voxel size), those obviously unreasonable artificially constructed connections are filtered out to ensure that the reconstructed map can reflect the real terrain trend and maintain good geometric stability.

[0063] Step S108: Integrate the enhanced point cloud model and image data to perform semantically constrained 3D reconstruction, and output a 3D model with semantic annotations.

[0064] Semantic constraints involve adapting deep neural network models commonly used in image recognition to aerial photographs. This involves extracting segmentation mask layers for various typical landform types, such as building rooftops, road boundaries, and vegetation cover. This high-level prior knowledge is then injected into the underlying mesh construction process, forcibly constraining vertices of specific semantic categories to remain within their planar domain. Simultaneously, the mapping error must be strictly controlled to not exceed one pixel to prevent stretching, deformation, tearing, or even damage. This not only significantly improves the model's realism but also endows it with rich attribute labels to support subsequent GIS analysis and data mining needs.

[0065] Step S109: Extract the curvature features of the point cloud based on the adaptive point cloud model, generate an encrypted verification point cloud, and calculate the elevation deviation between the encrypted verification point cloud and the three-dimensional model.

[0066] In order to verify whether the aforementioned series of processing has achieved the expected accuracy target, local areas with high surface curvature (curvature greater than 0.8) are selected as key verification objects, and more densified check points are set up around them to enhance statistical representativeness. These newly added points do not participate in the formal modeling process, but are only used to compare and check whether there is a systematic deviation in the final product.

[0067] Specifically, a temporary topological basis is reconstructed according to the original voxel division principle. Then, the vertical distance difference between the encrypted point cloud and the 3D model is compared to see if it exceeds the allowable tolerance range. Once an excess is found, the error correction procedure is immediately initiated.

[0068] Step S110: When the elevation deviation exceeds the preset deviation threshold, feedback is provided to adjust the dynamic voxel parameters and re-optimize the point cloud spatial distribution.

[0069] Specifically, after calculating the set of elevation deviation values ​​between the "encrypted verification point cloud" and the final output "3D model" through spatial location registration, the system does not merely treat it as a static accuracy report, but rather as a dynamic diagnostic signal reflecting whether the "dynamic voxel parameter" settings in the current processing pipeline are suitable for the current survey area environment and terrain features. When any of these elevation deviation values ​​exceed the "preset deviation threshold" (e.g., centimeter or millimeter tolerances set according to surveying grade requirements), it indicates that in the corresponding spatial area (usually a highly dynamic area with complex terrain and high curvature), the "point cloud spatial distribution optimization" guided by the current dynamic voxel parameters calculated based on height, slope, and wind speed has not achieved optimal results, potentially leading to the loss or deformation of local geometric details, which in turn can transmit unacceptable errors to the final model.

[0070] At this point, the system initiates a feedback adjustment process. Its core logic is to establish a reverse correlation model from the end-point "elevation deviation" to the source-point "dynamic voxel parameters." The system analyzes the spatial distribution, magnitude, and direction of the deviation exceeding limits, and correlates these areas with the corresponding original environmental indices, environmental monitoring data (such as instantaneous wind speed and point cloud density), and curvature distribution maps. Based on these correlation analyses, the system derives correction terms for the "dynamic voxel parameter" calculation function. For example, if deviations are found to be concentrated in steep slope areas under high wind speeds, it may be determined that the compensation coefficient for wind speed W in the current parameter function is insufficient, or that the voxel side length's contraction response with slope S is not agile enough. Therefore, the system will fine-tune the calculation logic of the dynamic voxel parameters by reducing the voxel size in the target area (i.e., increasing local sampling density) or adjusting the wind speed disturbance compensation weight. This adjustment is not globally uniform but rather specifically applied to problem areas identified through environmental indices and curvature characteristics.

[0071] Finally, the adjusted dynamic voxel parameters are re-inputted into the "point cloud spatial distribution optimization" stage. At this point, the system doesn't start entirely from scratch. Instead, based on the "preprocessed point cloud set," new parameters and an "airflow disturbance compensation model" are applied to perform a new, more targeted optimization of the point cloud's spatial distribution, generating an updated "adaptive point cloud model." This optimization process rebalances the density and regularity of the point cloud, aiming to more accurately resist major interference sources identified in areas exceeding deviation limits (such as strong winds and turbulence, and complex terrain occlusion). Subsequently, the entire downstream process, including topology reconstruction and semantically constrained 3D reconstruction, will be re-executed or partially updated based on this optimized point cloud model, and accuracy will be verified again, forming a closed loop of "perception-analysis-adjustment-verification." The technical significance of this mechanism lies in its ability to give the entire data processing system self-learning and adaptive capabilities. It can automatically approach the optimal processing parameter configuration for different tasks and environmental conditions through iterative feedback, thereby continuously ensuring and improving the reconstruction accuracy and reliability of the final 3D model under various complex real-world conditions without the need for repeated manual trial and error intervention.

[0072] The above implementation fully embodies the inevitable trend of modern remote sensing and mapping technology towards automation and intelligence. This technical solution not only effectively solves the problem of integrating multi-source heterogeneous data, but also incorporates various advanced algorithm tools and engineering practice experience, successfully connecting every key node from bottom-level sensing and acquisition to top-level visualization. In particular, by introducing an environment-aware, driven adaptive control mechanism and a multi-level feedback optimization strategy, it improves the comprehensive performance indicators of the final results in terms of spatial resolution, geometric fidelity, and semantic richness. This effectively addresses complex terrain and environmental interference, reduces the cost of manual intervention, and enhances the reliability and practicality of surveying and mapping results.

[0073] Reference Figure 2 As one implementation of step S104, the steps of dynamically partitioning the survey area according to the environmental index and performing partition point cloud preprocessing to obtain the preprocessed point cloud set with partition annotations include:

[0074] Step S201: Obtain point cloud data from the environmental index and augmented dataset;

[0075] The core objective of this step is to establish the foundational input conditions for subsequent processing. The environmental index is a comprehensive quantitative indicator reflecting the complexity or drastic changes in terrain and features within the target survey area. For example, in densely built-up urban areas, the environmental index is higher due to numerous vertical structures and shading effects, while it is relatively lower in open plains. This index serves as the basis for intelligent decision-making, guiding the application of different data processing strategies for different regions.

[0076] Meanwhile, the point cloud data in the augmented dataset refers to the data set after preliminary registration, noise reduction or other basic corrections. It not only contains the original spatial three-dimensional coordinate information (X,Y,Z), but may also include various attributes such as color, intensity, and timestamp. These data constitute the core operation objects in the entire process.

[0077] Step S202: Generate a partition mapping map based on the environmental index to characterize the processing level corresponding to each coordinate point in the test area;

[0078] In essence, the zoning mapping map is a geospatial index structure composed of two-dimensional raster maps or polygonal vector area features. Each pixel or geometric unit corresponds to an average environmental index within a specific range, and is assigned a corresponding processing level label (such as high frequency / low frequency). This mapping mechanism achieves an effective transformation from abstract numerical values ​​to visualized spatial divisions, enabling the system to quickly identify which areas require more refined data processing methods.

[0079] In this embodiment, if the environmental index of most sampling points in a certain area exceeds a set threshold of 0.8, it is determined to be a "high-frequency processing area," meaning that the area has high terrain undulation, richness of detail, or importance, and therefore should be processed with higher fidelity. Conversely, it is classified as a "low-frequency processing area," indicating that the computational overhead can be appropriately reduced without affecting the overall quality of the results. In this way, large-scale point cloud data of uniform specifications can be decomposed into several small-scale task modules with clear functional orientations as needed.

[0080] Step S203: Divide the point cloud data into multiple point cloud subsets according to the partition mapping map, with each subset corresponding to a processing level region;

[0081] During this process, the system combines spatial topological relationships with the aforementioned mapping results to perform cropping and classification operations on the original point cloud. Each generated subset of the point cloud represents a local geographic region and all measurement points within it, and they share the same processing priority attribute.

[0082] Specifically, to efficiently accomplish this task, spatial indexing algorithms such as KD-trees and octrees are often used to accelerate the retrieval and matching process, ensuring that each point can be accurately assigned to its respective partition category. It should be noted that although there may be some overlap between subsets, they generally maintain mutual exclusion to allow for independent parameter tuning. This divide-and-conquer strategy significantly improves the system's flexibility and resource utilization, avoiding the efficiency bottleneck caused by a single, high-load full computation.

[0083] Step S204: Perform differentiated preprocessing operations on point cloud subsets of regions with different processing levels;

[0084] The differentiation is mainly reflected in two aspects: first, selecting algorithm combinations with strong adaptability for different types of regions; and second, adjusting the configuration of key parameters to meet their respective accuracy requirements.

[0085] For the subset marked as the high-frequency processing area, since it carries greater information value, a more precise but also more time-consuming voxel filtering technique is adopted for downsampling and smoothing optimization. A typical approach is to construct a three-dimensional grid in units of 10 centimeters and find the centroid of the points falling into the same grid to replace the original scattered distribution, thereby achieving the purpose of compressing redundancy and improving uniformity.

[0086] For low-frequency processing areas, a statistical outlier removal method is applied. This method identifies and removes isolated noise samples in abnormally sparse locations by modeling the probability distribution of the number of neighboring points. This method is computationally inexpensive and suitable for coarse-grained filtering operations in large-area flat terrain scenarios. Although the two strategies have different approaches, they both achieve optimal performance in their respective applicable domains, demonstrating the advantages of intelligent scheduling.

[0087] Step S205: Merge the preprocessed point cloud subsets and embed the partition type identifier to generate a preprocessed point cloud set with partition annotation.

[0088] After each subset undergoes a customized cleaning and reconstruction process, it must ultimately be reintegrated into a globally consistent data framework. At this point, not only must the original spatial continuity and integrity be restored, but additional metadata tags related to the type of their source region must also be injected.

[0089] To address this, a standard 32-bit format label field is introduced, which includes, but is not limited to, two-digit encoding representing the processing level (e.g., 01 for high frequency, 02 for low frequency), with the remaining digits used to record boundary numbers, acquisition times, and other related auxiliary information. This additional descriptor greatly facilitates downstream applications' understanding and utilization of the received data, especially important in application scenarios involving multi-level modeling, progressive rendering, or multi-user collaborative editing.

[0090] The above implementation fully integrates spatial perception capabilities and adaptive control mechanisms, establishing a complete closed-loop control system based on scientifically sound hierarchical standards. This technical solution overcomes the problem of balancing efficiency and quality in traditional single-mode approaches, while effectively avoiding the economic burden risks associated with blindly expanding investment in high-performance computing power.

[0091] Reference Figure 3 As one implementation of step S105, the step of calculating dynamic voxel parameters by combining environmental indices and environmental monitoring data includes:

[0092] Step S301: Obtain wind speed data from environmental indices and environmental monitoring data;

[0093] Among them, environmental monitoring data refers to external physical conditions recorded in real time or near real time, which may include meteorological elements such as temperature, humidity, and wind speed. In particular, wind speed data has a significant impact on the quality of point cloud data.

[0094] Step S302: Analyze the terrain complexity parameter in the environmental index;

[0095] Specifically, the elevation gradient-related components in the environmental index are extracted, and these components are mapped to terrain complexity levels using a piecewise function. Essentially, this step extracts key discriminant factors affecting the difficulty of the surveying task from the static environmental background, thus providing a basis for further refined control.

[0096] Step S303: Determine the basic voxel size based on the terrain complexity parameter;

[0097] Specifically, when the terrain complexity component is greater than 0.8, the basic voxel size is set to 0.1 meters; when the terrain complexity component is less than or equal to 0.8, the basic voxel size is set to 0.5 meters.

[0098] In this context, a voxel refers to a basic sampling unit in three-dimensional space, similar to a two-dimensional pixel. Its size directly affects the final point cloud density and its corresponding storage overhead and processing performance. This application proposes the idea of ​​dynamically setting the voxel granularity based on terrain complexity. Specifically, when the terrain is flatter and features are sparser, the scale of a single voxel can be appropriately increased to reduce resource waste caused by redundant sampling; conversely, it should be reduced to a finer level to capture more detailed features.

[0099] It should be noted that the judgment threshold here (e.g., 0.8) is not arbitrarily chosen, but rather an empirical constant derived from extensive experiments. This constant indicates that when the elevation variation rate within a certain region reaches a critical level, a higher resolution mode should be used for coverage. Essentially, this method is an optimal resource allocation strategy oriented towards target quality constraints, maximizing equipment efficiency while ensuring the accuracy of the results.

[0100] Step S304: Acquire the drone's flight altitude data in real time, and calculate the airflow disturbance compensation coefficient based on the flight altitude data and wind speed data;

[0101] Specifically, the calculation of the airflow disturbance compensation coefficient includes: reading real-time wind speed values ​​and historical wind speed variances; calculating the altitude attenuation factor based on flight altitude; and fusing the wind speed values ​​and altitude attenuation factor using an exponential function. The calculation model for the altitude attenuation factor is as follows: H represents the flight altitude.

[0102] Understandably, this step is one of the proactive intervention measures taken to address external dynamic disturbances during flight operations. Because UAVs are often subjected to uncontrollable natural forces such as crosswinds and turbulence when performing aerial survey missions, this can lead to attitude deviations and even trajectory deviations, indirectly causing accumulated drift errors in the coordinates of the data acquisition points. To mitigate the impact of such problems, it is necessary to establish a reasonable compensation mechanism to correct for transient disturbances caused by airflow.

[0103] Against this backdrop, the system combines the measured wind speed at the current moment with the standard deviation information obtained from historical statistical data, further considering the vertical wind shear phenomenon caused by flight altitude, and derives a correction factor with timeliness and high relevance using a specific form of exponential decay model. This factor reflects the potential threat level of wind field intensity to flight stability under specific operating conditions, providing a quantitative reference benchmark for subsequent fusion calculations. It is worth noting that this type of compensation mechanism is not a simple superposition of a linear proportional term, but fully considers the trend of gradual weakening with increasing altitude within the atmospheric boundary layer, thus making it more scientifically sound.

[0104] Step S305: Combine the basic voxel size with the airflow disturbance compensation coefficient to generate dynamic voxel parameters.

[0105] Specifically, the fusion operation uses a weighted summation algorithm:

[0106] ;

[0107] In the above formula, S base Based on voxel size, C wind W1 and W2 are preset weights and are the airflow disturbance compensation coefficients.

[0108] Among them, the basic voxel size and the airflow disturbance compensation coefficient are two independent evaluation indicators that characterize the complexity of ground geometry and the risk of airborne disturbances, respectively. However, using either one alone is insufficient to fully reflect the changing patterns of overall mapping needs. Therefore, it is necessary to use a synthesis rule to coordinate them and form an optimal sampling strategy configuration that can both take into account terrain diversity and resist external interference.

[0109] In this embodiment, a classic weighted average method is used, whereby the two parameters are mixed in a certain proportion to obtain a new intermediate state parameter. The weights can be flexibly adjusted according to different application scenarios; for example, wind resistance stability can be prioritized in densely populated urban areas, while coverage can be increased in open suburban areas. The final result is the ideal voxel side length value recommended for this surveying cycle, which can be used to guide the operation of subsequent point cloud filtering, mesh reconstruction, and other related processing modules.

[0110] The above embodiments realize the intelligent control function of sampling density in three-dimensional space. The parameter linkage mechanism proposed by this technical solution not only conforms to the causal transmission chain that exists objectively in nature, but also meets the needs of efficient and accurate data acquisition in engineering practice, thereby improving the autonomous decision-making ability and on-site response level of the unmanned aerial photogrammetry system.

[0111] Reference Figure 4 As one implementation of step S106, the step of optimizing the spatial distribution of the preprocessed point cloud set by applying dynamic voxel parameters and a pre-configured airflow disturbance compensation model to generate an adaptive point cloud model includes:

[0112] Step S401: Obtain wind speed data from the preprocessing point cloud, dynamic voxel parameters, and environmental monitoring data;

[0113] Among them, the preprocessed point cloud refers to the set of scattered three-dimensional points obtained by lidar or photogrammetry after a series of standardized operations such as preliminary denoising, registration, and coordinate unification. These points constitute the basic objects of all subsequent spatial analysis. The dynamic voxel parameter is a core variable used to define the granularity of three-dimensional spatial division. The wind speed vector, as one of the external environmental disturbance terms, is introduced into the calculation process. It not only includes the horizontal velocity component, but may also cover the vertical airflow component, which provides the necessary physical input conditions for subsequent consideration of aerodynamic effects.

[0114] Step S402: Construct an adaptive spatial mesh structure based on dynamic voxel parameters;

[0115] The term "adaptive" means that the grid is not a fixed, traditional cube array, but a hierarchical spatial indexing architecture that can flexibly adjust its resolution according to changes in point cloud density within the region.

[0116] Specifically, the construction of the adaptive spatial grid structure includes: establishing a three-dimensional grid coordinate system with dynamic voxel parameters as the unit size; assigning a spatial index identifier to each grid unit; and associating the spatial mapping relationship between the grid unit and the point cloud.

[0117] In this embodiment, classic algorithms such as octrees or multi-level KD trees are used to determine whether to further subdivide nodes based on the point density within the current region. This avoids resource waste caused by over-refinement in sparse areas while ensuring sufficient expressive power in dense areas. During this process, each voxel is assigned a unique spatial index identifier, facilitating rapid retrieval and association with point sets within its corresponding range. This structured approach significantly improves the time efficiency of subsequent mapping, querying, and even update operations, and lays a solid foundation for accurate spatial projection in the next step.

[0118] Step S403: Map the preprocessed point cloud set to an adaptive spatial grid structure to generate an initial voxelized point cloud;

[0119] This step employs an efficient point cloud spatial partitioning strategy. By determining the voxel affiliation of each point (i.e., identifying its specific grid number), the originally unordered point sequence can be transformed into an ordered set of data blocks, ensuring that each voxel retains only a limited number of representative sample points or cluster centers. It is worth noting that due to the use of a dynamic voxel mechanism, sampling densities may differ between different regions. Therefore, appropriate interpolation methods must be used to maintain overall continuity and prevent significant boundary jumps.

[0120] Step S404: Call the pre-configured airflow disturbance compensation model and calculate the point cloud displacement compensation amount in combination with wind speed data;

[0121] In traditional surveying work, the impact of factors such as aircraft attitude fluctuations and atmospheric flow on the acquisition results is often ignored. However, in high-precision applications, these error sources may significantly reduce the quality of the results.

[0122] To address this issue, this application designs a mathematical model specifically for wind-induced displacement. This model comprehensively considers multiple dimensions of information, including ground material properties (such as roughness and reflectivity), elevation gradient, and real-time measured wind field vectors. Specifically, the model embeds a damping coefficient database, providing corresponding drag constant values ​​based on different surface types (e.g., grassland, concrete pavement, water surface), thereby participating in the quantitative assessment of wind pressure effects. The final output is a set of three-dimensional vector-based displacement compensation quantities. It can be used to correct the position of point groups within individual voxels, thereby offsetting the systematic deviation caused by wind.

[0123] Specifically, calling the airflow disturbance compensation model includes: inputting point cloud height data and surface material type code; querying the damping coefficient database based on the material type code; and calculating the displacement compensation amount under wind speed through vector operations.

[0124] The calculation of the displacement compensation amount satisfies the following: , Let K be the wind speed vector. d H is the material damping coefficient, and H is the point cloud height.

[0125] Step S405: Apply point cloud displacement compensation to the initial voxelized point cloud to generate a corrected point cloud;

[0126] This step does not simply involve translating the absolute coordinates of individual points; rather, it aims to perform a global spatial reconstruction while preserving the original topological relationships. To achieve this, a homogeneous transformation matrix can be used to encapsulate the combination of rotation and translation operations, allowing each cluster of points to move an appropriate distance in a specified direction without disrupting their relative positions. Furthermore, when points, after compensation, extend beyond the original grid boundaries, they should be reclassified to new neighboring cells to avoid misclassification in subsequent statistical analysis. This stage aims to restore the deformation distortion caused by external disturbances, making the point cloud as close as possible to the true terrain.

[0127] Step S406: Optimize the distribution density of the correction point cloud based on the curvature density function to output an adaptive point cloud model with spatial consistency.

[0128] Although the aforementioned compensation measures have significantly improved the overall accuracy of point clouds, there are still potential defects if they are directly regarded as the final product, especially in areas with drastic terrain undulations and complex textures, where local density may be too high or too low.

[0129] To resolve this contradiction, embodiments of this application introduce a nonlinear mapping function, namely the curvature-density mapping relationship, the specific formula of which is: The function describes the response curve between curvature c and its ideal sampling density ρ, with k and c0 as preset parameters. Essentially, this function simulates a sigmoid activation mechanism: when the surface curvature of a region is large (i.e., c > c0), the density of corresponding points is enhanced to capture more detailed features; conversely, the density is lower. This spatial sampling rule not only better reflects the true shape of natural terrain but also effectively suppresses noise interference, improving the robustness and reliability of subsequent visualization rendering and data analysis.

[0130] It should be noted that "spatial consistency" has multiple meanings here: First, at the macro level, the transition between regions is natural and seamless, with no obvious breakpoints; second, at the micro level, each local area maintains a reasonable point distribution pattern, neither too sparse nor too crowded; and third, in the time dimension, even multiple iterations will not cause problems such as cumulative drift.

[0131] Therefore, a series of post-processing steps are required before the formal output, such as removing isolated outliers that are outside the legal range, applying smoothing filtering to the boundaries of adjacent grids, and reconstructing missing adjacency links, so as to obtain an ideal 3D point cloud asset that has integrity, stability and practical value, suitable for various cutting-edge application scenarios such as urban planning, disaster early warning, and autonomous driving navigation.

[0132] The above implementation deeply integrates advanced concepts such as environmental perception, mechanical modeling, and intelligent scheduling, truly achieving the goal of controllability, predictability, and reproducibility throughout the entire process from source data collection to terminal presentation. Especially when facing complex field operation conditions, this technical solution demonstrates strong robustness and generalization ability.

[0133] Reference Figure 5 As one implementation of step S108, the step of fusing the enhanced point cloud model and image data to perform semantically constrained 3D reconstruction and outputting a semantically annotated 3D model includes:

[0134] Step S501: Obtain image data and spatial connectivity data corresponding to the enhanced point cloud model;

[0135] The image data is a sequence of aerial photographs that has undergone precise spatiotemporal synchronization through a precise clock protocol, interpolation, and timestamp correction matrix processing. Each frame of the image carries absolute geographic coordinates and attitude angle information provided by the GNSS / INS integrated navigation system, ensuring the image's position and attitude determinism in the global coordinate system.

[0136] Meanwhile, the "spatial connectivity data" comes from the output of the adaptive point cloud model after topological reconstruction (i.e., the "enhanced point cloud model with topological structure"). It is not the original point cloud, but a structured description that represents the adjacency relationship between data points in the point cloud, such as the initial triangulation based on the Delaunay triangulation principle, the point cloud normal vector field, and the connection graph between vertices.

[0137] Step S502: Extract semantic segmentation map from image data through image segmentation processing, and extract land cover category boundary information from semantic segmentation map;

[0138] Image segmentation typically relies on pre-trained deep convolutional neural networks (such as DeepLabv3+, UNet, etc.), which can perform pixel-level parsing of the input image and output a "semantic segmentation map". This segmentation map is a matrix with the same resolution as the input image, where each pixel position is no longer just a color value, but contains a probability distribution of predefined land cover categories such as buildings, vegetation, and roads.

[0139] Subsequently, "feature category boundary information" is extracted from the "semantic segmentation map" by applying methods such as Canny edge detection or calculating category boundary gradients from the segmentation probability map. This boundary information is represented by subpixel precision, continuous contour lines or closed polygons (such as the precise roof outline of a building), which identify the precise boundaries between different semantic categories on the two-dimensional image plane.

[0140] Step S503: Project the boundary information of the land cover category onto the three-dimensional space where the enhanced point cloud model is located, establish a semantic-geometric association mapping, and add corresponding semantic category labels to the point cloud data of the enhanced point cloud model;

[0141] Specifically, the perspective projection matrix is ​​calculated based on the camera pose parameters (rotation matrix R and translation vector t) and camera intrinsic parameters (intrinsic parameter matrix K) corresponding to each image. Through the inverse transformation of this projection matrix, the extracted two-dimensional land cover category boundary information (i.e., a series of boundary pixel coordinates) is back-projected into the three-dimensional space where the "enhanced point cloud model" is located.

[0142] In three-dimensional space, spatial search algorithms (such as KD-Tree nearest neighbor search) are used to find the best-matching 3D point cloud data point in the "enhanced point cloud model" for each projected boundary point, thus establishing a one-to-one or one-to-many "semantic-geometric association mapping." This means that specific 3D points are explicitly associated with semantic concepts such as "building boundaries" and "road edges." Based on this mapping relationship, the system "adds corresponding semantic category labels" to the point cloud data points in the "enhanced point cloud model," that is, the semantic category is added as a new attribute field of the point cloud.

[0143] This step endows the original geometric point cloud with rich semantic connotations, enabling subsequent mesh construction and simplification operations to "understand" and respect the geometric characteristics of different land cover categories, thus realizing the rigid guidance of semantic information on the geometric modeling process.

[0144] Step S504: Based on the spatial connectivity data, construct an initial triangular mesh surface on the enhanced point cloud model;

[0145] This method transforms discrete, semantically labeled point cloud data into continuous, surface representations capable of efficient computer graphics operations. The input data is spatial connectivity data, which defines the adjacency relationships between points in the point cloud data, essentially providing an initial, point-based topology graph.

[0146] Based on this, surface reconstruction algorithms such as constrained Delaunay triangulation or Poisson reconstruction can be used. Using these point cloud points as vertices, and strictly following the connection rules implicitly or explicitly defined in the "spatial connectivity data," an "initial triangular mesh surface" covering the entire point cloud region is generated. This initial mesh inherits the geometric details and spatial accuracy of the "enhanced point cloud model," while its triangular facet structure completely follows the original point cloud topology, ensuring the mesh's manifold properties (no holes, unambiguous boundaries) and geometric correctness. This step creates a continuous surface model that can be used for visualization, analysis, editing, and simplification, serving as a crucial bridge connecting discrete point cloud observations with the final structured 3D model.

[0147] Step S505: Simplify the initial triangular mesh surface, identify and lock the coordinates of the mesh vertices located on the boundary of the land cover category according to the semantic-geometric association mapping, and apply a non-uniform simplification strategy to the mesh cells on both sides of the boundary of the land cover category to generate a simplified mesh surface.

[0148] One approach is to introduce a semantic boundary protection mechanism into traditional mesh simplification algorithms to resolve the conflict between model simplification efficiency and feature preservation. Traditional mesh simplification algorithms (such as edge folding algorithms based on quadratic error metrics) globally reduce the number of triangular faces, but often produce smoothing distortions at sharp features such as building eaves and road curbs.

[0149] In this embodiment, the system first identifies the grid vertices corresponding to the semantic boundary. Their spatial coordinates are frozen during the simplification process and cannot be moved or merged, thereby geometrically anchoring these key feature lines. Under this constraint, a lower simplification degree (e.g., ≤40%) is forced in the grid cell regions on both sides of the land cover category boundary, allowing only the removal of secondary vertices that do not affect the boundary contour; while in flat or non-feature regions far from the semantic boundary, a higher simplification degree (e.g., ≤80%) is allowed, significantly reducing model redundancy through extensive vertex merging and edge folding.

[0150] Understandably, this differentiated processing can significantly compress the amount of model data and improve the efficiency of subsequent processing and rendering, while preserving the geometric sharpness and accuracy of the outlines of ground features with important semantic meaning and visual recognizability to the maximum extent, thus achieving the optimal balance between efficiency and fidelity.

[0151] Step S506: Based on the spatial correspondence between mesh vertices and image data, the image data is mapped onto the simplified mesh surface as a texture source to generate a mesh model with texture mapping.

[0152] Specifically, using a known camera projection model, the corresponding pixel coordinates (i.e., UV coordinates) of each grid vertex on the simplified grid surface in multiple image data are calculated. Since the positions of grid vertices (especially boundary vertices) are protected by semantic constraints, this correspondence is particularly accurate at feature edges.

[0153] Subsequently, texture mapping technology is used to attach the color information of the original high-resolution image onto the corresponding triangular facets. During this process, the system adjusts the position of texture seams based on semantic boundaries, ensuring that the seams are hidden as much as possible in non-feature areas. Advanced image fusion algorithms, such as Poisson fusion, are employed to smooth texture color differences and seams caused by variations in lighting and viewing angles from images from different perspectives, thereby generating a texture-mapped mesh model. Ultimately, a 3D model with both precise geometry and realistic visual texture is generated, greatly enhancing the model's realism and visualization, enabling it to meet high standards of visual display and interpretation.

[0154] Step S507: Fuse the semantic category labels carried in the enhanced point cloud model with the mesh model with texture mapping to output a 3D model with semantic annotation.

[0155] Specifically, the system passes semantic category labels and assigns them to each triangle or vertex of the texture-mapped mesh model. Simultaneously, these semantic labels can be mapped to standardized color values, generating a semantic annotation layer independent of the real texture to highlight different land cover categories.

[0156] The final output, a semantically annotated 3D model, is a comprehensive dataset integrating high-precision geometric meshes, high-fidelity image textures, and structured semantic attributes (such as category coding). This output can be directly used in GIS systems to filter buildings by semantic categories, calculate vegetation cover areas, or perform spatial relationship analysis, greatly expanding the in-depth application value of 3D models in fields such as smart cities, asset management, and environmental monitoring.

[0157] In the above implementation, the two-dimensional semantic boundaries extracted by deep learning are used as rigid constraints and deeply integrated into the entire process of 3D reconstruction and simplification from point cloud to mesh. This technical solution successfully solves the long-standing pain point of blurred and distorted feature edges caused by indiscriminate simplification in traditional 3D reconstruction methods. While ensuring the overall lightweight nature and processing efficiency of the model, it maintains the geometric accuracy of key feature contours. The final generated 3D model not only has a visually highly realistic texture appearance, but more importantly, it embeds structured semantic information that can be directly recognized by machines, providing a high-value data foundation for fields such as smart cities, digital twins, and refined planning and management.

[0158] Reference Figure 6 As one implementation of step S109, the steps of extracting point cloud curvature features based on the adaptive point cloud model, generating encrypted verification point cloud, and calculating the elevation deviation between the encrypted verification point cloud and the 3D model include:

[0159] Step S601: Obtain the adaptive point cloud model and the 3D model with semantic annotation;

[0160] Step S602: Perform curvature feature analysis on the adaptive point cloud model to generate a curvature distribution map;

[0161] Among them, curvature feature analysis is performed on the obtained adaptive point cloud to extract key geometric attributes of areas with dramatic topographic relief.

[0162] In this embodiment, the moving least squares method is used to perform local quadratic surface fitting on each point and several nearest neighbors to estimate the normal vector and principal curvature direction at that point. Furthermore, the Gaussian curvature formula k=k1·k2 and the average curvature H=(k1+k2) / 2 are combined to quantify the local morphological complexity, ultimately outputting a spatial distribution map representing the curvature magnitude of different parts of the entire scene. This process not only reveals potential abrupt terrain changes (such as cliff edges and building corners) but also provides a mathematical basis for selecting key areas in the next step.

[0163] Step S603: Identify high curvature feature regions based on the curvature distribution map;

[0164] Specifically, the identification of high curvature feature regions satisfies the following condition: when the curvature value is >0.8, it is marked as an encrypted region, and the boundary extension range of the encrypted region is 1.2 times the dynamic voxel parameter.

[0165] One approach involves setting a threshold condition: when the curvature value corresponding to a point is greater than 0.8, the location of that point is marked as a target area requiring focused encryption. This value is not arbitrarily chosen, but rather an empirical indicator derived from a comprehensive consideration of the balance between the intensity of typical terrain transitions and sensor resolution in real-world application scenarios.

[0166] In addition, to ensure a natural boundary transition without missing important information, this type of region needs to be expanded outward by a certain range. The expansion scale is determined by the dynamic voxel parameters in the current environment and multiplied by a coefficient of 1.2 to compensate for the possible edge effect, thereby obtaining a more robust and reliable encrypted candidate region.

[0167] Step S604: Perform point cloud encryption sampling in the high curvature feature region to generate an encrypted verification point cloud;

[0168] Specifically, point cloud encryption sampling includes: generating Poisson disk-distributed sampling points within the encryption region, setting the sampling density to 1.5 times the original point cloud density, and reconstructing the topological connectivity of the sampling points.

[0169] The study incorporates a Poisson disk sampling algorithm, which effectively avoids clustering issues associated with traditional random sampling while maintaining sample uniformity. The density of newly generated sampling points is set to 1.5 times the original point cloud density to enhance local detail representation and improve subsequent matching accuracy.

[0170] More importantly, a reasonable topological connection relationship needs to be established between these newly added points to support higher-level geometric computation requirements. To this end, an octree index architecture based on a dynamic voxel partitioning mechanism was adopted to organize all point information. Then, the local surface network structure was reconstructed with the help of Delaunay triangulation technology. Finally, abnormal edges with a length exceeding 3 times the size of a single voxel unit were removed to improve the overall network quality and robustness.

[0171] Step S605: Spatial registration of the encrypted verification point cloud with the 3D model, and calculation of the set of elevation deviation values ​​of the corresponding points after registration.

[0172] Specifically, the steps of spatial location registration include: extracting feature descriptors of the encrypted verification point cloud, performing feature matching with the vertices of the 3D model, and calculating the rigid body transformation matrix through singular value decomposition.

[0173] In this embodiment, modern feature extraction tools such as fast point feature histograms or SHOT descriptors are used to extract representative and highly discriminative local shape fingerprints from the two sets of data. Then, a nearest neighbor search algorithm is used to match similar feature point pairs. Finally, singular value decomposition is used to solve for the optimal rigid body transformation matrix (rotation and translation parameters), maximizing the geometric consistency between the two datasets in the global coordinate system. The entire registration process fully embodies the concept of cross-modal data collaboration.

[0174] Subsequently, for each point in the encrypted verification point cloud, the system can find its vertical projection point or nearest point on the registered 3D model surface and calculate the coordinate difference between the two in the vertical direction (usually the elevation direction, i.e., the Z-axis direction). This difference is the "elevation deviation" at that point. By traversing all verification points, a complete "elevation deviation value set" is obtained. This set not only provides overall error statistics (such as average error and root mean square error), but more importantly, it reveals the spatial distribution pattern of the error, clearly indicating which areas, under what terrain or environment, the geometric accuracy of the model may have problems, thus providing a direct, quantitative, and location-specific basis for the aforementioned feedback adjustment.

[0175] Specifically, a quantitative elevation difference analysis, i.e., elevation deviation calculation, is performed on the corresponding point pairs that have completed spatial alignment. The specific formula is as follows: , where P z M represents the Z-axis coordinate value from a specific point in the encrypted verification point cloud. z This represents the height reading of the projection point of the corresponding 3D model vertex on the same horizontal plane. By iterating the above calculation process repeatedly over all conjugate points, the overall elevation error of the entire survey area can be summarized.

[0176] In the above implementation, an objective and accurate internal quality inspection mechanism was established, which transformed the accuracy assessment from the traditional method of relying on external control points to a fully automated process that utilizes the high density of verification points on the data itself. This greatly improved the efficiency of quality inspection and the credibility of the results, and achieved an effective transformation from raw observation data to final evaluation conclusions.

[0177] Reference Figure 7 As a further implementation of the high-precision UAV mapping data processing method, after the step of outputting a semantically annotated 3D model, the method further includes:

[0178] Step S701: Obtain the semantically annotated 3D model and corresponding environmental index for historical surveying cycles;

[0179] The core logic of this step lies in incorporating the 3D model from a single survey into a dynamic time-series analysis framework. The semantically annotated 3D model of the historical surveying cycle refers to a standardized 3D digital twin constructed during the historical surveying process, which also embeds precise semantic tags for buildings, vegetation, roads, etc. Meanwhile, environmental indices include, but are not limited to, terrain complexity (TRI), wind speed disturbance coefficient, and atmospheric transmittance. These factors collectively reflect the natural environmental conditions faced by the surveying operation within a specific time period and the degree of their potential impact. For example, in high-wind conditions, lidar echo signals may be disturbed, leading to a decrease in point cloud density; and under heavy rainfall conditions, ground reflectivity changes, thus affecting reconstruction accuracy. Therefore, transforming these variables into environmental indices at a unified scale helps quantify the external differences between different observation periods, thereby more accurately determining which changes are caused by actual environmental variations rather than external interference.

[0180] Understandably, acquiring these two types of data establishes a complete historical baseline for change detection, encompassing geometric morphology, semantic attributes, and environmental context information. Without semantic annotation, computers cannot automatically identify and compare different land cover categories; without environmental indices, it is impossible to distinguish whether changes in land cover originate from their own evolution or are merely superficial differences caused by environmental disturbances from different surveying periods (such as point cloud jitter caused by strong winds or seasonal changes in vegetation).

[0181] In this embodiment of the application, a spatiotemporal coding mechanism is adopted to uniquely identify each record in order to effectively manage this batch of multi-source heterogeneous historical data resources. Specifically, by combining the model ID, timestamp, and spatial grid number into a composite coded string, the objects to be compared between any two periods can be quickly located, and their precise matching in time and geographical location can be ensured.

[0182] Step S702: Perform spatial benchmark unification processing on the 3D models of the current surveying cycle and the historical surveying cycle, and extract the structural feature parameters of the same type of land features in the two models based on semantic annotation;

[0183] Spatial benchmark unification is a prerequisite for ensuring comparability of data from multiple periods. Although each period's data undergoes high-precision GNSS / INS positioning, due to control point errors between different flight missions, minor differences in coordinate systems, or cumulative drift, the two models may exhibit overall translation, rotation, or scaling deviations in the absolute coordinate system. Direct comparison would misinterpret this systematic deviation as ground feature changes. The core principle of benchmark unification is to eliminate these benchmark differences through feature matching and least squares optimization.

[0184] To this end, this application adopts an affine transformation method based on common control points (GCPs) as the core means of unifying the benchmark. First, stable ground feature points common to both models (such as building corners, bridge supports, and fixed sign edges) are selected, and the optimal spatial transformation matrix, including rotation and translation components, is fitted using the least squares method to accurately map the historical model to the current coordinate system. Based on this, the Iterative Nearest Point (ICP) algorithm is further introduced to optimize the matching quality, and a KD-tree is used to accelerate the neighbor point search process, improving the speed and stability of point cloud registration. After this processing, the model is within the same reference frame, eliminating the risk of positional misalignment caused by coordinate system asynchrony, and laying a solid foundation for the next step of semantically guided feature extraction.

[0185] Subsequently, leveraging the rich semantic information already injected into the model, the computer is guided to extract the most representative quantitative indicators of key morphology and state for different land cover categories. For example, for a set of triangular facets labeled "buildings," the system calculates their "roof area" (reflecting changes in building floor plan size) and "facade inclination" (monitoring potential tilting or deformation of walls); for the "vegetation" category, it calculates their "canopy volume" (reflecting vegetation growth or deforestation) and "projection density" (characterizing vegetation density); for "roads," it extracts their "road width" (monitoring widening or encroachment) and "radius of curvature" (analyzing changes in route direction). These structural feature parameters condense complex three-dimensional geometric forms into mathematically comparable key indicators, allowing the computer to bypass pixels or point clouds that are difficult to compare directly and instead focus on attribute changes with clear physical meaning.

[0186] It should be noted that the corresponding feature extraction process will only be initiated once an area is confirmed to belong to a certain type of land cover. This method greatly reduces the burden of unnecessary computation, while improving the accuracy of change recognition and avoiding errors such as mistaking the shadow of leaves for road damage.

[0187] Step S703: Calculate the differences in structural characteristic parameters and identify areas of land cover change;

[0188] Specifically, based on statistical significance testing and thresholding segmentation, the system identifies the spatial range of changes with practical significance from continuous parameter differences. After completing spatial alignment and parameter extraction, the system calculates the difference in "structural feature parameters" for the same land cover in the two models for each category and element. This may be a relative rate of change (such as the percentage of area change) or an absolute difference (such as the number of meters of width change).

[0189] However, since 3D reconstruction itself has a certain degree of accuracy error, minor parameter fluctuations may fall into the category of noise. Therefore, the key to "identifying areas of change in ground features" lies in establishing scientifically reasonable judgment thresholds. These thresholds (e.g., roof area change rate > 5%, canopy volume change rate > 15%) are confidence intervals determined comprehensively based on the surveying accuracy level, ground feature type characteristics, and actual application requirements.

[0190] Subsequently, when the parameter difference of a certain land feature instance exceeds the threshold corresponding to its category, the system determines that the land feature has undergone significant change and marks its spatial extent as a "land feature change area". The logic of this process is to transform continuous numerical differences into discrete change patches with clear geographical boundaries and change type labels (such as "building expansion area" and "vegetation degradation area") through threshold filtering, providing clear objects for subsequent spatial analysis and attribution.

[0191] Step S704: Associate the changing trends of environmental indices with the spatiotemporal distribution of land cover change areas, and generate an analysis report containing the relationship between change types and environmental factors.

[0192] Understandably, the system will perform multidimensional correlation analysis between the spatiotemporal distribution information of identified land cover changes and the corresponding environmental indices and their trends from the two survey periods. For example, the system will analyze whether the corresponding historical and current environmental indices (especially the wind speed, temperature, and humidity components) of a region marked as having "significantly reduced vegetation canopy volume" have also experienced drastic fluctuations (such as encountering extreme windy weather). By using statistical methods (such as spatial overlay analysis and regression analysis) to correlate these co-occurrence relationships, a statistical model between environmental offsets and land cover changes can be established.

[0193] Specifically, by applying a sliding window average to meteorological observation data and other environmental factor observations over many years, a smooth trend function f(t) is obtained to characterize the long-term climate evolution trajectory of a certain region. This function is then cross-validated with the previously obtained land cover change map, and the Pearson correlation coefficient formula is used to measure the degree of agreement between the two.

[0194] ,

[0195] Among them, E i G represents the deviation value of the environmental index corresponding to the i-th observation time. i This represents the total change of a certain type of land cover in the region during the same period. Through such regression analysis, we can not only identify the most influential dominant environmental factors (such as the effect of temperature gradient on the aging rate of asphalt pavement), but also estimate their influence weights, providing theoretical support for future risk prediction and disaster prevention.

[0196] Ultimately, the generated analysis report is not merely a change detection map, but a data-driven research conclusion. The report clearly lists the location, type, and intensity of the changed areas, and displays corresponding environmental index curves, dynamic voxel parameter adjustment records, and other metadata. It even quantifies the potential correlation between specific environmental pressures (such as persistent strong winds) and specific types of land cover deformation (such as slight building displacement or vegetation morphology changes) using formats like an "environment-deformation coupling matrix." This correlation analysis has significant practical implications, helping to determine whether vegetation changes are caused by natural seasonal changes or pests and diseases, and whether building deformation is due to construction or environmental factors affecting the foundation. This provides a deeper level of decision-making support for disaster early warning, infrastructure health monitoring, and ecological environmental protection.

[0197] In this embodiment, the generated relationship matrix table comprehensively reflects the close relationship between various land features and their most important influencing factors, listing the correlation coefficient and influence weight of each set of control items, helping users quickly grasp the overall situation. For example, there is a high negative correlation between "reduction in tree canopy volume" and "increased wind speed disturbance coefficient". The relationship indicates that frequent extreme weather events have indeed, to some extent, suppressed the development vitality of forest ecosystems.

[0198] The above embodiments realize a profound evolution of UAV mapping from high-precision reconstruction of a single scene to multi-temporal dynamic monitoring and intelligent interpretation. This technical solution can not only automatically and accurately detect changes in the geometric shape and semantic attributes of surface targets, but also, by introducing an "environmental index"—a quantitative environmental background field that runs through the data acquisition and processing process—place the detected changes within their environmental context for correlation analysis, thereby achieving a synergistic insight into the "characteristics" and "causes" of changes. This application enhances the in-depth application value of mapping results in deformation monitoring, land surveys, ecological assessments, and other fields, transforming data from static geographic information archives into dynamic intelligent knowledge systems that reflect the interaction between humans and the environment, support trend prediction, and causal analysis.

[0199] As one implementation method for structural feature parameters, the extraction of structural feature parameters includes the following three types:

[0200] (1) Extract the roof area and facade slope angle for each building category;

[0201] Specifically, the extraction of roof area essentially involves boundary identification and area calculation of a two-dimensional orthogonal projection of a building's three-dimensional surface model onto a horizontal plane. The roof area directly quantifies the building's floor plan footprint, and its changes (e.g., >5%) can keenly indicate the building's expansion, addition, partial demolition, or complete disappearance, making it the most direct and stable indicator reflecting construction activities.

[0202] In this embodiment, firstly, all triangular facets belonging to the "building" are selected through semantic annotation. Then, by calculating the direction of the normal vectors of these facets, facets that are approximately horizontal (usually with a very small angle between the normal vector and the zenith direction) are identified as roof faces. Next, the boundaries of these roof facets are projected onto a horizontal reference plane. Through polygon merging and boundary optimization algorithms, a precise two-dimensional polygon reflecting the outermost contour of the building is obtained, and the roof area is then calculated.

[0203] The extraction of facade tilt angle focuses on the vertical surface of a building to monitor the health and safety of its structural posture. The principle is as follows: for triangular facets semantically labeled "facade" (typically with normal vectors close to horizontal), principal component analysis or fitting plane equations is used to calculate the average normal vector of that local facade or the entire wall surface, and then the angle between that facade and the absolute vertical direction (the direction of gravity) is calculated, i.e., the facade tilt angle. This parameter is extremely sensitive to uneven settlement of the building, foundation deformation, or tilting caused by external forces (such as strong winds or geological disasters). For example, a facade tilt angle that changes from 0° to 3° for a previously vertical wall may indicate potential structural risks.

[0204] Therefore, the combination of roof area and facade angle, from the two orthogonal dimensions of "planar scale" and "vertical posture", fully depicts the core characteristics of the building in terms of "quantitative change" and "qualitative change".

[0205] (2) Extract canopy volume and projection density for vegetation categories;

[0206] Canopy volume extraction aims to quantify the physical size occupied by vegetation in three-dimensional space. It is a key indicator reflecting vegetation biomass and growth status, and its technical principle is usually based on spatial segmentation and voxel counting. First, point cloud or model portions belonging to individual plants or specific areas of "vegetation" are isolated based on semantic annotation. Then, using the 3D AlphaShapes algorithm, convex hull algorithm, or the more commonly used voxelization method, the space occupied by vegetation is divided into tiny cubic units (voxels). By counting the number of voxels occupied by all vegetation and multiplying it by the volume of a single voxel, the canopy volume of the vegetation can be estimated. Significant changes in volume (e.g., >15%) directly indicate vigorous vegetation growth, deforestation, or damage caused by natural disasters.

[0207] Projection density, on the other hand, provides a complementary measure of vegetation density and vertical structural complexity from a vertical perspective. It works by vertically projecting a 3D vegetation model onto a horizontal plane, creating a 2D density image. In this process, by analyzing the contributions of points or patches from different height layers within the projected area, or by calculating the corresponding surface area (or point cloud quantity) of the 3D model per unit projected area, a value characterizing the vertical density of vegetation is obtained. High projection density indicates dense vegetation with rich layers; low density may indicate sparse vegetation or a period of leaf fall.

[0208] Therefore, the combination of canopy volume and projected density enables the coordinated measurement of vegetation's "overall size" and "internal fullness," effectively distinguishing between leaf density changes caused by seasonal alternation and the overall structural disappearance caused by removal or death.

[0209] (3) Extract road width and radius of curvature for road categories.

[0210] The core of road surface width extraction lies in accurately identifying and measuring its lateral boundaries from the road surface model. The technical principle is as follows: First, a triangular mesh of the road surface is obtained based on semantic annotation. Then, through mesh edge detection or cross-sectional slicing analysis of the road point cloud, boundary points on both sides of the road (such as curb edges) are identified on multiple cross-sections perpendicular to the general direction of the road. By connecting these boundary points and calculating the horizontal distance between them, a series of road surface width samples along the road direction can be obtained. Their average value or characteristic value (such as minimum width) directly reflects the road's traffic capacity, while width variations (such as >0.5 meters) indicate road widening, encroachment, or road surface wear.

[0211] The extraction of the radius of curvature is used to quantify the curvature of the road centerline, a core parameter for road design safety and alignment quality. Its principle involves centerline extraction and curve fitting. First, a three-dimensional centerline representing the road's direction is extracted from the road surface model using algorithms such as skeletonization or optimal path search. Then, this centerline is treated as a spatial curve, and points are taken on it at certain step sizes. An arc is fitted using three or more consecutive points (e.g., through a moving window), and the radius of this arc is the local radius of curvature at that point. A small radius of curvature indicates a sharp bend, while a large radius indicates a gentle bend or a straight section. Changes in the radius of curvature (e.g., a change rate > 10%) may originate from road realignment, alignment changes due to landslides, or initial design optimizations.

[0212] Therefore, road width and radius of curvature accurately depict the geometric design features of a road from the two dimensions of "cross section" and "longitudinal section", enabling its change detection to serve professional fields such as traffic planning, safety assessment and infrastructure maintenance.

[0213] In the above implementation, the complex 3D visual reconstruction results are transformed into "computable" knowledge units that can be directly used for advanced geospatial analysis and change attribution. This enables automated change detection systems to no longer rely solely on coarse pixel differences or voxel occupancy comparisons, but to perform professional and quantitative state assessments and change monitoring of different land features, just like domain experts.

[0214] As one implementation of the rules for identifying areas of change in land features, the rules specifically include: marking areas as building change zones when the roof area change rate is greater than 5% or the facade tilt angle change is greater than 3°; marking areas as vegetation change zones when the canopy volume change rate is greater than 15%; and marking areas as road change zones when the road width change is greater than 0.5 meters or the radius of curvature change rate is greater than 10%.

[0215] Specifically, for land features of different semantic categories, quantitative thresholds with clear physical meaning and statistical significance are set according to their inherent physical characteristics, change patterns and measurement uncertainties, thereby transforming continuous differences in feature parameters into a reliable binary judgment of "change" and "no change".

[0216] In this embodiment of the application, for buildings, the rule of roof area change rate > 5% focuses on capturing substantial expansion or demolition of its planar scale, and this threshold usually exceeds the method error of the three-dimensional reconstruction itself; while facade tilt angle change > 3° targets posture anomalies that may indicate structural safety risks, and this angle is significantly larger than conventional construction errors or point cloud registration noise.

[0217] For vegetation, the higher threshold of canopy volume change rate >15% mainly targets its obvious growth or removal, effectively filtering out smaller volume fluctuations caused by seasonal changes, wind swaying, or differences in point cloud density.

[0218] For roads, the absolute value rule for road width changes > 0.5 meters is directly related to significant changes in the standard lane width and has direct engineering significance; the rate of change of curvature radius > 10% is used to capture major adjustments in route orientation.

[0219] This application also discloses a high-precision unmanned aerial vehicle (UAV) mapping data processing system.

[0220] A high-precision UAV mapping data processing system, specifically comprising:

[0221] The multi-source data acquisition module is used to acquire raw monitoring data collected in real time during the flight of the UAV in the test area, including image data, laser point cloud data, attitude data, positioning data and environmental monitoring data;

[0222] The spatiotemporal synchronization processing module is used to perform spatiotemporal synchronization processing on the raw monitoring data to generate a spatiotemporally aligned enhanced dataset;

[0223] The environmental index calculation module is used to calculate terrain complexity parameters and environmental dynamic parameters based on the augmented dataset, and then fuse them to generate an environmental index.

[0224] The dynamic partitioning preprocessing module is used to dynamically partition the survey area according to the environmental index and perform partition point cloud preprocessing to obtain a preprocessed point cloud set with partition annotations.

[0225] The dynamic voxel parameter calculation module calculates dynamic voxel parameters by combining environmental indices and environmental monitoring data.

[0226] The point cloud spatial optimization module is used to optimize the spatial distribution of the preprocessed point cloud set by applying dynamic voxel parameters and a pre-configured airflow disturbance compensation model, and generate an adaptive point cloud model.

[0227] The topology reconstruction module is used to reconstruct the topology of the adaptive point cloud model, generating an enhanced point cloud model with topology.

[0228] The semantic 3D reconstruction module is used to fuse enhanced point cloud models and image data for semantically constrained 3D reconstruction, and outputs a 3D model with semantic annotations.

[0229] The accuracy verification and evaluation module is used to extract point cloud curvature features based on the adaptive point cloud model, generate encrypted verification point cloud, and calculate the elevation deviation between the encrypted verification point cloud and the 3D model.

[0230] The parameter feedback optimization module is used to adjust the dynamic voxel parameters and re-optimize the point cloud spatial distribution when the elevation deviation exceeds the preset deviation threshold.

[0231] The high-precision UAV mapping data processing system of this application embodiment can implement any of the above methods, and the specific working process of each module in the system can refer to the corresponding process in the above method embodiments.

[0232] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0233] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A high-precision UAV mapping data processing method, characterized in that, The data processing method includes: Acquire raw monitoring data in real time during the flight of the UAV in the survey area, including image data, laser point cloud data, attitude data, positioning data and environmental monitoring data; The original monitoring data is subjected to spatiotemporal synchronization processing to generate a spatiotemporally aligned enhanced dataset; Based on the enhanced dataset, terrain complexity parameters and environmental dynamic parameters are calculated and fused to generate an environmental index; The survey area is dynamically partitioned according to the environmental index, and partition point cloud preprocessing is performed to obtain a preprocessed point cloud set with partition annotations. Dynamic voxel parameters are calculated by combining the environmental index with the environmental monitoring data; The point cloud spatial distribution of the preprocessed point cloud set is optimized by applying the dynamic voxel parameters and the pre-configured airflow disturbance compensation model to generate an adaptive point cloud model. The adaptive point cloud model is reconstructed in topology to generate an enhanced point cloud model with topology. The enhanced point cloud model and the image data are fused together to perform semantically constrained 3D reconstruction, and a 3D model with semantic annotation is output. Based on the adaptive point cloud model, point cloud curvature features are extracted, and encrypted verification point cloud is generated. The elevation deviation between the encrypted verification point cloud and the three-dimensional model is calculated. When the elevation deviation exceeds the preset deviation threshold, feedback is provided to adjust the dynamic voxel parameters and re-optimize the point cloud spatial distribution.

2. The high-precision UAV mapping data processing method according to claim 1, characterized in that, The steps of dynamically partitioning the survey area based on the environmental index and performing partition point cloud preprocessing to obtain the preprocessed point cloud set with partition annotations include: Acquire point cloud data from environmental index and augmented datasets; Based on the environmental index, a partition mapping map is generated to characterize the processing level corresponding to each coordinate point in the test area. The point cloud data is divided into multiple point cloud subsets according to the partition mapping diagram, and each subset corresponds to a processing level region. Perform differentiated preprocessing operations on point cloud subsets of regions with different processing levels; Merge the preprocessed point cloud subsets and embed the partition type identifier to generate a preprocessed point cloud set with partition annotation.

3. The high-precision UAV mapping data processing method according to claim 1, characterized in that, The steps for calculating dynamic voxel parameters by combining the environmental index with the environmental monitoring data include: Obtain wind speed data from environmental indices and environmental monitoring data; Analyze the terrain complexity parameter in the environmental index; The basic voxel size is determined based on the terrain complexity parameters; The flight altitude data of the UAV is acquired in real time, and the airflow disturbance compensation coefficient is calculated based on the flight altitude data and wind speed data. The basic voxel dimensions and airflow disturbance compensation coefficients are combined to generate dynamic voxel parameters.

4. The high-precision UAV mapping data processing method according to claim 3, characterized in that, The steps of optimizing the spatial distribution of the preprocessed point cloud set by applying the dynamic voxel parameters and the pre-configured airflow disturbance compensation model to generate an adaptive point cloud model include: Obtain wind speed data from preprocessing point aggregates, dynamic voxel parameters, and environmental monitoring data; An adaptive spatial mesh structure is constructed based on the dynamic voxel parameters; The preprocessed point cloud is mapped to the adaptive spatial grid structure to generate an initial voxelized point cloud. The pre-configured airflow disturbance compensation model is invoked, and the point cloud displacement compensation is calculated in combination with the wind speed data; The point cloud displacement compensation amount is applied to the initial voxelized point cloud to generate a corrected point cloud; The distribution density of the corrected point cloud is optimized based on the curvature density function, and an adaptive point cloud model with spatial consistency is output.

5. The high-precision UAV mapping data processing method according to claim 1, characterized in that, The steps of fusing the enhanced point cloud model with the image data to perform semantically constrained 3D reconstruction and outputting a semantically annotated 3D model include: Acquire the image data and the spatial connectivity data corresponding to the enhanced point cloud model; Semantic segmentation maps are extracted from the image data through image segmentation processing, and land cover category boundary information is extracted from the semantic segmentation maps. The boundary information of the land cover category is projected onto the three-dimensional space where the enhanced point cloud model is located, a semantic-geometric association mapping is established, and corresponding semantic category labels are added to the point cloud data of the enhanced point cloud model. Based on the spatial connectivity data, an initial triangular mesh surface is constructed on the enhanced point cloud model; The initial triangular mesh surface is simplified, and the coordinates of the mesh vertices located on the boundary of the land cover category are identified and locked according to the semantic-geometric association mapping. A non-uniform simplification strategy is applied to the mesh cells on both sides of the boundary of the land cover category to generate a simplified mesh surface. Based on the spatial correspondence between the mesh vertices and the image data, the image data is used as a texture source and mapped onto the simplified mesh surface to generate a mesh model with texture mapping. By fusing the semantic category labels carried in the enhanced point cloud model with the textured mesh model, a semantically labeled 3D model is output.

6. The high-precision UAV mapping data processing method according to claim 5, characterized in that, The steps of extracting point cloud curvature features based on the adaptive point cloud model, generating encrypted verification point cloud, and calculating the elevation deviation between the encrypted verification point cloud and the 3D model include: Obtain the adaptive point cloud model and the semantically labeled 3D model; Perform curvature feature analysis on the adaptive point cloud model to generate a curvature distribution map; Identify high curvature feature regions based on the curvature distribution map; Point cloud encryption sampling is performed in the high curvature feature region to generate an encrypted verification point cloud; The encrypted verification point cloud is spatially registered with the 3D model, and the set of elevation deviation values ​​of the corresponding points after registration is calculated.

7. A high-precision UAV mapping data processing method according to any one of claims 1 to 6, characterized in that, After the step of outputting a semantically annotated 3D model, the following steps are also included: Obtain semantically annotated 3D models and corresponding environmental indices for historical surveying cycles; Spatial benchmark unification processing is performed on the 3D models of the current surveying cycle and the historical surveying cycle, and structural feature parameters of the same type of land cover in the two models are extracted based on semantic annotation; Calculate the differences in the structural feature parameters and identify areas of ground feature change; The changing trends of the associated environmental indices and the spatiotemporal distribution of the areas of change in the land features are used to generate an analysis report that includes the relationship between the change type and the environment.

8. The high-precision UAV mapping data processing method according to claim 7, characterized in that, The extraction of structural feature parameters includes: extracting roof area and facade tilt angle for building categories; extracting canopy volume and projected density for vegetation categories; and extracting road width and radius of curvature for road categories.

9. A high-precision UAV mapping data processing method according to claim 7, characterized in that, The identification rules for the change areas of ground features include: when the roof area change rate is greater than 5% or the facade tilt angle change is greater than 3°, it is marked as a building change area; when the canopy volume change rate is greater than 15%, it is marked as a vegetation change area; when the road width change is greater than 0.5 meters or the radius of curvature change rate is greater than 10%, it is marked as a road change area.

10. A high-precision UAV mapping data processing system, characterized in that, The data processing system includes: The multi-source data acquisition module is used to acquire raw monitoring data collected in real time during the flight of the UAV in the test area, including image data, laser point cloud data, attitude data, positioning data and environmental monitoring data; The spatiotemporal synchronization processing module is used to perform spatiotemporal synchronization processing on the original monitoring data to generate a spatiotemporally aligned enhanced dataset; The environmental index calculation module is used to calculate terrain complexity parameters and environmental dynamic parameters based on the enhanced dataset, and then fuse them to generate an environmental index. The dynamic partitioning preprocessing module is used to dynamically partition the test area according to the environmental index and perform partition point cloud preprocessing to obtain a preprocessed point cloud set with partition annotations. The dynamic voxel parameter calculation module calculates dynamic voxel parameters by combining the environmental index with the environmental monitoring data. The point cloud spatial optimization module is used to optimize the point cloud spatial distribution of the preprocessed point cloud set by applying the dynamic voxel parameters and the pre-configured airflow disturbance compensation model, and generate an adaptive point cloud model. The topology reconstruction module is used to reconstruct the topology of the adaptive point cloud model to generate an enhanced point cloud model with topology. The semantic 3D reconstruction module is used to fuse the enhanced point cloud model with the image data to perform semantically constrained 3D reconstruction and output a 3D model with semantic annotations. The accuracy verification and evaluation module is used to extract point cloud curvature features based on the adaptive point cloud model, generate encrypted verification point cloud, and calculate the elevation deviation between the encrypted verification point cloud and the three-dimensional model. The parameter feedback optimization module is used to adjust the dynamic voxel parameters and re-optimize the point cloud spatial distribution when the elevation deviation exceeds a preset deviation threshold.

Images