An intelligent obstacle avoidance system for drones based on lidar

Abstract

The present invention discloses an intelligent obstacle avoidance system for unmanned aerial vehicles (UAVs) based on laser radar, which belongs to the field of path planning technology. The system specifically comprises: a parameter acquisition module, which is used to acquire original point cloud data of a flight area; and performs obstacle geometric structure analysis to generate desensitized feature parameters including surface curvature and spatial density distribution; a feature analysis module, which is used to send the desensitized feature parameters to an edge computing node through a hierarchical encrypted transmission link, and perform initial aggregation of feature parameters of multiple UAVs in the same area to form a regional feature template; a cloud server, which is used to integrate obstacle features of different areas through a secure fusion protocol and construct a global obstacle feature library; a path planning module, which is used to obtain updated data of the global obstacle feature library through a secure synchronization channel, and generate a three-dimensional obstacle avoidance path by combining real-time scanning features with feature library matching results; the present invention realizes multi-machine collaborative obstacle avoidance path optimization based on obstacle avoidance information sharing.

Description

Technical Field

[0001] The present invention relates to the field of path planning technology, and in particular to an intelligent obstacle avoidance system for unmanned aerial vehicles based on laser radar. Background Art

[0002] With the widespread application of drones in fields such as agricultural plant protection, power inspections, and logistics and transportation, autonomous obstacle avoidance in complex environments has become a core requirement for flight safety. Traditional obstacle avoidance systems rely on single-machine LiDAR and visual sensors to construct local environmental maps through real-time perception. However, the operating range and data collection time of a single drone are limited, making it difficult to cover the entire lifecycle of obstacles, resulting in localized and time-sensitive environmental perception. In large-scale drone swarm operations, achieving multi-machine collaborative environmental perception and knowledge sharing has become a key challenge in improving the intelligence of obstacle avoidance systems.

[0003] Some existing solutions attempt to improve obstacle recognition capabilities through cloud-based data aggregation. For example, a centralized data processing architecture is employed to upload raw point cloud data from multiple drones to a cloud platform for unified modeling, and deep learning algorithms are used to construct an obstacle feature library. Other solutions propose local data sharing mechanisms based on edge computing, exchanging obstacle location information within a fleet of drones to optimize path planning.

[0004] In summary, the obstacle recognition model in existing technologies relies on static feature matching and is difficult to adapt to the morphological changes of dynamic obstacles, resulting in certain misjudgments of obstacle avoidance paths. Directly sharing raw point cloud data or obstacle coordinate information can easily expose sensitive information in the operating area, posing a risk of privacy leakage. Summary of the Invention

[0005] The purpose of this invention is to provide a UAV intelligent obstacle avoidance system based on laser radar to solve the following technical problems:

[0006] How to achieve multi-machine collaborative obstacle avoidance path optimization based on obstacle avoidance information sharing.

[0007] The purpose of the present invention can be achieved through the following technical solutions:

[0008] A laser radar-based intelligent obstacle avoidance system for unmanned aerial vehicles, comprising:

[0009] The parameter acquisition module uses the lidar sensors of multiple drones to synchronously collect raw point cloud data of the flight area. The local feature processing unit of each drone analyzes the geometric structure of obstacles in the raw point cloud data and generates desensitized feature parameters including surface curvature and spatial density distribution.

[0010] A feature analysis module is used to send the desensitized feature parameters to the edge computing node through a hierarchical encrypted transmission link. The edge computing node initially aggregates the feature parameters of multiple drones in the same area to form a regional feature template;

[0011] The cloud server is used to receive regional feature templates from multiple edge computing nodes, integrate obstacle features from different regions through a secure fusion protocol, and build a global obstacle feature library;

[0012] The path planning module is used to control each drone to obtain updated data from the global obstacle feature library through a secure synchronous channel, and generate a three-dimensional obstacle avoidance path by combining real-time scanning features with the feature library matching results.

[0013] As a further solution of the present invention: in the parameter acquisition module, the specific process of analyzing the obstacle geometry is as follows:

[0014] The original point cloud data is divided into three-dimensional cubic units of equal volume, and the gradient of the point cloud quantity change in each cubic unit is calculated to generate a spatial density distribution map;

[0015] Analyze the surface continuity characteristics between adjacent cubic units and identify rigid obstacles with regular geometric structures by calculating the direction difference values of the point cloud normal vectors between cubic units;

[0016] Extract the curvature mutation points at the boundaries of the cube units, construct the obstacle surface contour line, and distinguish independent obstacles from continuous obstacle groups based on the closed characteristics of the obstacle surface contour line;

[0017] The spatial density distribution map is fused and encoded with the obstacle surface contour features to generate a fixed-dimensional feature vector sequence. The feature vector sequence is subjected to spatial position confusion processing, and a random matrix permutation algorithm is used to eliminate the geographical location correlation of the original point cloud data.

[0018] As a further solution of the present invention: in the feature analysis module, the specific working process of the hierarchical encryption transmission link is:

[0019] Before sending the desensitized characteristic parameters, each drone uses the elliptic curve encryption algorithm to initially encrypt the data packet of the desensitized characteristic parameters, and the encrypted data packet is attached with an irreversible device anonymous identification code;

[0020] After receiving encrypted data packets from multiple drones in the same area, the edge computing node uses a hash chain-based key derivation method to generate a temporary decryption key, and then strips the device's anonymous identification code after decryption;

[0021] After receiving regional feature templates from different edge computing nodes, the cloud server uses a secure multi-party computing protocol to verify data integrity and employs homomorphic encryption technology to perform secondary obfuscation on cross-region features.

[0022] The obfuscated feature data is timestamped and regionally coded through a blind signature mechanism to generate a global feature set that cannot be traced back to the original device.

[0023] As a further solution of the present invention: In the cloud server, the specific process of constructing the global obstacle feature library is as follows:

[0024] Establish a cross-regional feature association model, compare the regional feature templates with the historical feature database for similarity, and identify obstacle categories with regional specificity;

[0025] Dynamic priority rules are established for recurring cross-regional obstacle features. Crop features that change over time in agricultural scenarios are assigned high update frequency weights, and equipment features with fixed spatial positions in power scenarios are assigned high matching accuracy weights.

[0026] Construct a three-dimensional feature projection space, map obstacle features in different areas to a unified coordinate system, and form an obstacle feature relationship network covering multiple operation scenarios. Establish spatiotemporal association links between obstacle features within the obstacle feature relationship network, record the morphological changes and position migration patterns of similar obstacles in different seasons, and update the obstacle morphology and position based on the current time point.

[0027] When the number of unidentified obstacles in a certain area reaches the characteristic library capacity ratio, the cloud server sends an enhanced scanning instruction to the drone group in the area, specifying the pitch angle change sequence and echo intensity acquisition parameters of the lidar;

[0028] After each drone collects multimodal feature data according to the enhanced scanning instructions, the edge computing node performs differential noise injection and spatial replacement processing on the original features; the processed feature data is uploaded to the cloud server through a security protocol, and the three-dimensional feature projection space is updated through an incremental learning algorithm; the fusion process of the newly added feature data and historical data performs double verification, including feature distribution consistency verification and spatiotemporal logical continuity verification.

[0029] As a further solution of the present invention: the dynamic priority rule is specifically:

[0030] Attach an environmental perception tag to each regional feature template. The environmental perception tag contains the lighting condition level and surface material reflection characteristic parameters at the time of acquisition.

[0031] Select a feature optimization strategy based on the parameters of the environmental perception tag: obstacle features in day-night scenes use a temporal sliding window optimization algorithm, and obstacle features in complex terrain scenes use a spatial feature enhancement algorithm;

[0032] Set the feature degradation coefficient in the three-dimensional feature projection space to automatically reduce the priority of historical features that have exceeded their survival period;

[0033] Periodically scan low-frequency feature nodes in the relationship network and perform archiving and storage operations on low-frequency feature nodes that have not been referenced for multiple consecutive periods.

[0034] As a further solution of the present invention: in the path planning module, the interaction method of the secure synchronization channel is:

[0035] When the cloud server sends a feature update request to the drone, it includes a digitally signed regional feature summary value;

[0036] After receiving the feature update request, the drone compares the locally stored feature summary value with the regional feature summary value sent by the cloud, and only uploads the subset of feature parameters that have differences;

[0037] Before receiving the updated data of the global obstacle feature library, the drone generates a zero-knowledge proof document containing the current environment feature parameters and submits it to the cloud server for validity verification;

[0038] After verification by the cloud server, a customized global obstacle signature library update fragment is distributed based on the geocoding and operation type of the drone’s area;

[0039] A block verification mechanism is used during the transmission of update fragments, and each data block contains a self-verified hash value and associated verification of adjacent data blocks.

[0040] As a further solution of the present invention: in the path planning module, the process of generating a three-dimensional obstacle avoidance path is:

[0041] The geometric structure features extracted in real time are matched with the global obstacle feature library for three-level verification:

[0042] The first level matches the basic geometric outline of the obstacle and compares the surface curvature distribution with the curvature templates of known obstacles in the feature library;

[0043] The second level matches the spatial density features to verify the point cloud distribution pattern and the density variation pattern of similar obstacles in the feature library;

[0044] The third level matches the dynamic change trend, comparing the real-time scanning data with the obstacle morphology evolution path recorded in the feature library;

[0045] For obstacles that are successfully matched at all three levels, the pre-stored avoidance strategy is called, and for obstacles that are partially matched, the multi-frame contour reconstruction process is started; a three-dimensional safe passage area is constructed based on the matching verification results, and the boundary of the three-dimensional safe passage area dynamically shrinks as the confidence level of obstacle recognition increases.

[0046] As a further solution of the present invention: the specific process of the multi-frame contour reconstruction process is:

[0047] Collect multi-angle point cloud data of the same obstacle during three consecutive flight control cycles to construct a time series model of the obstacle surface contour;

[0048] Analyze the volume change rate and center of mass displacement trajectory of the time series model of the obstacle surface contour to distinguish the motion properties of static obstacles from dynamic obstacles;

[0049] A contour completion algorithm is implemented for static obstacles, and missing areas of the point cloud of adjacent frames are interpolated and reconstructed. A motion trajectory prediction model is established for dynamic obstacles, and the spatial collision probability between dynamic obstacles and the UAV route is calculated.

[0050] The reconstructed complete contour and predicted trajectory are uploaded to the global obstacle feature library, triggering the incremental update mechanism of the feature library.

[0051] As a further solution of the present invention: the process of constructing the three-dimensional safe passage area is:

[0052] Safety levels are divided according to the confidence level of obstacle recognition: obstacles that are fully matched are defined as fixed safety boundaries, while obstacles that are partially matched are expanded by a preset ratio.

[0053] Establish a dynamic buffer zone calculation model to control the real-time flight speed and attitude adjustment capabilities of the UAV;

[0054] The overlapping parts of adjacent three-dimensional safe passage areas are fused to generate a continuous three-dimensional flight corridor; virtual navigation beacons are deployed in the three-dimensional flight corridor, and the positions of the virtual navigation beacons are dynamically arranged according to the maximum safe distance of the obstacle surface curvature; when it is detected that the deviation between the virtual navigation beacon and the drone's heading exceeds a threshold, a local path replanning instruction is triggered.

[0055] As a further solution of the present invention: the deployment method of the virtual navigation beacon is:

[0056] The central axis of the 3D flight corridor is extracted, and initial beacon positions are set at equal time intervals along the axis. The beacon spacing is adjusted based on the minimum safe diameter of the 3D flight corridor cross section; the smaller the diameter, the higher the beacon density.

[0057] A risk monitoring area is established at each beacon location, and the surface distance change rate between the risk monitoring area and the nearest obstacle is calculated in real time. When the surface distance change rate exceeds a safety threshold, a position adjustment instruction is sent to adjacent beacons to form a new heading guidance path.

[0058] The virtual navigation beacon position data is distributed to other drones through a secure synchronization channel, enabling collaborative obstacle avoidance path sharing.

[0059] Beneficial effects of the present invention:

[0060] The present invention builds a global obstacle feature library based on a hierarchical federated learning framework. Each drone completes the geometric structure analysis and spatial position obfuscation processing of point cloud data locally, generates desensitized feature parameters, and then encrypts them with elliptic curves and aggregates them with edge nodes to achieve secure sharing of obstacle features across devices. Through three-dimensional feature projection space and dynamic priority rules, multi-source heterogeneous obstacle features are mapped to a unified coordinate system and a spatiotemporal correlation link is established. Combined with a three-level matching verification mechanism, matching analysis is performed from the basic geometric contours, spatial density features, and dynamic change trends to improve the recognition accuracy of complex obstacles. A multi-frame contour reconstruction process is adopted. Combined with virtual navigation beacon deployment technology, surface contour interpolation reconstruction and motion trajectory prediction are performed on unmatched obstacles, and a three-dimensional flight corridor with dynamic buffer areas is constructed to achieve deep coupling of path planning and obstacle recognition. With the help of incremental update trigger mechanism and directional feature learning tasks, enhanced scanning and differential noise injection processing are automatically started when the number of unidentified obstacles in a specific area reaches a threshold, and the spatiotemporal evolution modeling capabilities of the global obstacle feature library are continuously optimized, thereby significantly reducing the misjudgment rate of new obstacles while avoiding the leakage of original data, and solving the problems of redundant obstacle avoidance paths and low operational efficiency caused by data silos in traditional solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] The present invention will be further described below with reference to the accompanying drawings.

[0062] Figure 1 It is a module schematic diagram of the present invention. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0064] See also Figure 1 As shown, the present invention is a UAV intelligent obstacle avoidance system based on laser radar, comprising:

[0065] Parameter Acquisition Module: Multiple drones simultaneously collect raw point cloud data from their flight areas using lidar. Local feature processing units then perform geometric structural analysis on this data. The specific process includes: dividing the point cloud into three-dimensional cubic units, calculating the point cloud gradient within each unit to generate a spatial density map; identifying rigid obstacles through normal vector differences; extracting curvature mutation points to construct surface contours to distinguish between independent obstacles and obstacle clusters; finally, fusing the density map with contour features to encode a feature vector, and using random matrix permutation to eliminate geographic correlation, generating desensitized feature parameters.

[0066] The Feature Analysis Module is responsible for the secure transmission and regional aggregation of desensitized feature parameters. The drone initially encrypts the feature parameters using elliptic curve cryptography and adds an anonymous identifier. Upon receiving the feature parameters, the edge computing node decrypts and aggregates the data for the same region using a hash chain key, forming a regional feature template. The cloud server then verifies the integrity of the template using secure multi-party computation and uses homomorphic encryption to obfuscate the features again, ultimately generating a global feature set whose origin cannot be traced.

[0067] Cloud Server: Builds and maintains a global obstacle feature library. It uses a cross-regional feature association model to identify regionally specific obstacles and manages feature weights based on dynamic priority rules, such as high update frequency in agricultural scenarios and high matching accuracy in power scenarios. It constructs a three-dimensional feature projection space, maps multi-source features to a unified coordinate system, and establishes spatiotemporal correlation links to record seasonal changes in obstacle morphology and location. When the number of unidentified obstacles in a region reaches a threshold, it issues enhanced scanning instructions and continuously optimizes the feature library through incremental learning.

[0068] Path Planning Module: This module incrementally updates the feature library through a secure synchronous channel. For example, when comparing summary values, only the difference data is transmitted. A 3D obstacle avoidance path is generated by combining real-time scan data. This module specifically includes: Verifying obstacles through three-level matching, including geometric outline, spatial density, and dynamic trend. Multi-frame outline reconstruction is initiated for partially matched obstacles. Safety levels are assigned based on matching confidence. A 3D flight corridor with a dynamic buffer zone is constructed. Virtual navigation beacons are deployed, and the beacon spacing dynamically adjusts with corridor width to support collaborative path sharing among multiple aircraft.

[0069] In a preferred embodiment of the present invention, in the parameter acquisition module, the specific process of the obstacle geometric structure analysis is:

[0070] 1. Quantitative modeling of spatial density characteristics

[0071] The raw point cloud data collected by the LiDAR is divided into equal-volume cubic cells in three-dimensional space. The cell size can be adjusted based on the operational scenario, such as 1m³ cells for agricultural scenarios and 0.5m³ cells for power inspections. A spatial density distribution map is generated by calculating the gradient change in the number of point clouds within each cell—the ratio of the difference in point cloud counts between adjacent cells to their spatial distance. This map visually displays the spatial distribution of obstacles in the form of a heat map. For example, dense point cloud areas exhibit high gradient values, such as building walls; sparse point cloud areas exhibit low gradient values, such as low-altitude vegetation; and areas without point clouds are marked as traversable space.

[0072] 2. Geometric regularity recognition of rigid obstacles

[0073] For adjacent cubic units, the system calculates the directional difference between the point cloud normal vectors and uses the cosine of the angle between these two units to determine surface continuity. When the angle between the normal vectors is less than a preset threshold (e.g., 30°), the adjacent units are considered to belong to the same smooth surface, such as a flat wall. If the angle exceeds the threshold, the unit is considered to have a sudden surface change, such as a wall corner or obstacle corner. This allows for rapid identification of rigid obstacles with regular geometric structures, such as rectangular buildings and cylindrical towers, while eliminating interference from irregular obstacles such as natural terrain and dynamic vegetation.

[0074] 3. Obstacle outline construction and cluster classification

[0075] At the boundaries of the cube cells, curvature mutation points are extracted through curvature calculations, such as the curvature values of quadratic surfaces fitted based on local point clouds. These points are typically located at the edges or vertices of obstacles, such as the edge of a tree canopy or the top corner of a vehicle. These curvature mutation points are sequentially connected to form the obstacle surface contour line, and the obstacle type is determined by the closure of the contour line: closed contour lines correspond to independent obstacles, such as single trees and isolated buildings, while open or chained contour lines correspond to continuous obstacle groups, such as forest boundaries and building clusters. This classification mechanism provides key topological information for subsequent path planning. For example, independent obstacles can be circumvented, while continuous obstacle groups require global detour planning.

[0076] 4. Feature fusion and privacy protection processing

[0077] The numerical features of the spatial density distribution map (such as the gradient values of each unit) are encoded and fused with the geometric features of the surface contour lines (such as the length of the contour lines and the coordinates of the extreme points of curvature) to generate a fixed-dimensional feature vector sequence, such as a 128-dimensional vector, to achieve dimensionality reduction representation of obstacle features. To prevent the original data from leaking the geographic location of the operation area, a random matrix permutation algorithm is used to obfuscate the spatial coordinate components of the feature vector: by generating a random orthogonal matrix and linearly transforming the vector, the transformed feature vector loses its association with the original coordinate system, while retaining the relative relationships of the geometric structure, such as distance and angle invariance. This process ensures that the desensitized feature parameters only contain the morphological characteristics of the obstacle and do not contain any geographic positioning information.

[0078] In another preferred embodiment of the present invention, in the feature analysis module, the specific working process of the hierarchical encryption transmission link is:

[0079] 1. Initial encryption and anonymization on the device

[0080] Before sending desensitized signature parameters, drones use elliptic curve cryptography (ECC) to encrypt data packets. Based on the elliptic curve discrete logarithm problem, the ECC algorithm offers the advantages of short key length and high computational efficiency, making it suitable for terminal devices like drones with limited computing resources. The encrypted data packet is appended with an irreversible anonymous device identifier. This identifier is generated by hashing the device's physical address (MAC address) using a hashing operation (e.g., SHA-256), ensuring that attackers cannot trace the identifier back to the original device. For example, a drone's MAC address is hashed to generate a 64-bit anonymous identifier "a1b2c3...", which is valid only for the current transmission cycle and is regenerated for the next communication.

[0081] 2. Edge Node Decryption Aggregation and De-identification

[0082] After receiving encrypted data packets from multiple drones in the same region, edge computing nodes use a hash chain-based key derivation method to generate a temporary decryption key. This hash chain generates a key sequence, H(k), H(H(k)), …, through iterative hashing operations. Each key is used only for data decryption within the current cycle, ensuring key timeliness. After decryption, the edge node strips the device's anonymous identifier, retaining only the characteristic parameters. It then performs an initial aggregation of these characteristic parameters within the same region. By calculating metrics such as average curvature and density distribution mean, it generates a regional feature template that reflects regional commonality, such as the average density feature of crop rows in a farmland scenario. This template removes noise data unique to a single device, improving feature robustness.

[0083] 3. Cloud-based secondary obfuscation and global feature generation

[0084] After receiving the cross-region regional feature templates, the cloud server first verifies the data integrity using a secure multi-party computation (MPC) protocol. MPC allows participants (edge nodes) to collaborate on computations without sharing the original data. For example, interactive verification ensures that the template has not been tampered with. After verification, the cross-region features are obfuscated again using homomorphic encryption technology. This allows numerical operations (such as addition and multiplication) to be performed directly on the encrypted data, enabling the cloud to aggregate features without accessing the plaintext. The obfuscated feature data is then timestamped and encoded using a blind signature mechanism, such as "20250527-01" for data from the first region on May 27, 2025. This ultimately generates a global feature set that is untraceable to the original device. This set retains only common cross-regional obstacle features, such as transmission tower structures shared across different regions, enabling the cloud to build a comprehensive obstacle feature library covering multiple scenarios.

[0085] In another preferred embodiment of the present invention, the specific process of constructing the global obstacle feature library in the cloud server is as follows:

[0086] 1. Cross-regional feature association and regional specificity identification

[0087] Using cross-regional feature association models, such as the K-nearest neighbor algorithm based on cosine similarity, regional feature templates uploaded by edge nodes are compared with a historical feature database. For example, if the similarity between the "crop row density feature" uploaded from a particular agricultural region and the historical crop features of similar crops falls below a threshold (e.g., 60%), the obstacle is identified as a region-specific obstacle (such as a new crop variety or a unique planting pattern). The system automatically flags such features and triggers a manual review process to ensure accurate modeling of rare obstacles.

[0088] 2. Dynamic Priority Rules: Context-Aware Feature Management

[0089] Dynamically adjust the handling strategy based on obstacle characteristics in different operation scenarios:

[0090] Agricultural scenarios: Crop growth is seasonal, with sparse growth during the seedling stage and dense growth during the flowering stage. Crop features that change over time, such as point cloud density and stem curvature, are weighted with a higher update frequency. For example, a mandatory update of the crop feature template every seven days ensures that the feature library reflects the crop growth status in real time.

[0091] Power Utility Scenarios: Transmission towers, substations, and other equipment are located in fixed locations but have complex structures. High matching accuracy weights are assigned to fixed spatial features, such as tower geometry and insulator curvature. Increasing the dimension of the feature vector, for example from 128 to 256 dimensions, improves recognition accuracy for complex structures.

[0092] 3. 3D Feature Projection Space: Unified Modeling of Multi-Scene Features

[0093] A three-dimensional feature projection space is constructed, and based on principal component analysis (PCA) or t-SNE dimensionality reduction technology, obstacle features of different regions and scales are mapped to a unified coordinate system. For example, features such as "crop rows" in agricultural scenarios, "towers" in power scenarios, and "buildings" in logistics scenarios are projected into the same space to form an obstacle feature relationship network covering multiple operation scenarios. Spatiotemporal association links are established in the network to record the morphological changes of similar obstacles in different seasons, such as the change in point cloud density from flowering to fruiting of a tree, as well as position migration, such as dynamic obstacles on the migration path of migratory birds. The system automatically retrieves the feature template for the corresponding season based on the current time point and updates the real-time morphology and position of the obstacle.

[0094] 4. Incremental learning mechanism: active detection of unknown obstacles

[0095] When the number of unidentified obstacles in a certain area reaches 10% of the feature library capacity (configurable threshold), the cloud triggers an enhanced scanning command:

[0096] Command parameters: specify the pitch angle change sequence of the lidar, such as scanning every 10° from -30° to +60°, and echo intensity acquisition parameters, such as enhancing the signal gain of weakly reflecting targets, to ensure that multi-angle and multi-intensity characteristics of obstacles are obtained.

[0097] Data Processing: Multimodal feature data collected by drones is processed by edge nodes. This data is further desensitized through differential noise injection (adding Laplace-distributed noise to maintain differential privacy) and spatial permutation (randomly scrambling the order of point cloud coordinates). The processed feature data is uploaded to the cloud via a secure transmission protocol. Incremental learning algorithms (such as online gradient descent) are used to update the 3D projection space, avoiding the computational overhead of retraining the entire model.

[0098] Double verification: When new features are integrated with historical data, feature distribution consistency verification is performed, such as the KS test to determine whether the distribution of new and old features is consistent, as well as spatiotemporal logical continuity verification, such as whether the change in obstacle position conforms to the law of geographical migration, to ensure the reliability of the new data.

[0099] In a preferred embodiment of the present invention, the dynamic priority rule is specifically as follows:

[0100] 1. Environmental Perception Labeling: Scenario-Based Feature Annotation

[0101] Attach environmental awareness labels to each regional feature template, including:

[0102] Light condition levels: divided into three levels: low light (≤500 lux), medium light (500-3000 lux), and strong light (≥3000 lux), and collected in real time by the light sensor on the drone.

[0103] Surface material reflectance characteristics, such as soil (reflectivity 10-20%), vegetation (20-40%), and metal (60-80%), are obtained through LiDAR echo intensity inversion. This labeling data is used for dynamic selection of subsequent feature optimization strategies.

[0104] 2. Feature Optimization Strategy: Adaptive Processing of Spatiotemporal Dimensions

[0105] Temporal sliding window optimization (day-night scenarios): To address the shadow changes caused by the alternation of day and night, a sliding window algorithm is used with a window size set to 24 hours. Only the feature data within the current window is retained, and outdated day and night features are eliminated, improving matching efficiency in dynamic environments.

[0106] Spatial dimension feature enhancement: In terrains such as mountains and canyons, feature enhancement algorithms (such as feature vector normalization and noise filtering) are used to highlight the three-dimensional structural features of obstacles (such as height difference and slope), thereby suppressing the interference of terrain noise on obstacle avoidance.

[0107] 3. Feature Degradation and Archiving: Automated Lifecycle Management

[0108] Feature degradation coefficient: Set a survival period for each feature in the three-dimensional projection space, such as 30 days by default. Features that exceed the period will automatically reduce their priority weight, such as exponentially decaying to 50% of the original value, to prevent outdated features from affecting the matching of new features.

[0109] Low-frequency feature archiving: Regularly scan the feature relationship network and perform archiving and storage operations on low-frequency features that have not been referenced by drones for five consecutive cycles, such as temporary obstacles formed by rare natural disasters. These features are migrated from the real-time matching library to the historical database, freeing up computing resources while retaining data traceability.

[0110] In another preferred embodiment of the present invention, in the path planning module, the interaction method of the secure synchronization channel is:

[0111] 1. Comparison of feature update request and summary

[0112] Cloud-initiated request: When the global obstacle feature library is updated, such as by adding a new regional feature template or optimizing an existing feature, the cloud server sends a feature update request to the drone swarm in the target area, along with a digitally signed regional feature digest value. The digest value is generated using a hash algorithm (such as SHA-3) on key parameters of the updated feature (such as the curvature template mean and density distribution variance). The length is fixed to 256 bits to ensure data integrity and verifiability.

[0113] Drone Local Verification: After receiving the request, the drone extracts the locally stored feature summary value, periodically synchronizes it to the local cache, and compares it bit by bit with the summary value sent from the cloud. Only when the two values are inconsistent is a feature difference determined, triggering the subsequent differential update process to avoid transmitting redundant data. For example, if the local summary is "d41d8cd98f00b204e9800998ecf8427e," indicating no feature update, the upload and download operations are skipped.

[0114] 2. Zero-knowledge proof: trusted verification of environmental characteristics

[0115] Before receiving updated data from the cloud, the drone needs to prove the legitimacy of its current operating environment to the cloud without revealing specific environmental parameters. This is achieved through zero-knowledge proof (ZK-SNARKs) technology:

[0116] Proof Generation: The drone collects environmental parameters, such as light levels, surface reflectivity, and the geocoded prefix of the real-time location, and generates a proof document containing these parameters through an encryption function. The proof document simply indicates that the current environment meets the characteristic update conditions, such as being within the preset operating area, but does not disclose the specific parameter values.

[0117] Cloud Verification: The cloud server uses a pre-deployed verification key to verify the validity of the certification document. For example, this can verify whether the drone is within an authorized operating area, such as by verifying the coordinate range of a geofence, or whether it is operating within a specific operation type, such as allowing agricultural plant protection drones to only receive crop-related feature updates. Once verification is successful, the cloud server unlocks the corresponding feature update snippet download permission.

[0118] 3. Customized feature distribution: scenario-sensitive data push

[0119] The cloud performs differentiated feature update strategies based on the drone's geocoding and operation type tags:

[0120] Geocoding matching: The geocoding of the drone's real-time location (e.g., the province-city-district three-level coding based on the BD-09 coordinate system) is matched with the regional labels in the feature library (e.g., "North China-Beijing-Chaoyang District"), and only obstacle features in that area and adjacent areas (e.g., common local building and tree types) are pushed, reducing irrelevant data transmission.

[0121] Operation type filtering: Customized feature subsets are distributed for different operation scenarios, including agriculture, power generation, and logistics. For example, agricultural drones only receive features related to crops and irrigation equipment, while power generation drones receive features of equipment such as transmission towers and insulators. This avoids feature library redundancy that affects matching efficiency.

[0122] Updated fragment compression: Customized feature fragments are losslessly compressed (such as the gzip algorithm), with a compression ratio of up to 3:1, further reducing transmission bandwidth requirements. This is suitable for UAV operations in remote areas with weak network environments.

[0123] 4. Block verification mechanism: double guarantee for reliable transmission

[0124] During transmission, the update fragment is divided into fixed-size data blocks (e.g., 512KB per block) and double-checked information is appended:

[0125] Self-verifying hash value: Each data block generates a unique hash value using the SHA-256 algorithm. After receiving it, the drone recalculates the hash value and compares it with the transmitted value to ensure that the individual data block has not been tampered with or damaged.

[0126] Adjacent Block Linked Checksum: This introduces a chained checksum mechanism where each data block includes the hash value of the previous data block. For example, the checksum field of block n contains the hash value of block n-1, forming a checksum chain. If a data block fails to transmit, adjacent block checksums can quickly locate the error, triggering a retransmission of a single block rather than the entire data block, improving transmission efficiency.

[0127] In another preferred embodiment of the present invention, in the path planning module, the process of generating a three-dimensional obstacle avoidance path is:

[0128] The generation of a 3D obstacle avoidance path is based on a multi-level feature matching mechanism, which improves the accuracy and robustness of obstacle recognition through step-by-step verification:

[0129] 1. First level matching: basic geometric contour verification

[0130] The real-time scanned obstacle point cloud data is used to extract surface curvature distribution features, such as Gaussian curvature and mean curvature, through the principal curvature calculation algorithm, and compared with the curvature templates of known obstacles in the global feature library. For example:

[0131] For spherical obstacles, such as water tanks, verify whether their surface curvature is evenly distributed and close to the theoretical spherical curvature value;

[0132] For rectangular obstacles, such as buildings, check whether their edge curvature presents a right-angle mutation feature.

[0133] The matching threshold is dynamically adjusted according to the operation scenario. The power inspection scenario requires a curvature error of ≤5%, and the agricultural scenario can be relaxed to 10% to accommodate the natural variation of crop morphology.

[0134] 2. Second level matching: spatial density feature verification

[0135] Analyze the spatial density distribution pattern of the real-time point cloud, such as the trend of the number of point clouds changing with height, the geometric shape of sparse areas, and compare it with the density change pattern of similar obstacles in the feature library. For example:

[0136] For low-altitude vegetation obstacles, verify whether their point cloud density shows the natural growth characteristics of "dense at the bottom and sparse at the top";

[0137] For transmission tower obstacles, verify whether their density distribution conforms to the structural characteristics of "dense tower base and sparse tower frame".

[0138] The mean and variance of the density distribution are calculated using the sliding window statistical method. When the difference between the real-time data and the feature library template exceeds the preset threshold, such as when the variance deviation is greater than 15%, it is determined to be a partial match.

[0139] 3. Third-level matching: dynamic change trend analysis

[0140] Retrieve the spatiotemporal evolution records of obstacles in the feature library, such as the morphological changes of crops in different growth cycles and the seasonal movement of migratory bird routes, and compare them with the dynamic trends of real-time scanning data:

[0141] For static obstacles, such as buildings, check whether their location and shape are consistent with historical records;

[0142] For dynamic obstacles, such as moving vehicles, predict whether their motion trajectory in the next 3 seconds is consistent with the average speed and turning pattern of similar objects in the feature library.

[0143] The Kalman filter algorithm is used to fit the deviation between the real-time trajectory and historical data. When the prediction error exceeds a safe distance (such as 2 meters), it is determined to be a dynamic feature mismatch.

[0144] 4. Matching result processing

[0145] Level 3 Full Match: Directly invokes pre-existing standard avoidance strategies, such as executing a fixed path of "circling half a tower while maintaining a 5-meter vertical distance" for a known transmission tower. Partial Match: Initiates a multi-frame contour reconstruction process to avoid misjudgments caused by missing data from a single frame.

[0146] In a preferred embodiment of the present invention, the specific process of the multi-frame contour reconstruction process is as follows:

[0147] 1. Multi-angle data collection and time series modeling

[0148] The drone collects point cloud data of the same obstacle at a yaw angle of ±15° over three consecutive flight control cycles (typically 0.3 seconds). This data then constructs a 3D contour time series model, consisting of the contour coordinates at times t-1, t, and t+1. For example, for trees obscured from the side, the complete crown outline is restored by stitching together the front, left, and right view frames.

[0149] 2. Motion attribute identification: static and dynamic obstacle classification

[0150] Volume change rate calculation: Compare the volume differences of the contours of adjacent frames. If the change rate is less than 5%, it is determined to be a static obstacle, such as a fixed building. Otherwise, it is determined to be a dynamic obstacle, such as mobile machinery.

[0151] Center of mass offset analysis: Tracks the coordinate changes of the contour center of mass and, combined with the drone's own flight speed, calculates the relative speed and direction of the obstacle. For example, if the center of mass offset trajectory forms an angle greater than 90° with the drone's flight path, it is considered a lateral obstacle crossing and requires avoidance priority.

[0152] 3. Differentiated reconstruction and prediction algorithm

[0153] Static obstacle completion: Using a triangulation interpolation algorithm, surface fitting is performed on missing areas of the point cloud between adjacent frames (such as the backlit side of a building) to complete the outline details. The completed outline error must be less than 0.2 meters (for agricultural scenarios) or less than 0.05 meters (for power generation scenarios).

[0154] Dynamic Obstacle Prediction: A linear regression model is used to establish a trajectory prediction equation. This equation combines obstacle size and drone speed to calculate spatial collision probability, such as the probability of the distance falling below a safety threshold within the next five seconds. When the probability exceeds 60%, an emergency obstacle avoidance plan is triggered.

[0155] 4. Incremental update of feature library

[0156] The reconstructed complete contour data and dynamic prediction model parameters (mean velocity, volume change rate threshold) are uploaded to the cloud. A federated learning mechanism is used to update the global feature library, enabling "one-time recognition, network-wide sharing." For example, if a new type of agricultural machinery obstacle appears in a certain area for the first time, its reconstructed features will be synchronized to all drones in the area to avoid repeated misjudgments.

[0157] In another preferred embodiment of the present invention, the process of constructing the three-dimensional safe passage area is as follows:

[0158] 1. Security level division and dynamic boundary adjustment

[0159] Complete obstacle matching: A fixed safety boundary is defined based on the standard size recorded in the feature library. For example, the obstacle perimeter is extended by 1.5 meters, and the boundary remains unchanged during flight.

[0160] Partially matched / reconstructed obstacles: The safety margin is expanded by a preset percentage, such as 200%, to create a buffer zone to reduce the risk of collisions caused by incomplete outlines. For example, for reconstructed tree obstacles, the margin is expanded from the standard 1 meter to 2 meters until the confidence level reaches above 80%.

[0161] 2. Dynamic buffer area and flight control coupling

[0162] Establish a buffer area calculation model, and the input parameters include:

[0163] Obstacle recognition confidence, ranging from 0-100%;

[0164] The drone's current speed and attitude adjustment capabilities, such as maximum turn rate, are monitored. A safety buffer distance is output, such as 3 meters at a 50% confidence level and 1.5 meters at an 80% confidence level. The model ensures smooth obstacle avoidance by adjusting the drone's speed limit in real time, such as limiting it to 3 meters per second when entering a high buffer zone, as well as its turning radius.

[0165] 3. 3D flight corridor generation and beacon deployment

[0166] Corridor Fusion: Boolean operations are used to merge overlapping safety zones of adjacent obstacles into a continuous 3D flight corridor, avoiding path breakage caused by discrete boundaries. For example, the safety zones of two adjacent forests are merged to form a continuous "tree corridor."

[0167] Dynamic arrangement of virtual navigation beacons:

[0168] Initial beacons are set up along the central axis of the corridor at equal time intervals (e.g., every 0.5 seconds of flight distance);

[0169] Adjust the beacon spacing based on the minimum safe diameter of the corridor cross section. When the diameter is less than 2 meters, the beacon density is doubled to ensure navigation accuracy.

[0170] The beacon has a built-in risk monitoring module that calculates the rate of change of the surface distance to the nearest obstacle in real time. When the rate of change is greater than 0.5m / s, it automatically sends position adjustment instructions to adjacent beacons to form an adaptive obstacle avoidance path.

[0171] 4. Local Path Replanning Trigger Mechanism

[0172] When the deviation between the drone's heading and the beacon's guidance direction exceeds a threshold, the system determines it as a path deviation and triggers local replanning:

[0173] Based on the A-star algorithm, an alternative path is searched within the current corridor, with priority given to channels with small curvature changes and high beacon density. The replanning time must be less than 200ms to ensure the real-time obstacle avoidance response.

[0174] It is worth noting that the deployment method of the virtual navigation beacon is:

[0175] 1. Initial beacon layout: axis extraction and dynamic density adjustment

[0176] The deployment of virtual navigation beacons is based on three-dimensional flight corridors and intelligently arranged through geometric feature analysis:

[0177] A skeleton extraction algorithm is used to generate a central axis from the point cloud data of the 3D flight corridor. This axis represents the corridor's geometric centerline, ensuring symmetrical beacon distribution along the path. For example, in an L-shaped corridor or a curved forest path, the axis extends along the corridor's direction, maintaining a safe distance from obstacles on both sides.

[0178] Initial beacon positions are set at fixed intervals along the axis, for example, a beacon is deployed every 0.5 seconds of the drone's flight. Assuming the drone's flight speed is 6m / s, the initial spacing between adjacent beacons is 3 meters, forming a uniformly distributed sequence of navigation nodes.

[0179] The minimum safe diameter of the flight corridor cross section is calculated in real time. This is the shortest distance between two obstacles in a plane perpendicular to the axis. The beacon density is dynamically adjusted based on the diameter:

[0180] When the diameter is ≥5 meters, maintain the initial distance of 3 meters;

[0181] When 3 meters ≤ diameter < 5 meters, the beacon spacing is reduced to 1.5 meters and the density is doubled;

[0182] When the diameter is less than 3 meters, the spacing is further reduced to 0.8 meters to ensure dense beacon coverage in narrow spaces, such as gaps between transmission towers and urban alleys, thereby improving navigation accuracy.

[0183] 2. Risk Monitoring and Dynamic Beacon Adjustment: Real-time Environmental Response Mechanism

[0184] Each beacon acts as an independent intelligent sensing node and has the ability to adapt to the dynamic environment:

[0185] A spherical risk monitoring area is established with the beacon position as the center, and the radius is dynamically adjusted according to the beacon spacing. For example, when the spacing is 1.5 meters, the radius is set to 2 meters to ensure that the monitoring range covers potential obstacles between adjacent beacons.

[0186] The system continuously collects point cloud data of obstacles within the monitoring area through LiDAR, and calculates the speed of change in the distance between the beacon center and the nearest obstacle surface in real time. If the distance rapidly decreases or increases, such as when the rate of change exceeds the safety threshold of 0.6m / s, a collision risk or path redundancy is determined.

[0187] If the distance is rapidly decreasing, such as when an obstacle approaches, the beacon will shift 0.3-0.8 meters to a safe direction perpendicular to the axis and send coordinated adjustment instructions to the upstream and downstream adjacent beacons to form a detour curve;

[0188] If the distance increases significantly, such as when entering an open area, the beacon moves forward 1-2 meters along its axis, compressing redundant paths and improving flight efficiency. The adjusted beacon position is quickly synchronized to adjacent nodes via a distributed communication protocol, ensuring the consistency of the path adjustment.

[0189] 3. Collaborative Obstacle Avoidance Path Sharing: Multi-Device Data Distribution and Navigation Synchronization

[0190] The location data of the virtual navigation beacon is shared in real time within the drone cluster through a secure synchronization channel, enabling collaborative perception of dynamic paths:

[0191] Beacon location information (including 3D coordinates, adjustment time, and risk level) is encrypted and sent via multicast to drone swarms in the same area, reducing repeated transmission overhead. Data compression technology can reduce transmission volume by approximately 50%, meeting the communication needs of drones in weak network environments.

[0192] After receiving the beacon data, the drone integrates it with the obstacle point cloud scanned locally in real time and displays it as a visual icon on the onboard interface. For example, a green beacon represents a safe path, a yellow beacon indicates the need to slow down, and a red beacon warns of a dangerous area, assisting the drone in predicting the path.

[0193] The lead drone is responsible for detecting the path ahead and generating beacons, and subsequent drones directly follow the beacon to reduce repeated scanning energy consumption;

[0194] When a drone triggers a beacon adjustment due to dynamic changes in obstacles, the new beacon data is automatically synchronized to all members of the cluster, achieving a collaborative obstacle avoidance effect of "single-unit perception, group response", and avoiding path conflicts caused by independent planning of multiple drones.

[0195] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A UAV intelligent obstacle avoidance system based on laser radar, characterized in that: include: The parameter acquisition module uses the lidar sensors of multiple drones to synchronously collect raw point cloud data of the flight area. The local feature processing unit of each drone analyzes the geometric structure of obstacles in the raw point cloud data and generates desensitized feature parameters including surface curvature and spatial density distribution. A feature analysis module is used to send the desensitized feature parameters to the edge computing node through a hierarchical encrypted transmission link. The edge computing node initially aggregates the feature parameters of multiple drones in the same area to form a regional feature template; The cloud server is used to receive regional feature templates from multiple edge computing nodes, integrate obstacle features from different regions through a secure fusion protocol, and build a global obstacle feature library; The path planning module is used to control each drone to obtain updated data from the global obstacle feature library through a secure synchronous channel, and generate a three-dimensional obstacle avoidance path by combining real-time scanning features with the feature library matching results; In the cloud server, the specific process of building a global obstacle feature library is as follows: Establish a cross-regional feature association model, compare the regional feature templates with the historical feature database for similarity, and identify obstacle categories with regional specificity; Dynamic priority rules are established for recurring cross-regional obstacle features. Crop features that change over time in agricultural scenarios are assigned high update frequency weights, and equipment features with fixed spatial positions in power scenarios are assigned high matching accuracy weights. Construct a three-dimensional feature projection space, map obstacle features in different areas to a unified coordinate system, and form an obstacle feature relationship network covering multiple operation scenarios. Establish spatiotemporal association links between obstacle features within the obstacle feature relationship network, record the morphological changes and position migration patterns of similar obstacles in different seasons, and update the obstacle morphology and position based on the current time point. When the number of unidentified obstacles in a certain area reaches the characteristic library capacity ratio, the cloud server sends an enhanced scanning instruction to the drone group in the area, specifying the pitch angle change sequence and echo intensity acquisition parameters of the lidar; After each drone collects multimodal feature data according to the enhanced scanning instructions, the edge computing node performs differential noise injection and spatial displacement processing on the original features. The processed feature data is uploaded to the cloud server through a secure protocol, and the three-dimensional feature projection space is updated through an incremental learning algorithm. The fusion process of newly added feature data and historical data performs double verification, including feature distribution consistency verification and spatiotemporal logical continuity verification.

2. The intelligent obstacle avoidance system for unmanned aerial vehicles based on laser radar according to claim 1, characterized in that: In the parameter acquisition module, the specific process of obstacle geometric structure analysis is as follows: The original point cloud data is divided into three-dimensional cubic units of equal volume, and the gradient of the point cloud quantity change in each cubic unit is calculated to generate a spatial density distribution map; Analyze the surface continuity characteristics between adjacent cubic units and identify rigid obstacles with regular geometric structures by calculating the direction difference values of the point cloud normal vectors between cubic units; Extract the curvature mutation points at the boundaries of the cube units, construct the obstacle surface contour line, and distinguish independent obstacles from continuous obstacle groups based on the closed characteristics of the obstacle surface contour line; The spatial density distribution map is fused and encoded with the obstacle surface contour features to generate a fixed-dimensional feature vector sequence; The spatial position confusion processing is performed on the feature vector sequence, and the random matrix permutation algorithm is used to eliminate the geographical location correlation of the original point cloud data.

3. The intelligent obstacle avoidance system for unmanned aerial vehicles based on laser radar according to claim 1, characterized in that: In the feature analysis module, the specific working process of the hierarchical encryption transmission link is as follows: Before sending the desensitized characteristic parameters, each drone uses the elliptic curve encryption algorithm to initially encrypt the data packet of the desensitized characteristic parameters, and the encrypted data packet is attached with an irreversible device anonymous identification code; After receiving encrypted data packets from multiple drones in the same area, the edge computing node uses a hash chain-based key derivation method to generate a temporary decryption key, and then strips the device's anonymous identification code after decryption; After receiving regional feature templates from different edge computing nodes, the cloud server uses a secure multi-party computing protocol to verify data integrity and employs homomorphic encryption technology to perform secondary obfuscation on cross-region features. The obfuscated feature data is timestamped and regionally coded through a blind signature mechanism to generate a global feature set that cannot be traced back to the original device.

4. The intelligent obstacle avoidance system for unmanned aerial vehicles based on laser radar according to claim 1, characterized in that: The dynamic priority rules are specifically as follows: Attach an environmental perception tag to each regional feature template. The environmental perception tag contains the lighting condition level and surface material reflection characteristic parameters at the time of acquisition. Select a feature optimization strategy based on the parameters of the environmental perception tag: obstacle features in day-night scenes use a temporal sliding window optimization algorithm, and obstacle features in complex terrain scenes use a spatial feature enhancement algorithm; Set the feature degradation coefficient in the three-dimensional feature projection space to automatically reduce the priority of historical features that have exceeded their survival period; Periodically scan low-frequency feature nodes in the relationship network and perform archiving and storage operations on low-frequency feature nodes that have not been referenced for multiple consecutive periods.

5. The intelligent obstacle avoidance system for unmanned aerial vehicles based on laser radar according to claim 1, characterized in that: In the path planning module, the interaction method of the secure synchronization channel is: When the cloud server sends a feature update request to the drone, it includes a digitally signed regional feature summary value; After receiving the feature update request, the drone compares the locally stored feature summary value with the regional feature summary value sent by the cloud, and only uploads the subset of feature parameters that have differences; Before receiving the updated data of the global obstacle feature library, the drone generates a zero-knowledge proof document containing the current environment feature parameters and submits it to the cloud server for validity verification; After verification by the cloud server, a customized global obstacle signature library update fragment is distributed based on the geocoding and operation type of the drone’s area; A block verification mechanism is used during the transmission of update fragments, and each data block contains a self-verified hash value and associated verification of adjacent data blocks.

6. The intelligent obstacle avoidance system for unmanned aerial vehicles based on laser radar according to claim 1, characterized in that: In the path planning module, the process of generating a three-dimensional obstacle avoidance path is as follows: The geometric structure features extracted in real time are matched with the global obstacle feature library for three-level verification: The first level matches the basic geometric outline of the obstacle and compares the surface curvature distribution with the curvature templates of known obstacles in the feature library; The second level matches the spatial density features to verify the point cloud distribution pattern and the density variation pattern of similar obstacles in the feature library; The third level matches the dynamic change trend, comparing the real-time scanning data with the obstacle morphology evolution path recorded in the feature library; For obstacles that are successfully matched at all three levels, the pre-stored avoidance strategy is called, and for obstacles that are partially matched, the multi-frame contour reconstruction process is started; a three-dimensional safe passage area is constructed based on the matching verification results, and the boundary of the three-dimensional safe passage area dynamically shrinks as the confidence level of obstacle recognition increases.

7. The UAV intelligent obstacle avoidance system based on laser radar according to claim 6, characterized in that: The specific process of the multi-frame contour reconstruction process is as follows: Collect multi-angle point cloud data of the same obstacle during three consecutive flight control cycles to construct a time series model of the obstacle surface contour; Analyze the volume change rate and center of mass displacement trajectory of the time series model of the obstacle surface contour to distinguish the motion properties of static obstacles from dynamic obstacles; A contour completion algorithm is implemented for static obstacles, and missing areas of the point cloud of adjacent frames are interpolated and reconstructed. A motion trajectory prediction model is established for dynamic obstacles, and the spatial collision probability between dynamic obstacles and the UAV route is calculated. The reconstructed complete contour and predicted trajectory are uploaded to the global obstacle feature library, triggering the incremental update mechanism of the feature library.

8. The laser radar-based intelligent obstacle avoidance system for unmanned aerial vehicles according to claim 6, characterized in that: The construction process of the three-dimensional safe passage area is as follows: Safety levels are divided according to the confidence level of obstacle recognition: obstacles that are fully matched are defined as fixed safety boundaries, while obstacles that are partially matched are expanded by a preset ratio. Establish a dynamic buffer zone calculation model to control the real-time flight speed and attitude adjustment capabilities of the UAV; The overlapping parts of adjacent three-dimensional safe passage areas are fused to generate a continuous three-dimensional flight corridor; virtual navigation beacons are deployed in the three-dimensional flight corridor, and the positions of the virtual navigation beacons are dynamically arranged according to the maximum safe distance of the obstacle surface curvature; when it is detected that the deviation between the virtual navigation beacon and the drone's heading exceeds a threshold, a local path replanning instruction is triggered.

9. The UAV intelligent obstacle avoidance system based on laser radar according to claim 8, characterized in that: The deployment method of the virtual navigation beacon is: The central axis of the 3D flight corridor is extracted, and initial beacon positions are set at equal time intervals along the axis. The beacon spacing is adjusted based on the minimum safe diameter of the 3D flight corridor cross section; the smaller the diameter, the higher the beacon density. A risk monitoring area is established at each beacon location, and the surface distance change rate between the risk monitoring area and the nearest obstacle is calculated in real time. When the surface distance change rate exceeds a safety threshold, a position adjustment instruction is sent to adjacent beacons to form a new heading guidance path. The virtual navigation beacon position data is distributed to other drones through a secure synchronization channel, enabling collaborative obstacle avoidance path sharing.

Images