Efficient terrain data acquisition method and system for unmanned aerial vehicle based on adaptive sampling and collaborative compression

Abstract

The application discloses a kind of unmanned plane efficient terrain data acquisition method and system of adaptive sampling and collaborative compression, belong to unmanned plane remote sensing and three-dimensional terrain reconstruction technical field.It includes: each unmanned plane real-time calculation dynamic terrain entropy value, and compared with dynamic threshold to trigger high-density adaptive sampling;According to entropy, data is divided into structure, texture, detail three layers, and is compressed differently, to generate compressed data stream and metadata abstract;Each unmanned plane exchanges metadata abstract through self-organizing network, and dynamically elects temporary coordination node to generate task instruction, realizes no center collaboration;Ground station receives compressed data stream, after space-time registration, layered fusion, finally reconstructs three-dimensional terrain model.The system includes corresponding module.The application solves the problems of rigid sampling, large data transmission pressure, unrobust collaborative architecture and low data processing efficiency in the prior art, and realizes efficient, high-precision and high-robustness unmanned plane collaborative terrain mapping.

Description

Technical Field

[0001] This invention belongs to the field of UAV remote sensing and 3D terrain reconstruction technology, specifically relating to an efficient UAV terrain data acquisition method and system based on adaptive sampling and collaborative compression. Background Technology

[0002] Due to their flexibility, efficiency, and low cost, drones have become an important data acquisition platform in fields such as topographic mapping, environmental monitoring, and disaster assessment. However, when facing large-scale missions with complex and varied terrain, the operational efficiency, data coverage, and endurance of a single drone all face severe challenges. Employing multiple drones in collaborative operations is an inevitable trend for solving these problems.

[0003] Existing multi-UAV terrain data acquisition methods mainly suffer from the following drawbacks:

[0004] Rigid sampling strategies result in both data redundancy and information loss: Pre-set fixed routes and uniform sampling densities are typically used, failing to adapt to terrain complexity. This leads to a large amount of redundant data in simple areas, increasing transmission and processing burdens, while in critical complex areas, insufficient sampling may result in the loss of important terrain details, affecting the accuracy of the reconstruction model.

[0005] High data transmission pressure makes communication bandwidth a bottleneck: The amount of data generated by sensors such as lidar and oblique photography on UAVs is enormous. Multiple UAVs simultaneously transmit raw high-definition data back to the ground station, placing extremely high demands on the bandwidth of inter-UAV and air-to-ground communication links. This can easily cause data transmission delays, packet loss, or even link congestion, making it difficult to meet the needs of real-time or near-real-time 3D modeling.

[0006] The centralized coordination mechanism suffers from insufficient robustness and flexibility: Most systems adopt a model of centralized control from ground stations or task allocation by designated lead drones. This centralized architecture is prone to single-point failures and struggles to dynamically respond to unexpected situations involving individual drones, resulting in poor overall system robustness and adaptability.

[0007] Inefficient data processing and fusion: Post-processing of massive amounts of raw data at ground stations is lengthy and lacks real-time performance. When fusing data collected by different UAVs, insufficient registration accuracy or inconsistent feature representation often leads to seams, ghosting, or loss of accuracy in the fused model.

[0008] To address the shortcomings of existing technologies, such as unintelligent sampling, high data transmission pressure, unsustainable collaborative architecture, and poor real-time data processing, this invention aims to provide a UAV terrain data acquisition method and system that enables on-demand adaptive sampling, efficient hierarchical compression of airborne data, decentralized autonomous collaboration, and rapid ground fusion and reconstruction. Summary of the Invention

[0009] In view of this, the purpose of this invention is to provide an efficient terrain data acquisition method and system for UAVs with adaptive sampling and collaborative compression, so as to solve the problems of low data efficiency caused by rigid sampling strategies, high pressure of massive data transmission, insufficient robustness of centralized collaborative architecture, and poor real-time performance of data processing in the prior art.

[0010] To achieve the above objectives, the present invention adopts the following technical solution;

[0011] An efficient UAV terrain data acquisition method and system based on adaptive sampling and collaborative compression, executed by multiple UAVs as peer nodes, includes the following steps:

[0012] S100. Each UAV calculates the dynamic terrain entropy value in real time based on multi-sensor fusion data; when the dynamic terrain entropy value exceeds the dynamic threshold associated with the flight altitude, a high-density sampling mode is triggered, and the sampling density and sensor configuration are adjusted according to the dynamic terrain entropy value, while the original terrain data is collected.

[0013] S200. Each UAV dynamically divides the raw terrain data it collects into three layers of feature data: structure, texture, and detail, based on entropy values. It then performs triangulation simplification compression, wavelet transform compression, and sparse coding compression on each layer of feature data to generate a compressed feature data stream. At the same time, it generates a feature layer metadata summary that represents the spatial distribution and entropy statistics of each layer of feature data.

[0014] S300. Each UAV exchanges feature layer metadata digests through a self-organizing network. Based on a distributed consensus algorithm, a temporary coordination node is dynamically elected with the real-time status of each UAV as input. The temporary coordination node generates and issues task allocation instructions based on the global feature layer metadata digest. Each UAV adjusts its high-density sampling mode according to the instructions.

[0015] S400. Layered Data Fusion and Reconstruction: The ground station receives compressed feature data streams sent by each UAV and performs spatiotemporal registration based on the spatiotemporal labels embedded in the data streams and the feature layer metadata summaries obtained through exchange. Multi-source triangulation fusion is performed on the registered structural layer data, wavelet domain image fusion is performed on the texture layer data, and sparse reconstruction fusion is performed on the detail layer data. Finally, the three-layer fusion results are superimposed to generate a complete three-dimensional terrain model.

[0016] As a further aspect of the present invention, the dynamic terrain entropy value The calculation formula is:

[0017] ;

[0018] The weighting coefficients are configured according to task requirements and satisfy the following conditions: .

[0019] As a further aspect of the present invention, the dynamic threshold function Represented as:

[0020] ;

[0021] This is the ground reference threshold; The attenuation coefficient; This refers to the flight altitude.

[0022] As a further aspect of the present invention, after the high-density sampling mode is triggered, the sampling density is dynamically adjusted according to the extent to which the dynamic terrain entropy value exceeds the dynamic threshold, and the adjusted sampling density is subject to an upper limit limit.

[0023] As a further aspect of the present invention, the step of dynamically dividing the raw terrain data collected by each UAV into three layers of feature data—structure, texture, and detail—based on entropy values ​​includes:

[0024] Read the raw data block from the cache and its associated dynamic terrain entropy value. There are two preset dynamic entropy threshold values. and ,satisfy The hierarchical rules are as follows:

[0025] Structural layer data: composed of entropy values The data blocks are composed of structural layer data, which correspond to flat or gently sloping areas with gentle terrain undulations and simple features.

[0026] Texture layer data: satisfied by entropy values The data blocks are composed of texture layer data, which corresponds to moderately complex terrain.

[0027] Detail layer data: composed of entropy values The data is composed of blocks, with the detailed layer data corresponding to highly complex regions;

[0028] threshold and The accuracy of the model is dynamically set according to the task's requirements.

[0029] As a further aspect of the present invention, the simplified compression of the triangular mesh adopts an edge folding algorithm based on a quadratic error metric, wherein the error cost from folding an edge to a new vertex is determined by the sum of the quadratic error matrices of the two endpoints of the edge.

[0030] The wavelet transform compression includes threshold quantization of the high-frequency detail subband coefficients, setting coefficients whose absolute values ​​are less than a preset quantization threshold to zero.

[0031] The sparse coding compression uses an octree structure for spatial partitioning and encoding of the geometric point cloud in the detail layer, and employs product quantization technology to compress the high-dimensional feature descriptors.

[0032] As a further aspect of the present invention, S300 includes:

[0033] Each UAV periodically broadcasts its own generated feature layer metadata digest to neighboring nodes through a self-organizing wireless mesh network, and receives digest information from other UAVs in order to maintain and update the global digest table locally.

[0034] The cluster is based on a distributed consensus algorithm, which scores each drone based on its real-time status and dynamically recommends the drone with the highest score as a temporary coordination node.

[0035] The comprehensive scoring function is:

[0036] ;

[0037] for The overall score of the drone; Weighted by remaining battery power; As a weight for processor idle rate; Weights are based on proximity. The weights are random perturbations. For the first The remaining battery percentage of the drone; For the first The current idle rate of the drone's processor; For the first Normalized distance from the drone to the geometric center of the cluster; For random disturbance terms;

[0038] Temporary coordination nodes aggregate and analyze metadata summaries from the entire cluster, divide the task area into virtual grids, identify undersampled grids, potentially redundant grids, and hotspots of interest based on the summary information, and generate task allocation instructions by combining the real-time position and status of each UAV.

[0039] The temporary coordination node broadcasts the generated task instruction set to the entire cluster. Each UAV receives the instruction and makes its own decision. If the evaluation is successful, it adjusts its adaptive sampling logic locally according to the instruction.

[0040] As a further aspect of the present invention, S400 includes:

[0041] The ground station receives compressed feature data streams transmitted by each UAV and feature layer metadata summaries obtained through collaborative exchange. It uses the spatiotemporal information embedded in the data streams to perform time synchronization, spatial alignment, and fine registration on all data. The received compressed feature data streams are decompressed, and corresponding fusion algorithms are executed on the structural layer, texture layer, and detail layer data respectively.

[0042] Specifically: For the structural layer data, multi-source triangulation fusion is performed to generate a seamless backbone terrain mesh; for the texture layer data, wavelet domain image fusion is performed to generate a globally consistent high-resolution orthophoto; for the detail layer data, sparse reconstruction fusion is performed, including feature matching, deduplication, and geometric embedding, to obtain a detail-enhanced 3D mesh; the fused high-resolution orthophoto is mapped onto the surface of the detail-enhanced 3D mesh, and for mesh areas covered by multiple original images, viewpoint-dependent texture fusion is used to finally generate and output a complete 3D terrain model containing geometric, texture, and attribute information.

[0043] A second aspect of this application provides an efficient terrain data acquisition system for unmanned aerial vehicles (UAVs) that employs adaptive sampling and cooperative compression, the system comprising:

[0044] The real-time entropy calculation and adaptive sampling module is configured to calculate the dynamic terrain entropy value in real time by each UAV based on multi-sensor fusion data, and trigger a high-density sampling mode when the dynamic terrain entropy value exceeds a dynamic threshold associated with the flight altitude. The sampling density and sensor configuration are adjusted according to the dynamic terrain entropy value, while the original terrain data is collected.

[0045] The feature layering and differential compression module is configured to dynamically divide the raw terrain data collected by each UAV into three layers of feature data: structure, texture, and detail according to the entropy value. It performs triangular mesh simplification compression, wavelet transform compression, and sparse coding compression on each layer of feature data to generate a compressed feature data stream. At the same time, it generates feature layer metadata summaries that characterize the spatial distribution and entropy value statistics of each layer of feature data.

[0046] The decentralized coordination module is configured to allow each UAV to exchange feature layer metadata summaries through a self-organizing network. Based on a distributed consensus algorithm, a temporary coordination node is dynamically elected with the real-time status of each UAV as input. The temporary coordination node generates and issues task allocation instructions based on the global feature layer metadata summaries. Each UAV adjusts its high-density sampling mode according to the instructions.

[0047] The layered data fusion and reconstruction module is configured to receive compressed feature data streams sent by various UAVs from the ground station. Based on the spatiotemporal labels embedded in the data streams and the feature layer metadata summaries obtained through exchange, it performs spatiotemporal registration. It then performs multi-source triangulation fusion on the registered structural layer data, wavelet domain image fusion on the texture layer data, and sparse reconstruction fusion on the detail layer data. Finally, it overlays the three-layer fusion results to generate a complete three-dimensional terrain model.

[0048] In summary, due to the adoption of the above technical solution, the beneficial technical effects of the invention are as follows:

[0049] Due to the adoption of the above technical solution, the adaptive sampling and collaborative compression UAV efficient terrain data acquisition method and system described in this invention have the following outstanding advantages compared with the prior art:

[0050] First, it significantly improves overall data acquisition efficiency and model quality, solving the problem of rigid sampling. This invention introduces a dynamic terrain entropy value calculation in real time and an adaptive sampling mechanism, enabling each UAV to dynamically switch sampling modes and adjust sampling density based on the actual complexity of the terrain below it. This ensures automatic high-density focused sampling in key areas with complex terrain, preserving fine details; while in flat and simple areas, an economical cruise mode is used, greatly reducing redundant data. The on-demand sampling strategy fundamentally avoids the problem of data redundancy and information loss coexisting in existing technologies, greatly improving overall data acquisition efficiency while ensuring the accuracy of the 3D reconstruction model.

[0051] Secondly, it effectively alleviates data transmission pressure and breaks through the communication bandwidth bottleneck. This invention performs dynamic feature layering based on entropy values ​​on the raw data collected at the UAV end, and applies differentiated, high-efficiency compression algorithms to the characteristics of each layer. These high-efficiency compression algorithms can significantly compress massive amounts of raw terrain data while preserving core information, generating a feature data stream with a significantly reduced volume. This fundamentally reduces the communication link load when multiple UAVs transmit data back to the ground station, lowers transmission latency and packet loss risks, lays a solid foundation for real-time or near-real-time data transmission and processing, and solves the bandwidth bottleneck problem caused by raw data transmission in existing technologies.

[0052] Third, it significantly enhances the robustness and flexibility of system collaboration, overcoming the inherent defects of centralized architectures. This invention constructs a decentralized collaboration mechanism based on feature metadata exchange. Each UAV exchanges lightweight metadata summaries through a self-organizing network and dynamically elects temporary coordinating nodes based on real-time status to generate task allocation suggestions, which are ultimately executed autonomously by each UAV. The distributed architecture eliminates the risk of single-point failures, enabling the cluster to dynamically respond to sudden situations of individual UAVs and intelligently identify undersampled or redundant areas globally for task reallocation. This greatly improves the overall robustness, adaptability, and task completion efficiency of multi-UAV systems in complex and dynamic environments, effectively solving the vulnerability problems of existing centralized or master-slave collaborative architectures.

[0053] Fourth, it achieves rapid and high-quality data fusion and 3D reconstruction, improving post-processing efficiency. This invention employs a layered data fusion and reconstruction process at the ground station that matches the front-end acquisition, compression, and collaboration. By utilizing the spatiotemporal tags embedded in the data stream and the metadata digests obtained through exchange for efficient registration, and performing targeted fusion algorithms on the structural, texture, and detail layers of data respectively, the three layers of results are finally superimposed to generate a complete, seamless, and high-precision 3D terrain model. This ground processing workflow fully utilizes the results of front-end processing, achieving an efficient closed loop from data acquisition, compression, transmission to fusion and reconstruction, significantly shortening the overall time from data acquisition to model generation, and ensuring high quality in geometric accuracy and texture consistency of the final model. It solves the problems of long data processing flows and poor fusion effects in existing technologies. Attached Figure Description

[0054] Figure 1 A flowchart illustrating an efficient terrain data acquisition method for UAVs that employs adaptive sampling and collaborative compression.

[0055] Figure 2 A flowchart of S100, a method for efficient terrain data acquisition by UAVs that employs adaptive sampling and collaborative compression;

[0056] Figure 3 A flowchart of S200, a method for efficient terrain data acquisition by UAVs that employs adaptive sampling and collaborative compression;

[0057] Figure 4 A flowchart of S300, an efficient terrain data acquisition method for UAVs that employs adaptive sampling and collaborative compression;

[0058] Figure 5 This is a flowchart of S400, an efficient terrain data acquisition method for UAVs that employs adaptive sampling and collaborative compression. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the invention clearer, the invention 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 invention.

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0061] The core logic of this invention is:

[0062] This invention proposes an efficient UAV terrain data acquisition method and system based on adaptive sampling and collaborative compression. Its core logic lies in constructing a technical framework that uses dynamic terrain entropy as a unified quantification standard, and integrates front-end intelligent sampling, airborne layered compression, cluster collaborative scheduling, and ground-based fusion reconstruction across the entire data acquisition chain.

[0063] First, by using multi-sensor data to calculate dynamic terrain entropy values ​​in real time, and automatically triggering and adjusting sampling density based on a threshold associated with flight altitude, data acquisition shifts from uniform and rigid to adaptive based on terrain complexity, reducing redundant data from the source.

[0064] Second, based on the same entropy index, the original data is intelligently divided into three layers: structure, texture, and detail. For each layer of features, the optimal compression algorithm is used for efficient airborne compression to alleviate the pressure on transmission bandwidth.

[0065] Third, by exchanging lightweight metadata summaries, the drone swarm dynamically elects temporary coordinating nodes based on real-time status and intelligently allocates tasks according to the global data coverage situation, forming a self-organizing, highly elastic collaborative network and avoiding single points of failure.

[0066] Fourth, guided by metadata, the ground station performs targeted fusion algorithms on the received three-layer compressed data to synthesize a high-precision, seamless 3D terrain model, ensuring reconstruction quality.

[0067] This invention achieves systematic optimization of data efficiency, transmission efficiency, system robustness, and final modeling quality in multi-UAV collaborative operations by integrating the core metric of dynamic terrain entropy into all stages from data generation, processing, scheduling to the final product.

[0068] like Figure 1 As shown, it illustrates the overall flow of an exemplary UAV-based efficient terrain data acquisition method based on adaptive sampling and cooperative compression, specifically including the following steps:

[0069] S100. Real-time entropy calculation and adaptive sampling;

[0070] S200. Feature layering and differential compression;

[0071] S300. Decentralized collaboration based on feature metadata;

[0072] S400. Layered data fusion and reconstruction.

[0073] Specifically, the following steps are included:

[0074] S100. Adaptive sampling driven by multi-sensor data fusion and real-time entropy calculation:

[0075] The purpose of this step is to enable each drone to autonomously and in real time assess the complexity of the terrain below it, and dynamically adjust its data acquisition strategy accordingly, so as to minimize global data redundancy while ensuring the accuracy of data in key areas.

[0076] Please refer to Figure 2 It illustrates an exemplary flowchart of S100 real-time entropy calculation and adaptive sampling for an efficient UAV terrain data acquisition method based on adaptive sampling and cooperative compression. The specific steps are as follows:

[0077] S110. Multi-sensor data synchronous acquisition and fusion: During flight, the UAV synchronously acquires lidar point cloud, multispectral / hyperspectral camera images, and pose data of inertial measurement unit (IMU); through pre-calibrated sensor extrinsic parameters and time synchronization mechanism, the data of different modes are initially fused to form raw data blocks with spatial position and preliminary attribute information;

[0078] S120. Real-time calculation of local dynamic terrain entropy: For data within the current perception window (such as the area covered by the most recent N point cloud frames or image frames), three core terrain complexity metrics are calculated in parallel:

[0079] Calculate the local elevation variation coefficient The elevation values ​​of all laser points in the area are extracted. First, the standard deviation of these elevation values ​​is calculated, and then divided by their arithmetic mean to obtain the dimensionless coefficient of variation. The coefficient of variation effectively eliminates the influence of absolute elevation and purely reflects the relative undulation of the terrain. The coefficient of variation in flat areas approaches zero, while the coefficient of variation in canyons and steep slopes increases significantly.

[0080] Estimating the spatial rate of change of surface curvature Based on the point cloud within the window, principal component analysis is used to estimate the normal vector and curvature of the neighborhood surface for each point. Then, the gradient vector of the curvature value on the two-dimensional horizontal plane is calculated, and the average value of the gradient magnitude of all points within the entire window is obtained. The surface curvature spatial change rate index is sensitive to abrupt changes in terrain curvature and can effectively identify feature lines such as ridgelines and gully edges.

[0081] Extracting multispectral texture information entropy The multispectral image region corresponding to the window is divided into several small blocks. For each small block, the entropy of the numerical distribution of all its pixels in the multidimensional spectral space is calculated. The entropy value is low in areas with homogeneous land cover, while the entropy value is high in areas containing multiple land features or with fragmented structures.

[0082] Finally, the three core terrain complexity indicators are linearly weighted and fused to generate a comprehensive dynamic terrain entropy value. The weighting coefficients are pre-set according to the task type. For example, geometric features are given higher weight in geomorphological mapping, while spectral features are given higher weight in ecological environment monitoring.

[0083] Dynamic terrain entropy The calculation formula is:

[0084] ;

[0085] The weighting coefficients are configured according to task requirements and satisfy the following conditions: ;

[0086] S130. Adaptive Sampling Decision and Execution Based on Entropy Threshold Comparison:

[0087] A dynamic threshold function related to flight altitude is preset. The dynamic threshold function indicates that when flying at low altitude, the system is more sensitive to terrain complexity, the threshold is lower, and fine sampling is more easily triggered. As the flight altitude increases, the threshold is increased accordingly to adapt to a more macroscopic observation perspective.

[0088] Dynamic threshold function The formula is expressed as:

[0089] ;

[0090] This is the ground reference threshold; The attenuation coefficient; Flight altitude;

[0091] The dynamic threshold function shows that when flying at low altitudes, the system is more sensitive to terrain complexity, with a lower threshold that is more likely to trigger fine sampling; as the flight altitude increases, the threshold increases accordingly to adapt to a more macroscopic observation perspective.

[0092] Real-time comparison of the calculated dynamic terrain entropy values Dynamic threshold at current height If the entropy value continues to exceed the threshold, that is... If the current flying area is determined to be a complex terrain area or a region of critical interest, the system will automatically switch from the standard economic cruise sampling mode to the high-density focus sampling mode.

[0093] After entering high-density sampling mode, the system dynamically adjusts multiple acquisition parameters based on the extent to which the entropy value exceeds the threshold.

[0094] First, increasing the scanning frequency of the lidar or reducing the shooting interval of the multispectral camera will increase the sampling density of the core sensor by a certain percentage. They show a positive linear correlation.

[0095] The relative increase in sampling density Determined by the following formula:

[0096] ;

[0097] ;

[0098] in, It is a proportional adjustment coefficient greater than zero, whose value is preset during system calibration based on sensor performance and task requirements. The actual sampling density is increased relative to the ratio. Amplitude limiting is required:

[0099] ;

[0100] in, The maximum allowable sampling density increase ratio is determined by the physical limits of the sensor or the load capacity of the platform.

[0101] Based on the calculated relative increase in sampling density Dynamically increase the scanning frequency of the lidar or decrease the shooting interval of the multispectral camera to bring it up to the original reference value. First, adjust the gimbal angle to align the camera with the side of the terrain for more information, or adjust the deflection range of the laser scanning mirror. Simultaneously, appropriately reduce the drone's flight speed to ensure the sensors have sufficient dwell time in critical areas.

[0102] Before being stored in the cache, all raw data blocks are tagged with complete metadata, including a precise GNSS timestamp, a unique UAV identifier, the current 3D attitude, and the calculated entropy value. The current sampling mode identifier, the relative increase ratio of sampling density, and sensor configuration parameters.

[0103] Please refer to Figure 3 The diagram illustrates an exemplary S200 feature layering and differential compression flowchart of this application.

[0104] Step S200 runs on various UAV platforms and aims to perform intelligent preprocessing on the collected raw massive data, stratify it according to its information content and feature type, and apply the most suitable compression algorithm to different layers to significantly reduce the data volume while retaining key information, thus laying the foundation for efficient wireless transmission.

[0105] Specifically, it includes:

[0106] S210. Dynamic feature hierarchies based on entropy:

[0107] Read the raw data block from the cache and its associated dynamic terrain entropy value. There are two preset dynamic entropy threshold values. and ,satisfy This is used to dynamically divide the collected raw terrain data into three layers of feature data—structure, texture, and detail—based on entropy values; the layering rules are as follows:

[0108] Structural layer data: composed of entropy values The data consists of blocks, with the structural layer data corresponding to flat or gentle slope areas with gentle terrain undulations and simple features. Its core information is to describe the macroscopic outline and overall skeleton of the terrain. The absolute requirements for geometric accuracy are relatively low, but a large range of continuity must be maintained.

[0109] Texture layer data: satisfied by entropy values The data consists of blocks, with texture layer data corresponding to moderately complex terrain, such as hills and continuous slopes. Its characteristics include obvious changes in land cover, spectral features, and medium-scale geometric undulations, which constitute the main body of the visual realism of the terrain.

[0110] Detail layer data: composed of entropy values The data consists of blocks, with detailed layer data corresponding to highly complex areas such as cliffs, gullies, buildings, and rock outcrops. It contains key subtle geometric changes, edge and corner features, as well as rich hyperspectral texture information, which are crucial to the precision and accuracy of 3D reconstruction.

[0111] In one possible implementation, the threshold and The accuracy of the model can be dynamically set according to the task's requirements. For example, in a detailed modeling task, the accuracy can be appropriately reduced so that more data can be included in the detail layer.

[0112] S220. Hierarchical Feature Data Extraction and Representation: For each layer of data after partitioning, adaptive feature extraction and data format conversion are required to prepare for subsequent compression; among which:

[0113] Structural layer feature extraction: For the structural layer point cloud, voxel grid-based downsampling and normal vector estimation are performed to construct an incremental triangular mesh and extract key terrain feature points such as mountain peaks and saddle points to form structured triangular mesh data;

[0114] Texture layer feature extraction: For the multispectral image patch corresponding to the texture layer, apply two-dimensional discrete wavelet transform. This transform decomposes the image into a low-frequency approximation sub-band that carries the main texture information and a high-frequency detail sub-band that captures details, forming a set of wavelet coefficients.

[0115] Detail layer feature extraction: For the detail layer point cloud and the corresponding image patch, extract geometric features such as edge points and corner points based on curvature, as well as local visual feature descriptors such as SIFT or ORB, to form a sparse feature set;

[0116] S230. Differentiated compression algorithm execution, the compression algorithm is customized for the characteristics of each layer of data:

[0117] Structural layer data compression: For irregular triangular meshes in the structural layer, an edge folding simplification algorithm based on quadratic error metric is adopted. The edge folding simplification algorithm based on quadratic error metric iteratively merges vertex pairs that have the least impact on the mesh geometry and records the simplification operation sequence, thereby significantly reducing the number of triangular faces within a controllable geometric error and generating a simplified mesh data stream.

[0118] For the quadratic error metric moment of a vertex, the error cost from the folded edge to the new vertex is:

[0119] ;

[0120] in, Let be the homogeneous coordinate vector of the new vertex. In the edge folding operation, the selected edge will be folded to a new vertex position, which is the coordinate of this new vertex. The folding error cost function measures the cost of folding an edge to a new vertex. When the folding operation causes a total error to the geometry of the 3D model, the smaller the value, the less impact the folding operation has on the model's accuracy. for The transpose of converts a column vector into a row vector, which is used for matrix multiplication operations of quadratic forms; Let be the quadratic error matrix of one endpoint of the edge; Let be the quadratic error matrix of the other endpoint of the edge;

[0121] For wavelet coefficients in the texture layer, different treatments are applied: low-frequency subbands use lossy compression with high fidelity, such as JPEG2000 encoding; high-frequency subbands first use threshold quantization to set coefficients with smaller amplitudes to zero. The quantization operation is as follows:

[0122] ;

[0123] These are the quantized wavelet coefficients, the result of thresholding the original wavelet coefficients, with some coefficients set to zero to achieve compression. These are the original wavelet coefficients, obtained by performing wavelet transform on an image or signal, containing low-frequency approximation information and high-frequency detail information; The absolute value of the wavelet coefficient is used to measure the importance of the information carried by the coefficient; the larger the absolute value, the more important the information. The wavelet compression threshold for the texture layer is a preset positive number used to determine which coefficients are retained and which are set to zero. If so, then retain that coefficient (considered important information); if If so, then set the coefficient to zero;

[0124] Detail layer data compression: For sparse features in the detail layer, the geometric point cloud is spatially partitioned and encoded using an octree structure, recording only the node information containing feature points; the high-dimensional feature descriptor is divided into multiple sub-vectors using product quantization technology and quantized onto a pre-trained codebook, with short codeword indices replacing the original vectors to achieve high compression ratio.

[0125] S240. Feature Layer Metadata Summary Generation: During compression, a metadata summary describing the data at each layer is generated synchronously. The summary contains key descriptive information:

[0126] The spatial extent of the data coverage is usually represented by the corner coordinates of the axial bounding box; the statistical characteristics of the data block entropy, such as mean, maximum and standard deviation; the data size information, including the original size, compressed size and compression ratio; the key parameters of the compression algorithm used; and spatiotemporal tags such as the data acquisition timestamp and UAV identification.

[0127] After compression, each UAV will generate three independent compressed feature data streams and corresponding feature layer metadata summaries.

[0128] Please refer to Figure 4 This illustrates an exemplary S300 of this application.

[0129] The goal of step S300 is to build a decentralized autonomous collaborative mechanism for drone swarms. Each drone perceives the global mission situation by exchanging lightweight metadata summaries, dynamically elects a temporary coordinator to generate task allocation suggestions, and finally, each drone makes autonomous decisions to execute them, thereby improving overall sampling efficiency and system robustness.

[0130] Specifically, it includes:

[0131] S310. Metadata digest exchange:

[0132] Each drone periodically broadcasts its generated feature layer metadata digest to neighboring nodes via a self-organizing wireless mesh network. At the same time, it continuously listens to the network and receives digest information from other drones. The network routing protocol ensures that the digest information can still be effectively propagated within the cluster when drone movement causes topology changes. Each drone maintains a global digest table locally to store and update the data coverage and status information of all known companions.

[0133] S320. Dynamic election of temporary coordinating nodes:

[0134] The cluster uses a distributed consensus algorithm to dynamically elect a temporary coordinating node, rather than designating a fixed leader. The election process is triggered periodically or when the coordinating node fails. The election is based on a comprehensive score of each drone's real-time status, with factors including remaining battery percentage, current processor idle rate, and proximity to the cluster's geometric center. Small random perturbations are introduced to avoid deadlock. The comprehensive scoring function is:

[0135] ;

[0136] for The overall score of the drone is used to evaluate whether the drone is suitable to serve as a temporary coordination node. The higher the score, the more likely it is to be selected as the coordinator. Weighted by remaining battery power; As a weight for processor idle rate; Weights are based on proximity. The weights are random perturbations. For the first The remaining battery percentage of the drone; the higher the battery percentage, the more suitable it is to serve as a coordination node. For the first The current idle rate of the drone's processor; For the first Normalized distance from the drone to the geometric center of the cluster; For the random perturbation term, a small random number is used to avoid election deadlock;

[0137] The drone with the highest overall score will assume the role of a temporary coordinating node, responsible for coordinating cluster tasks in the next cycle.

[0138] S330. Global task allocation instruction generation:

[0139] The selected interim coordinating node aggregates and analyzes metadata summaries from the entire cluster. It divides the task area into virtual grids on a horizontal plane and analyzes the state of each grid based on the summary information: identifying undersampled grids with high entropy values ​​but lacking detailed layer data coverage, marking potentially redundant grids that are repeatedly sampled at high density by multiple drones, and locating the hotspots of interest with the highest global entropy values. Combining the real-time positions, remaining endurance, and task load of each drone, the coordinating node generates a series of specific task allocation instructions; each instruction explicitly specifies the target grid, the expected action such as focused sampling or avoidance, the suggested sampling density adjustment factor, the task priority, and a list of drones recommended to perform the task.

[0140] S340. Instruction Distribution and Local Execution Adjustment:

[0141] The temporary coordination node broadcasts the generated task instruction set to the entire cluster. Upon receiving the instructions, each drone does not obey unconditionally but makes its own decisions, assessing whether it is on the list of suggested executors and comprehensively considering the compatibility of the instruction target with its current flight path, remaining battery power, and sensor status. If the assessment is successful, the drone will adopt the instruction and adjust its adaptive sampling logic accordingly, such as temporarily modifying its entropy trigger threshold when flying towards the target grid to ensure that the sampling actions required by the instruction are executed. The drone can also report the instruction execution status back to the network.

[0142] Please refer to Figure 5 This illustrates an exemplary S400 of this application.

[0143] Step S400 is completed at the ground station, which is responsible for receiving, decompressing, registering and fusing the layered and compressed data streams from multiple UAVs, and finally reconstructing a unified, complete and high-precision three-dimensional terrain model.

[0144] Specifically, it includes:

[0145] S410. Data reception and spatiotemporal registration:

[0146] The ground station receives compressed feature data streams transmitted by each UAV in parallel, along with feature layer metadata summaries obtained through collaborative exchange. First, using the precise GNSS timestamps and UAV pose information embedded in the data streams, all data is time-synchronized and roughly spatially aligned. Then, fine-grained spatiotemporal registration is performed.

[0147] Registration based on structural layer features: Significant terrain feature points (such as mountain peaks and saddle points) in the triangulation of each structural layer are used to perform feature matching in overlapping areas. The fine relative pose transformation parameters between different data sources are solved by the iterative nearest point algorithm, and the error is distributed to each data source for overall adjustment optimization, finally obtaining the accurate transformation matrix from each UAV coordinate system to the global coordinate system.

[0148] Detail layer feature-assisted verification: High-precision feature points (such as SIFT feature points) of the decompressed detail layer are used to perform secondary matching in the overlapping area to verify and fine-tune the registration results of the structure layer, ensuring sub-pixel level registration accuracy.

[0149] S420. Layered Data Decompression and Fusion:

[0150] S421. Structural Layer Fusion - Multi-Source Triangular Network Fusion:

[0151] The compressed data streams of each structural layer are decompressed to restore the simplified triangulation meshes collected by each UAV. These triangulation meshes are then transformed to the global coordinate system using registration transformation; a multi-source mesh fusion algorithm based on constrained Delaunay triangulation is employed.

[0152] Align the boundaries of triangular meshes from different sources in overlapping areas;

[0153] The overlapping areas are uniformly triangulated using Delaunay triangulation to eliminate mesh cracks;

[0154] Merge non-overlapping regions directly;

[0155] The final result is a seamless backbone terrain mesh covering the entire region.

[0156] S422. Texture Layer Blending - Wavelet Domain Image Blending:

[0157] The compressed data stream of the texture layer is decompressed, the wavelet coefficients of each image patch are recovered, and wavelet domain fusion is performed in the registered image space.

[0158] Low-frequency subband fusion employs a weighted average method to fuse the low-frequency approximation subband coefficients.

[0159] ;

[0160] The low-frequency subband coefficients after fusion represent the position. The low-frequency approximation subband coefficient value after weighted average fusion; Represents the coordinate position of wavelet coefficients, and the row index in the low-frequency subband coefficient matrix. and column indexes ; Indicates the number of drones participating in the integration; For drone indexes, the value range is 1 to... This is used to iterate through all data sources involved in the fusion process; for The fusion weight of each data source; For the first Data sources in location The low-frequency subband coefficient at that location;

[0161] High-frequency subband fusion uses the method of maximizing the absolute value of the coefficients of high-frequency detail subbands to retain the clearest edge details.

[0162] ;

[0163] The high-frequency subband coefficients after fusion represent the values ​​at position. The high-frequency detail subband coefficient values ​​after fusion using the absolute value taking method; These are the high-frequency subband coefficients after fusion; For drones with the highest absolute value high-frequency coefficient; Indicates the first Data sources in location High-frequency subband coefficient at the location;

[0164] The fused wavelet coefficients are inversely transformed to generate a globally consistent high-resolution orthophoto image.

[0165] S423. Detail Layer Blending - Sparse Reconstruction Blending:

[0166] Decompress the compressed data stream at the detail layer to recover geometric feature points and visual descriptors:

[0167] Feature matching and deduplication: For feature points from different sources, deduplication is performed using descriptor matching within a spatial proximity threshold.

[0168] ;

[0169] To match the judgment function, determine the two feature points. and Does it match? For feature descriptors and The Euclidean distance between them; For descriptor distance threshold; Representing feature points and The Euclidean distance between spatial locations; This is the spatial distance threshold; This indicates a match between two feature points; The two feature points do not match;

[0170] The average position of the matched feature points is taken as the fused feature point.

[0171] Detailed geometric embedding: Using the fused high-precision feature point set as a constraint, local geometric enhancement is performed in the key areas of the backbone mesh through constrained Delaunay triangulation or moving least squares surface fitting method to obtain a detailed enhanced 3D mesh.

[0172] S430. Model Overlay and Generation:

[0173] Texture mapping: The fused high-resolution orthophoto is mapped onto the detail-enhancing mesh surface. The texture coordinates of each vertex are calculated based on the camera model and registration parameters to establish a texture-geometry correspondence.

[0174] Multi-image fusion region processing: For grid regions covered by multiple original images, viewpoint-dependent texture fusion is employed.

[0175] ;

[0176] in, Indicates the texture coordinates after fusion. The final texture pixel value after fusion; Indicates image index; Indicates the first The fusion weight of Zhang's images, fusion weight The angle between the line of sight and the grid normal vector, image quality, etc., are used to determine the angle. Show the first Zhang Image in Texture Coordinates Pixel value at;

[0177] Model output: The final output is a complete 3D terrain model containing geometric, texture, and attribute information. The 3D terrain model is output in multiple standard formats, supporting visualization, analysis, and application.

[0178] Example 2

[0179] An efficient terrain data acquisition system for UAVs with adaptive sampling and cooperative compression includes:

[0180] Real-time entropy calculation and adaptive sampling module:

[0181] Dynamic terrain entropy calculation unit: calculates local elevation variation coefficient in real time. Estimate the rate of change of surface curvature in space. Extracting multispectral texture information entropy And merged into dynamic terrain entropy value ;

[0182] Adaptive Decision and Execution Unit: Comparing Dynamic Terrain Entropy Values With dynamic threshold function The sampling density, sensor parameters, and flight speed are dynamically adjusted.

[0183] 2. Feature layering and differential compression module:

[0184] Entropy hierarchical unit: based on two preset dynamic entropy thresholds and The data is divided into three layers: structure, texture, and detail.

[0185] Hierarchical feature extraction unit:

[0186] Structural layer: Triangular mesh construction and feature point extraction;

[0187] Texture layer: wavelet transform and coefficient decomposition;

[0188] Detail layer: Curvature feature and visual descriptor extraction;

[0189] Differentiated compression unit:

[0190] Structural layer: Simplified edge folding based on QEM;

[0191] Texture layer: Wavelet coefficient threshold quantization and encoding;

[0192] Detail layer: Octree coding and product quantization;

[0193] 3. Decentralized Collaboration Module:

[0194] Metadata digest exchange unit: broadcasting and receiving feature layer metadata digests based on self-organizing networks.

[0195] Dynamic coordination node election unit: based on Establish temporary coordination nodes.

[0196] Task allocation and adjustment unit: Coordinates nodes to analyze the global summary, generates and issues task instructions, and each machine makes autonomous decisions to execute them.

[0197] 4. Layered Data Fusion and Reconstruction Module:

[0198] Spatiotemporal registration unit: Fine registration is achieved by matching GNSS timestamps with feature points.

[0199] Layered fusion unit:

[0200] Structural layer: Multi-source triangular network fusion;

[0201] Texture layer: Wavelet domain weighted average and maximum value fusion;

[0202] Detail layer: Feature matching and sparse reconstruction;

[0203] Model generation and output unit: texture mapping, multi-image fusion, outputting a standard 3D terrain model.

[0204] 5. Hardware Platform:

[0205] Unmanned aerial vehicle (UAV) carrier: equipped with high-precision positioning, a stable gimbal, and a multi-sensor payload.

[0206] Onboard processor: supports real-time entropy calculation, hierarchical compression, and collaborative decision-making.

[0207] Communication equipment: Supports self-organizing mesh networks and air-to-ground data transmission.

[0208] Ground station: It has the ability to receive multiple data streams, decompress in parallel, and fuse and reconstruct data.

[0209] Example 3

[0210] Application example of this invention in a topographic surveying task in a certain mountainous area

[0211] 1. Task Area and Equipment Configuration Area: A typical hilly-mountain transition zone in Southwest China was selected as the test area, covering an area of ​​approximately 10 km². 2 The region has a complex topography, consisting of gentle valleys (about 30%), continuous hills (about 50%), and steep gullies and rock outcrops (about 20%).

[0212] Equipment: Deploy 4 rotary-wing drones (model: DJI Motrice 350RTK), each carrying:

[0213] LiDAR: RoboSenseRS-LiDAR-16 (scanning frequency 10–20Hz, accuracy ±2cm)

[0214] Multispectral camera: Parrot Sequoia+ (5 bands: green, red, red edge, near-infrared, RGB)

[0215] Positioning and attitude determination system: NovAtelPwrPak7D-E2 (GNSS / IMU tight combination, positioning accuracy 1cm+1ppm horizontally, 2cm+1ppm vertically)

[0216] Onboard computing unit: NVIDIA Jetson AGX Orin (32GB memory, 200 TOPS computing power)

[0217] Communication: The drones use a Mesh self-organizing network based on IEEE 802.11ax, and the air-to-ground link uses 4G / 5G dual-mode communication.

[0218] Preset parameters (based on task type: high-precision topographic mapping):

[0219] Entropy value is used to calculate weights: ;

[0220] Dynamic threshold parameters: ; ;

[0221] Feature stratification threshold: , ;

[0222] Election weights: , , , ;

[0223] Compression parameters: Simplified structure layer retains 30% of target surface area; texture layer wavelet compression threshold. The octree depth of the detail layer is 10, and the product quantization codebook size is 256.

[0224] 2. Implementation process and data recording

[0225] Adaptive sampling process:

[0226] Flat valley area: Flight altitude 120m, calculated average Below the threshold The system maintains the economic cruise mode, with the lidar scanning frequency at 10Hz and the camera shooting interval at 2s.

[0227] In gully and steep cliff areas: the flight altitude was reduced to 80m, and the following measurements were taken. It rose to 0.28, exceeding the threshold. The system automatically switches to high-density focused sampling mode. Calculated according to the formula:

[0228] ;

[0229] After being limited ,have to The laser radar frequency was increased to approximately 10.7 Hz, and the camera interval was shortened to approximately 1.87 seconds. Simultaneously, the flight speed was reduced by 30%, and the gimbal was adjusted to -30° pitch to collect lateral information.

[0230] Layered compression effect (taking a typical data block as an example):

[0231] Feature layer Original data size (MB) Compressed data size (MB) Compression ratio Key parameters Structural layer 150.2 45.1 3.33:1 Simplification rate 70%, QEM error <0.1m Texture layer 320.5 21.4 14.97:1 Wavelet threshold, PSNR>38dB Detail layer 85.7 4.3 19.93:1 The octree has a depth of 10 and a feature point retention rate of 96.2%.

[0232] Decentralized Collaboration Event Log: 25 minutes after the task begins, the cluster conducts a dynamic election (based on comprehensive scoring). (Calculation) Determine UAV-2 as a temporary coordination node.

[0233] UAV-2 analyzed the global metadata summary and found that grid G-07 (located on an unplanned steep cliff) had an entropy value as high as 0.31 but a detail layer coverage of 0, indicating it was an undersampled grid. An instruction was then generated, assigning the nearest UAV-4 to perform focused sampling.

[0234] After receiving the instruction, UAV-4 autonomously adjusted its course, temporarily lowering its local entropy trigger threshold by 20% during its flight towards G-07 to ensure high-density sampling of the area. The entire process, from event recognition to adjustment, took approximately 12 seconds.

[0235] Ground convergence and reconstruction: All data is back to the ground station via 5G links (total data transmission volume is approximately 4.7GB, a reduction of 64.4% compared to the original estimated data volume of 13.2GB).

[0236] Spatiotemporal registration accuracy: Based on feature point matching and ICP optimization, the RMS residual of multi-source point cloud registration is 0.037m. Fusion result: Structural layer fusion generates a global continuous triangular network with approximately 1.2 × 10⁻⁶ vertices. 7 The number of dough pieces is approximately 2.4 × 10. 7 Texture layer fusion generates an orthophoto with a ground resolution of 0.08m, and no obvious seams or ghosting are observed in overlapping areas. Detail layer fusion adds approximately 5.3×10 feature points to areas such as gullies and cliffs. 5 The backbone mesh is embedded using MLS surface fitting, achieving a local detail resolution of 0.02m. The final model is a 3D terrain model in OSGB format, approximately 8.3GB in size, containing geometric, texture, and layered attribute information.

[0237] 3. Quantitative comparison of results (compared with traditional fixed-route uniform sampling method):

[0238] Performance indicators Traditional methods Method of the present invention Improvement effect Total data collection time 4.5 hours 3.2 hours Reduced by 28.9% Data transfer volume 13.2 GB 4.7 GB Reduced by 64.4% Model reconstruction time 6.8 hours 2.1 hours Reduced by 69.1% Richness of detail (density of feature points) in key areas (steep cliffs, gullies) 12.5 points / m² 28.7 points / m² An increase of 129.6% Data redundancy rate in flat areas 85.2% 32.6% Reduced by 52.6% Overall model geometric accuracy (RMSE) 0.25 m 0.12 m An increase of 52.0% Visual consistency of texture (SSIM index) 0.78 0.91 An increase of 16.7% System robustness (task completion rate after single-machine failure) 65% 92% An increase of 41.5%

[0239] in conclusion

[0240] In this typical mountainous terrain mapping task, the method and system proposed in this invention demonstrate significant advantages:

[0241] Efficiency improvement: Through adaptive sampling and hierarchical compression, the total task time is reduced by about 30%, and the data transmission burden is reduced by about 64%.

[0242] Quality Assurance: Detail richness in key areas increased by over 129%, and overall model accuracy improved by 52%.

[0243] Furthermore, it exhibits excellent visual consistency in texture.

[0244] System intelligence and robustness: The decentralized collaborative mechanism effectively responded to the data collection needs in steep cliff areas that were not planned in advance, and ensured 92% area coverage through autonomous task adjustment when the communication of a single machine was temporarily interrupted.

[0245] This embodiment fully verifies that the present invention achieves high efficiency, high precision, and high robustness in complex terrain environments.

[0246] The ability to acquire and reconstruct 3D terrain data has significant value for widespread application.

[0247] The sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0248] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for efficient UAV terrain data acquisition using adaptive sampling and collaborative compression, characterized in that, Includes the following steps: S100. Each UAV calculates the dynamic terrain entropy value in real time based on multi-sensor fusion data; when the dynamic terrain entropy value exceeds the dynamic threshold associated with the flight altitude, a high-density sampling mode is triggered, and the sampling density and sensor configuration are adjusted according to the dynamic terrain entropy value, while the original terrain data is collected. S200. Each UAV dynamically divides the raw terrain data it collects into three layers of feature data: structure, texture, and detail, based on entropy values. It then performs triangulation simplification compression, wavelet transform compression, and sparse coding compression on each layer of feature data to generate a compressed feature data stream. At the same time, it generates a feature layer metadata summary that represents the spatial distribution and entropy statistics of each layer of feature data. S300. Each UAV exchanges feature layer metadata digests through a self-organizing network. Based on a distributed consensus algorithm, a temporary coordination node is dynamically elected with the real-time status of each UAV as input. The temporary coordination node generates and issues task allocation instructions based on the global feature layer metadata digest. Each UAV adjusts its high-density sampling mode according to the instructions. S400. Layered Data Fusion and Reconstruction: The ground station receives compressed feature data streams sent by each UAV and performs spatiotemporal registration based on the spatiotemporal labels embedded in the data streams and the feature layer metadata summaries obtained through exchange. Multi-source triangulation fusion is performed on the registered structural layer data, wavelet domain image fusion is performed on the texture layer data, and sparse reconstruction fusion is performed on the detail layer data. Finally, the three-layer fusion results are superimposed to generate a complete three-dimensional terrain model.

2. The method for efficient UAV terrain data acquisition using adaptive sampling and collaborative compression according to claim 1, characterized in that, The dynamic terrain entropy value The calculation formula is: ； The weighting coefficients are configured according to task requirements and satisfy the following conditions: .

3. The method for efficient UAV terrain data acquisition using adaptive sampling and collaborative compression according to claim 1, characterized in that, The dynamic threshold function Represented as: ； This is the ground reference threshold; The attenuation coefficient; This refers to the flight altitude.

4. The method for efficient UAV terrain data acquisition using adaptive sampling and collaborative compression according to claim 1, characterized in that, After the high-density sampling mode is triggered, the sampling density is dynamically adjusted according to the extent to which the dynamic terrain entropy value exceeds the dynamic threshold, and the adjusted sampling density is subject to an upper limit limit.

5. The method for efficient UAV terrain data acquisition using adaptive sampling and collaborative compression according to claim 1, characterized in that, The steps by which each UAV dynamically divides the raw terrain data it collects into three layers of feature data—structure, texture, and detail—based on entropy values ​​include: Read the raw data block from the cache and its associated dynamic terrain entropy value. There are two preset dynamic entropy threshold values. and ,satisfy The hierarchical rules are as follows: Structural layer data: composed of entropy values The data blocks are composed of structural layer data, which correspond to flat or gently sloping areas with gentle terrain undulations and simple features. Texture layer data: satisfied by entropy values The data blocks are composed of texture layer data, which corresponds to moderately complex terrain. Detail layer data: composed of entropy values The data is composed of blocks, with the detailed layer data corresponding to highly complex regions; threshold and The accuracy of the model is dynamically set according to the task's requirements.

6. The method for efficient UAV terrain data acquisition using adaptive sampling and collaborative compression according to claim 1, characterized in that, The simplified compression of the triangular mesh adopts an edge folding algorithm based on quadratic error metric, wherein the error cost from folding an edge to a new vertex is determined by the sum of the quadratic error matrices of the two endpoints of the edge; The wavelet transform compression includes threshold quantization of the high-frequency detail subband coefficients, setting coefficients whose absolute values ​​are less than a preset quantization threshold to zero. The sparse coding compression uses an octree structure for spatial partitioning and encoding of the geometric point cloud in the detail layer, and employs product quantization technology to compress the high-dimensional feature descriptors.

7. The method for efficient UAV terrain data acquisition using adaptive sampling and collaborative compression according to claim 1, characterized in that, The S300 includes: Each UAV periodically broadcasts its own generated feature layer metadata digest to neighboring nodes through a self-organizing wireless mesh network, and receives digest information from other UAVs in order to maintain and update the global digest table locally. The cluster is based on a distributed consensus algorithm, which scores each drone based on its real-time status and dynamically recommends the drone with the highest score as a temporary coordination node. The comprehensive scoring function is: ； for The overall score of the drone; Weighted by remaining battery power; As a weight for processor idle rate; Weights are based on proximity. The weights are random perturbations. For the first The remaining battery percentage of the drone; For the first The current idle rate of the drone's processor; For the first Normalized distance from the drone to the geometric center of the cluster; For random disturbance terms; Temporary coordination nodes aggregate and analyze metadata summaries from the entire cluster, divide the task area into virtual grids, identify undersampled grids, potentially redundant grids, and hotspots of interest based on the summary information, and generate task allocation instructions by combining the real-time position and status of each UAV. The temporary coordination node broadcasts the generated task instruction set to the entire cluster. Each UAV receives the instruction and makes its own decision. If the evaluation is successful, it adjusts its adaptive sampling logic locally according to the instruction.

8. The method for efficient UAV terrain data acquisition using adaptive sampling and collaborative compression according to claim 1, characterized in that, The S400 includes: The ground station receives compressed feature data streams transmitted by each UAV and feature layer metadata summaries obtained through collaborative exchange. It utilizes the spatiotemporal information embedded in the data streams to perform time synchronization, spatial alignment, and fine registration on all data. The received compressed feature data streams are decompressed, and corresponding fusion algorithms are executed on the structural layer, texture layer, and detail layer data respectively. Specifically: For the structural layer data, multi-source triangulation fusion is performed to generate a seamless backbone terrain mesh; for the texture layer data, wavelet domain image fusion is performed to generate a globally consistent high-resolution orthophoto; for the detail layer data, sparse reconstruction fusion is performed, including feature matching, deduplication, and geometric embedding, to obtain a detail-enhanced 3D mesh. The fused high-resolution orthophoto is mapped onto the surface of the detail-enhanced 3D mesh. For mesh areas covered by multiple original images, viewpoint-dependent texture fusion is used to finally generate and output a complete 3D terrain model containing geometric, texture, and attribute information.

9. A high-efficiency terrain data acquisition system for unmanned aerial vehicles (UAVs) with adaptive sampling and cooperative compression, characterized in that, The system for implementing the method of any one of claims 1 to 8 comprises: The real-time entropy calculation and adaptive sampling module is configured to calculate the dynamic terrain entropy value in real time by each UAV based on multi-sensor fusion data, and trigger a high-density sampling mode when the dynamic terrain entropy value exceeds a dynamic threshold associated with the flight altitude. The sampling density and sensor configuration are adjusted according to the dynamic terrain entropy value, while the original terrain data is collected. The feature layering and differential compression module is configured to dynamically divide the raw terrain data collected by each UAV into three layers of feature data: structure, texture, and detail according to the entropy value. It performs triangular mesh simplification compression, wavelet transform compression, and sparse coding compression on each layer of feature data to generate a compressed feature data stream. At the same time, it generates feature layer metadata summaries that characterize the spatial distribution and entropy value statistics of each layer of feature data. The decentralized coordination module is configured to allow each UAV to exchange feature layer metadata summaries through a self-organizing network. Based on a distributed consensus algorithm, a temporary coordination node is dynamically elected with the real-time status of each UAV as input. The temporary coordination node generates and issues task allocation instructions based on the global feature layer metadata summaries. Each UAV adjusts its high-density sampling mode according to the instructions. The layered data fusion and reconstruction module is configured to receive compressed feature data streams sent by various UAVs from the ground station. Based on the spatiotemporal labels embedded in the data streams and the feature layer metadata summaries obtained through exchange, it performs spatiotemporal registration. It then performs multi-source triangulation fusion on the registered structural layer data, wavelet domain image fusion on the texture layer data, and sparse reconstruction fusion on the detail layer data. Finally, it overlays the three-layer fusion results to generate a complete three-dimensional terrain model.

Images