Lidar airport bird flock detection and early warning method and system based on dgnn

Abstract

The application discloses a laser radar airport bird group detection and early warning method and system based on a DGNN, relates to the technical field of airport operation safety monitoring, and comprises the following steps: collecting continuous three-dimensional point cloud data of an airport runway, a taxiway and adjacent airspace by using a laser radar device and performing preprocessing to form a continuous multi-frame structured point cloud sequence; taking points or point cloud clusters in the sequence that meet preset space and intensity conditions as graph nodes, establishing a connection edge between the nodes according to a spatial proximity relationship between the nodes, a time continuity relationship and operation similarity, and forming a dynamic graph structure that evolves over time; jointly analyzing the features of the graph nodes and the mutual relationship thereof by using a DGNN, and identifying bird targets in the nodes; aggregating the identified bird targets to construct a bird group cluster set, combining a bird strike risk assessment model, calculating a potential risk value of the bird group cluster to the operation of the airport, and performing graded early warning according to the risk value, so that integrated bird group detection, behavior analysis and risk early warning are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of airport operation safety monitoring, and particularly relates to a laser radar airport bird group detection and early warning method and system based on a DGNN. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] With the continuous expansion of civil aviation transportation scale, the number of aircraft takeoff, landing and taxiing in the airport flight area is increasing, and the threat of bird activity to aviation safety is increasingly prominent. After the bird group invades the flight area, it may cause bird strike accidents such as engine suction, aircraft structure damage or windshield rupture, and even endanger flight safety and cause large-scale flight delays, which has become one of the safety hazards that need to be focused on and prevented in airport operation management.

[0004] Traditional airport bird monitoring mainly relies on manual patrol, video monitoring, infrared imaging, bird radar and other technologies. However, the manual patrol method is limited by personnel experience and patrol range, and it is difficult to achieve continuous and all-weather monitoring; the video monitoring system has significantly decreased imaging quality under complex weather conditions such as night, backlight, haze and rain and snow, and has limited recognition ability for long-distance and small-volume bird targets; infrared imaging is prone to false detection in high-temperature background or low-temperature difference environment; some airports use traditional bird radar equipment for medium and long distance detection, but such equipment is usually large in size, high in cost and complex in maintenance, and the detection accuracy in low altitude, small target and high-speed motion scene is still insufficient, which is difficult to meet the needs of multi-point deployment and fine monitoring in the airport flight area. At present, laser radar is gradually introduced into airport perimeter monitoring and low-altitude target detection scenarios due to its high-precision distance measurement capability, strong anti-light interference performance and good adaptability to low-visibility environment. However, due to the characteristics of small volume, fast flight speed and weak reflection of birds, the point cloud data obtained by laser radar often presents the characteristics of sparseness, discontinuity and local clustering, and is easily disturbed by noise such as raindrops, insects and ground reflection, so it is difficult to realize stable and reliable bird target recognition by simply relying on point cloud geometric threshold or static rule method.

[0005] In addition, airport operation management not only needs to judge whether there is a bird target, but also needs to comprehensively evaluate the scale, spatial distribution state, motion trend of the bird group and its potential threat to the flight area, so as to timely issue early warning information and link up the bird driving or scheduling measures. However, most of the existing detection methods only focus on single-frame or single-target recognition of birds, lack the modeling capability of the overall structural features of the bird group and its time evolution behavior, and are difficult to effectively depict the group motion law of the bird group, thereby limiting the accuracy and foresight of bird strike risk warning. SUMMARY

[0006] To address the shortcomings of existing airport bird monitoring technologies, such as insufficient stability under complex weather conditions, difficulty in accurately identifying low-altitude small bird targets, and lack of comprehensive assessment of the overall behavioral characteristics and risk situation of bird flocks, which limits bird strike risk warning, this invention proposes a lidar-based airport bird flock detection and warning method and system based on DGNN. This method fully utilizes lidar point cloud data, operates stably under complex weather and low visibility conditions, and jointly models the spatial relationships and temporal correlations between bird targets, achieving integrated bird flock detection, behavioral analysis, and risk warning, thereby improving the intelligence and precision of airport bird strike risk prevention and control.

[0007] Firstly, this invention proposes a method for detecting and warning of bird flocks at airports using lidar based on DGNN.

[0008] A method for bird flock detection and early warning at airports based on DGNN lidar, comprising:

[0009] Based on lidar equipment deployed in the airport flight area and its surrounding areas, continuous three-dimensional point cloud data of airport runways, taxiways and adjacent airspace are collected, and the collected point cloud data is preprocessed to form a continuous multi-frame structured point cloud sequence.

[0010] Using points or point cloud clusters in the sequence that meet the preset spatial and intensity conditions as graph nodes, and establishing connecting edges between nodes based on spatial proximity, temporal continuity, and operational similarity, a dynamic graph structure that evolves over time is formed.

[0011] Based on the dynamic graph structure, bird targets in the nodes are identified by jointly analyzing the features of the graph nodes and their interrelationships using DGNN.

[0012] The identified bird targets are aggregated to construct bird clusters, and combined with a bird strike risk assessment model, the potential risk value of the bird clusters to airport operations is calculated, and graded early warnings are issued based on the risk value.

[0013] A further technical solution is that the 3D point cloud data of frame t includes the 3D coordinates, reflection intensity, and timestamp of each point in the current frame; the collected point cloud data is preprocessed, including noise filtering, density constraint removal, reflection intensity correction, ground point removal, and time synchronization resampling processing performed sequentially.

[0014] A further technical solution involves constructing a dynamic graph structure based on a continuous multi-frame structured point cloud sequence, including:

[0015] In each frame of point cloud, points or point cloud clusters that meet the preset height range, reflection intensity threshold, and local density constraints are selected as graph nodes. The graph nodes are recorded as the spatial center position of the node, the node feature vector, and the node timestamp.

[0016] Based on the constructed set of nodes, connecting edges between nodes are established according to the spatial proximity and temporal correlation between nodes. For any two nodes, a connecting edge is established if at least one constraint condition is met. The edge weight is a weighted combination of spatial distance, temporal continuity and motion similarity between nodes. Among them, the constraints include: spatial proximity constraint, temporal continuity and motion consistency constraint, and reflection intensity change consistency constraint.

[0017] The nodes and their connections at different times are dynamically updated to construct a dynamic graph structure that evolves over time.

[0018] A further technical solution is that the node feature vector includes the average reflection intensity of the point cloud corresponding to the node, the spatial distribution scale, and the local motion features estimated based on the changes in the center position of the node in adjacent frames.

[0019] A further technical solution, based on a dynamic graph structure, uses DGNN to jointly analyze the features of graph nodes and their interrelationships to identify bird targets within the nodes, including:

[0020] Based on a dynamic graph structure, the features of each node are combined with the features of its neighboring nodes, and the features of the current node are updated through the graph structure feature propagation mechanism.

[0021] Based on the updated node features, a discriminant function is used to identify targets in each node, and bird targets in the nodes are identified. After bird flock target aggregation and continuity constraints, the set of bird targets at each time step and their distribution results in the spatial and temporal dimensions are output.

[0022] A further technical solution involves aggregating identified bird targets to construct bird flock clusters, and combining this with a bird strike risk assessment model to calculate the potential risk value of these bird flock clusters to airport operations, including:

[0023] Based on the identified set of bird targets, bird targets with continuous correlation in spatial location and time dimension are aggregated to construct a bird flock cluster set;

[0024] Based on the number of targets contained in a bird flock and their spatial distribution range, calculate the bird flock size density index, the bird flock spatial proximity index, and the movement trend index.

[0025] By combining bird flock size and density indicators, spatial proximity indicators, and movement trend indicators, a bird strike risk assessment model is constructed. The values ​​of the three indicators are weighted and calculated to obtain the final potential risk value of bird flocks to airport operations.

[0026] Secondly, this invention proposes a lidar system for detecting and warning flocks of birds at airports based on DGNN.

[0027] A lidar-based airport bird flock detection and early warning system using DGNN, comprising:

[0028] The data acquisition module includes lidar equipment deployed in the airport flight area and its surrounding areas, used to collect continuous three-dimensional point cloud data of the airport runway, taxiway and adjacent airspace, and to preprocess the collected point cloud data to form a continuous multi-frame structured point cloud sequence.

[0029] The dynamic graph structure construction module is used to establish connecting edges between nodes based on spatial proximity, temporal continuity, and operational similarity, using points or point cloud clusters in the sequence that meet preset spatial and intensity conditions as graph nodes, thus forming a dynamic graph structure that evolves over time.

[0030] The bird target recognition module is used to identify bird targets in nodes based on a dynamic graph structure and by jointly analyzing the features of graph nodes and their interrelationships using DGNN.

[0031] The early warning module is used to aggregate identified bird targets to construct bird clusters, combine them with a bird strike risk assessment model, calculate the potential risk value of bird clusters to airport operations, and issue graded early warnings based on the risk value.

[0032] Thirdly, the present invention also provides an electronic device, comprising: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the above-described DGNN-based lidar airport bird flock detection and early warning method.

[0033] Fourthly, the present invention also provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, implement the above-described method for detecting and warning flocks of birds at airports using a DGNN-based lidar.

[0034] Fifthly, the present invention also provides a computer program product comprising executable instructions stored in a computer-readable storage medium; wherein, when the processor of the electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, the above-mentioned method for detecting and warning flocks of birds at airports based on DGNN is implemented.

[0035] The above one or more technical solutions have the following beneficial effects:

[0036] 1. This invention proposes a method and system for airport bird flock detection and early warning based on DGNN lidar. Based on the lidar deployed at the airport to acquire stable three-dimensional point cloud data, combined with multi-level preprocessing algorithms, it has the advantages of strong anti-light interference capability and adaptability to night and low visibility environments. It can achieve stable airport bird monitoring under complex weather conditions and solve the problem of decreased detection performance of traditional video surveillance and infrared imaging in harsh environments.

[0037] 2. This invention introduces the DGNN (Dynamic Graph Neural Network) method to construct a dynamic graph structure that integrates spatial adjacency relationships and temporal evolution relationships based on collected 3D point cloud data. This enables joint modeling of the spatiotemporal correlation of bird targets, effectively characterizing the group structure characteristics and movement consistency of bird flocks. Based on this, a comprehensive assessment of bird flock size, spatial distribution, and movement trends is conducted, providing reliable data support for bird strike risk assessment and graded early warning. By jointly analyzing the features of graph nodes and their interrelationships using DGNN, bird targets within the nodes are identified. Furthermore, a comprehensive assessment of bird flock size, spatial distribution, and proximity trends is performed, achieving a functional upgrade from "target detection" to "risk assessment." This enhances the foresight of bird strike risk early warning and addresses the shortcomings of bird flock behavior analysis and comprehensive risk assessment.

[0038] 3. This invention also designs a three-level early warning mechanism applicable to airport operation scenarios. The early warning information can be directly linked with the airport's existing operation management system and bird deterrence equipment, providing airport management personnel with intuitive and actionable decision-making basis, effectively improving the overall efficiency and intelligence level of airport bird strike risk prevention and control.

[0039] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0040] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0041] Figure 1 This is an overall flowchart of the DGNN-based lidar airport bird flock detection and early warning method in Embodiment 1 of the present invention. Detailed Implementation

[0042] It should be noted that the following detailed descriptions are exemplary and are intended only to describe specific embodiments and to provide further explanation of the invention, and are not intended to limit the scope of exemplary embodiments of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0043] Example 1

[0044] To address the shortcomings of existing airport bird monitoring technologies, such as insufficient stability under complex weather conditions, difficulty in accurately identifying low-altitude small bird targets, and lack of comprehensive assessment of overall bird flock behavior and risk profiles, this embodiment proposes a lidar-based airport bird flock detection and early warning method using DGNN (Directional Geographical Array Neural Network). This method is suitable for real-time bird monitoring of airport runways and surrounding areas at night or in low visibility conditions. It includes lidar point cloud data acquisition, point cloud data preprocessing, dynamic graph structure construction, bird target identification, and early warning information generation. Figure 1 As shown, the method specifically includes the following steps:

[0045] Step S1: Based on the lidar equipment deployed in the airport flight area and its surrounding areas, collect continuous three-dimensional point cloud data of the airport runway, taxiway and adjacent airspace, and preprocess the collected point cloud data to form a continuous multi-frame structured point cloud sequence.

[0046] Specifically, lidar equipment is deployed in and around the airport's flight area. LiDAR can acquire stable 3D point cloud data in low-visibility environments such as nighttime, fog, haze, rain, or snow. This lidar equipment is then used to acquire continuous 3D point cloud data covering the airport runway, taxiway, and adjacent airspace in real time. The acquired multi-frame point cloud data undergoes preprocessing, including noise filtering, density constraint removal, reflection intensity correction, ground point removal (i.e., ground segmentation), and time synchronization resampling (i.e., time alignment), to enhance the separability of bird targets in the point cloud. Ultimately, this results in a multi-frame structured point cloud sequence that can be used for spatiotemporal feature construction. The data acquisition and preprocessing process specifically includes:

[0047] Step S1.1: Point Cloud Acquisition and Data Representation. In this embodiment, one or more LiDARs are used, with a scanning frequency of not less than 10Hz, which can stably cover a spatial range of 200–300m. The point cloud data acquired during scanning... The original point cloud data of a frame is defined as follows:

[0048] ;

[0049] in, Let i be the coordinates of the i-th point. Reflection intensity, For timestamps, This represents the number of points in that frame.

[0050] To facilitate subsequent statistics and segmentation, the radial distance from the radar location to the radar is defined as follows, with the radar location as the origin:

[0051] ;

[0052] in, This represents the spatial distance between the laser point and the radar sensor, used for subsequent reflection intensity correction and distance-related feature calculation.

[0053] Step S1.2, noise filtering. First, for each point... Pick Nearest neighbor set The average nearest neighbor distance is calculated as follows:

[0054] ;

[0055] in, Indicates the current laser point; Indicates that it is located at The first in the nearest neighbor set One laser point; Indicates Euclidean distance; Point The average spatial distance between it and its nearest neighbors.

[0056] Next, the average nearest neighbor distance of all laser points is statistically analyzed to obtain the global mean. with standard deviation When the following conditions are met: When, determine the point Outlier noise points were identified and removed. This represents the number of nearest neighbors, with a value range of [value missing]. ; The global mean of the average nearest neighbor distance; The table represents the corresponding global standard deviation; This is an empirical coefficient used to adjust the strictness of outlier detection. The value is determined based on the statistical distribution characteristics of the average nearest neighbor distance of the point cloud. The preferred value range is 1.0 to 3.0, so as to ensure that noise points are effectively removed while avoiding excessive filtering of effective target points. This embodiment is not limited to a specific single value.

[0057] Step S1.3, Density Constraint Removal. After removing statistical outliers, to further suppress local sparse noise points, this embodiment introduces spatial density-based constraints into the point cloud. For each point... In a radius of The local point cloud density within its neighborhood is calculated as follows:

[0058] ;

[0059] in, Point Local point cloud density; The neighborhood radius calculated for density is used to define the local spatial extent; This represents the number of points that satisfy the condition.

[0060] On point The local density satisfies At this point, the point is identified as sparse noise and removed. After this step, the set of point clouds that satisfies the density constraint is obtained as follows:

[0061] ;

[0062] in, This represents the point cloud data after noise filtering in step S1.2; The density threshold ρ is used to distinguish between effective target points and sparse noise points. min Based on the statistical characteristics of local point cloud density, the preferred value range is [5, 20], so as to effectively suppress the interference of local sparse noise points while ensuring the integrity of the bird target point cloud. This embodiment is not limited to a specific single value.

[0063] Step S1.4, Reflection Intensity Correction. Considering that bird targets in lidar point clouds typically have weak reflection intensity and are easily obscured by the background, this embodiment performs point cloud reflection intensity correction. First, the reflection intensity is normalized as follows:

[0064] ;

[0065] in, This represents the normalized reflection intensity; and These represent the maximum and minimum values ​​of reflection intensity in the current point cloud frame, respectively.

[0066] Subsequently, distance attenuation compensation is introduced to correct the normalized reflection intensity, as follows:

[0067] ;

[0068] in, This indicates the reflection intensity after distance compensation; The radial distance from the point to the radar; This is a reference distance, used as a benchmark for reflection intensity compensation; The intensity compensation coefficient γ is determined based on the statistical characteristics of the attenuation of lidar reflection intensity with distance. Its preferred value range is [0.5, 2.0], so as to compensate for the attenuation of reflection intensity of distant targets while avoiding excessive amplification of the intensity of nearby point clouds. This embodiment is not limited to a specific single value.

[0069] After completing the above intensity correction, the point cloud data with corrected reflection intensity is obtained as follows:

[0070] .

[0071] Step S1.5, Ground Point Removal. To eliminate interference from the ground and runway surface on subsequent target detection, this embodiment performs ground point removal processing on the point cloud. First, based on the height threshold... The point cloud is coarsely segmented as follows:

[0072] ;

[0073] in, Represents the set of candidate ground points; Represents the set of candidate obstacle points; This is a height threshold used to initially distinguish between ground points and non-ground points. The value is determined based on the undulation characteristics of the airport runway and surrounding ground, and its preferred range is [0.2, 0.5], so as to effectively remove point clouds on the ground and runway surface while retaining low-altitude bird targets. This embodiment is not limited to a specific single value.

[0074] Subsequently, the candidate ground point set was... Performing RANSAC plane fitting yields the following planar model:

[0075] ;

[0076] Simultaneously, the distance from each point to the fitted plane is calculated as follows:

[0077] ;

[0078] in, For planar model parameters; Point The perpendicular distance to the fitted plane. When At that time, the corresponding point is identified as an in-ground point and removed, thereby updating the obstacle candidate point cloud. ,in The in-plane point determination threshold δ0 is determined based on the statistical characteristics of the plane fitting residual of the ground point cloud. Its preferred value range is [0.05, 0.20], so as to ensure the accuracy of ground point removal while avoiding the accidental removal of low-altitude bird target point clouds. This embodiment is not limited to a specific single value.

[0079] Step S1.6, Time Synchronization and Resampling. To ensure the consistency of multiple point cloud frames in the time dimension, this embodiment resamples the point cloud data according to a unified time axis, resulting in the following point cloud frame sequence:

[0080] ;

[0081] The sampling interval between adjacent frames is .

[0082] For points with time deviations, corrections are made using interpolation or merging methods to ensure they meet the following requirements: .

[0083] By performing unified coordinate transformation and time synchronization processing on point cloud data from different acquisition nodes, spatial deviations caused by acquisition from multiple devices and at multiple times are eliminated, ultimately outputting a structured point cloud frame sequence. This serves as the input for subsequent spatiotemporal feature construction and target recognition steps.

[0084] Step S2: Using points or point cloud clusters in the sequence that meet the preset spatial and intensity conditions as graph nodes, establish connecting edges between nodes based on spatial proximity, temporal continuity, and operational similarity to form a dynamic graph structure that evolves over time.

[0085] Specifically, the acquired multi-frame LiDAR point cloud data is transformed into a graph structure representation suitable for dynamic graph structure learning. By structurally modeling candidate bird reflector units and their spatiotemporal relationships (i.e., spatial proximity and temporal continuity) in the point cloud, the dynamic distribution of bird targets within the airport perimeter area is depicted, providing a stable and associative input data format for subsequent bird target identification and flock behavior analysis. In this embodiment, point cloud units composed of multiple LiDAR points satisfying spatial connectivity conditions are used as graph nodes for dynamic graph structure modeling, specifically:

[0086] Step S2.1: Point Cloud Node Construction. Specifically, the continuous multi-frame point cloud data obtained in step S1 is processed in chronological order. For the first frame... Frame point cloud data, its point set is denoted as:

[0087] ;

[0088] in, These represent the three-dimensional spatial coordinates of the point and the corrected reflection intensity, respectively.

[0089] In each frame of the point cloud, points or point cloud clusters (referred to as point cloud clusters) that meet the preset height range, reflection intensity threshold, and local density constraints are filtered, and the filtered points or point cloud clusters are used as graph nodes. Each graph node is denoted as:

[0090] ;

[0091] in, Indicates the spatial center position of the node. Let represent the node feature vector, and t be the node's timestamp. The node feature vector... It should include at least the following: the average reflection intensity of the point cloud corresponding to the node, the spatial distribution scale, and the local motion features estimated based on the changes in the center position of the nodes in adjacent frames. These motion features can be expressed as:

[0092] ;

[0093] in, This indicates the time interval between adjacent point cloud frames.

[0094] Through the above processing, the original multi-frame point cloud data is transformed into a node set consisting of multiple nodes, which can be represented as:

[0095] ;

[0096] The above node definitions provide a foundation for subsequent relationship modeling.

[0097] Step S2.2: Node Relationship Modeling and Edge Construction. Based on the constructed node set, connecting edges are established between nodes according to their spatial proximity and temporal correlation to characterize the continuity of bird targets in spatial and temporal dimensions.

[0098] Specifically, for any two nodes and If at least one of the following constraints is satisfied, then a connecting edge is established between the two. The constraints include:

[0099] (1) Spatial proximity constraint is:

[0100] ;

[0101] in, The spatial neighborhood radius r is a preset value. This spatial neighborhood radius r is determined based on the spatial resolution of the lidar point cloud and the scale characteristics of the bird target. Its preferred value range is [0.5, 3.0], so as to ensure the spatial correlation of nodes while avoiding the introduction of interference from too far irrelevant nodes. This embodiment is not limited to a specific single value.

[0102] (2) The constraints on time continuity and motion consistency are:

[0103] ;

[0104] in, This is the directional consistency threshold. The value is determined based on the characteristics of the bird target's movement direction change at adjacent times. The preferred value range is [15, 45], so as to ensure the stability of the motion continuity determination while avoiding the interruption of target association due to instantaneous direction disturbance. This embodiment is not limited to a specific single value.

[0105] (3) Consistency constraint for reflection intensity variation:

[0106] ;

[0107] in, This is the threshold for the change in reflection intensity. The value range is determined based on the normalized variation characteristics of the lidar point cloud reflection intensity, preferably [0.1, 0.3], so as to maintain the consistency of the reflection characteristics of bird targets while suppressing the interference of instantaneous high-frequency reflection noise such as raindrops and insects. This embodiment is not limited to a specific single value.

[0108] Next, each connecting edge To describe the strength of the association between nodes, the weight of the connection edge can be represented as a weighted combination of spatial distance, temporal continuity, and motion similarity between nodes. Spatial distance can be determined based on the distance relationship between the spatial centers of nodes; temporal continuity can be determined based on the difference between node timestamps; and motion similarity can be determined based on the displacement changes or consistency of motion directions of nodes in adjacent time frames. Based on this, according to the actual application scenario, the above indicators are normalized and then weighted according to preset weights to obtain the weight of the connection edge, which is used in the subsequent feature propagation and aggregation process.

[0109] In this embodiment, regarding spatial distance, for any two graph nodes from adjacent time frames, their spatial proximity is first determined based on the distance between their spatial center positions. When this distance is within a preset neighborhood range (i.e., a preset spatial neighborhood radius), a higher spatial association value is assigned. Conversely, when the distance approaches the neighborhood boundary, the corresponding spatial association value decreases. Regarding temporal continuity, the temporal continuity is determined based on the timestamp difference between the nodes. When two nodes come from adjacent time frames or the time interval is within a preset time window, they are considered to have a strong temporal continuity relationship and are assigned a higher temporal association value. Regarding motion similarity, by comparing the displacement direction or motion trend of nodes in adjacent time frames, when the difference in their motion directions is less than a preset threshold, they are considered to have high motion similarity and are assigned a higher motion association value. Finally, the spatial association value, temporal association value, and motion association value are normalized and weighted according to preset weights to obtain the weight of the corresponding connecting edge, which is used to characterize the overall association strength between nodes.

[0110] Step S2.3: Dynamic graph structure generation. As the point cloud data is continuously updated over time, the node set... and its corresponding set of connecting edges This also changes accordingly. By dynamically updating the nodes and connections at different times, a dynamic graph structure that evolves over time is constructed, which can be represented as:

[0111] ;

[0112] The constructed dynamic graph structure can simultaneously reflect the spatial distribution characteristics of bird targets and the motion evolution patterns in the temporal dimension, and can be used to depict the dynamic distribution and movement process of bird flocks in airport airspace.

[0113] Through the above steps, continuous multi-frame LiDAR point cloud data are transformed into a dynamic graph structure representation, and this dynamic graph structure is used as the input for subsequent bird target recognition and flock behavior analysis steps based on dynamic graph structure learning.

[0114] Step S3: Based on the dynamic graph structure, identify bird targets in the nodes by jointly analyzing the features of the graph nodes and their interrelationships through DGNN.

[0115] In this embodiment, the Dynamic Graph Neural Network (DGNN) used is a spatiotemporal dynamic graph network structure that integrates spatial relationship modeling and temporal evolution modeling. Its network hierarchy and other structures are the same as existing graph network architectures. This DGNN is used to perform feature learning and node discrimination on the dynamic graph constructed from LiDAR point clouds. Furthermore, the application methods of this DGNN functional modules are described in detail in the context of bird flock detection and early warning applications. The basic framework of the DGNN includes the following functional modules:

[0116] (1) Dynamic graph construction module (or data processing module): This module constructs a dynamic graph from the original acquired LiDAR point cloud data. Specifically, in this module, a dynamic graph structure that updates over time is constructed based on the constraints of spatial proximity, temporal continuity, and motion consistency of the point cloud nodes. Specifically, through a neighborhood search unit and a relationship determination unit, graph edge connections are established for node pairs that meet the spatial distance threshold and temporal continuity conditions. Each graph edge is assigned a corresponding spatial or temporal relationship identifier, thus forming a dynamic graph structure composed of nodes, edges, and relationships, providing a topological foundation for subsequent feature propagation and updates. This module corresponds to the execution process of step S2 above.

[0117] (2) The spatiotemporal graph learning module includes a spatial feature propagation module and a temporal feature update module. The spatial feature propagation module consists of at least one spatial feature aggregation layer, which is used to fuse the features of nodes that are in spatial neighborhood at the same time. Furthermore, the spatial feature aggregation layer performs weighted aggregation of the node's own features and the features of its neighboring nodes based on the adjacency relationship of nodes in the dynamic graph, and generates an updated node spatial feature representation through linear transformation and nonlinear mapping units to characterize the feature information of bird targets in local spatial distribution and group structure. The temporal feature update module consists of a time series feature update layer, which is used to fuse and update the node features that are related in adjacent time periods. Furthermore, the time series feature update layer combines the node features of the current time with the corresponding node features of the previous time based on the connection relationship of nodes across time periods in the dynamic graph, and generates a node representation that reflects the characteristics of temporal evolution through feature update units, thereby characterizing the continuous change trend of node features in the time dimension and enhancing the continuity and stability of target recognition results in the time series.

[0118] (3) The node discrimination module consists of a feature mapping layer and a discrimination layer, which is used to discriminate targets on the node representation after fusing spatial and temporal features. Furthermore, the feature mapping layer performs dimensional transformation on the node features through a linear mapping unit, and the discrimination layer outputs the confidence value of the corresponding node belonging to the bird target through an activation function, thereby realizing the determination of the target attribute of the node.

[0119] Through the synergistic effect of the above modules, stable identification of bird target nodes and continuous tracking of them in time series can be achieved.

[0120] As a further technical solution, the point cloud data used in this embodiment is acquired by a VLP-32C lidar. This lidar has multi-line scanning capabilities and can stably acquire three-dimensional point cloud data and reflection intensity information of small low-altitude targets. Based on the acquisition of several frames of lidar point cloud data, the acquired point cloud data is labeled manually or semi-automatically, and divided according to time series or data segments. After preprocessing (i.e., the preprocessing process in step S1 above), sample data labeled as either bird or not is formed, constructing a sample dataset. Next, this sample dataset is divided, with approximately 70% of the sample data used to construct a training dataset for parameter learning of the dynamic graph neural network model, and the remaining approximately 30% of the sample data used as a validation test dataset for model parameter adjustment and technical effect verification. Preferably, during the data division process, the training dataset and the validation test dataset are ensured not to overlap in the time dimension to avoid data leakage affecting the validation results.

[0121] In the actual detection process, based on the dynamic graph structure constructed in step S2, the trained Dynamic Graph Neural Network (DGNN) is used to jointly analyze the relationships between each graph node and its neighboring nodes, thereby automatically distinguishing between avian and non-avian targets and outputting the spatial and temporal distribution results of avian targets, providing reliable input for subsequent bird flock behavior analysis and risk assessment. The specific process includes:

[0122] Step S3.1: Graph Node Feature Update and Information Propagation. The spatiotemporal graph learning module based on DGNN (including a spatial feature propagation module and a temporal feature update module) addresses the dynamic graph structure constructed in step S2, performing feature updates for each time step. graph nodes By combining the node's own characteristics with the characteristics of its neighboring nodes, the node's features are updated through a graph structure feature propagation mechanism. Let node... The initial feature vector is The set of its adjacent nodes is denoted as Then the node feature update process can be represented as:

[0123] ;

[0124] in, Represents a node With neighboring nodes The correlation weight between them This represents the feature fusion function.

[0125] During feature propagation, nodes incorporate information related to spatial proximity, temporal continuity, and motion similarity from their neighboring nodes. This allows node features to reflect not only individual attributes but also their relative position within the overall structure of the bird flock. Through multiple rounds of feature propagation and aggregation, node features gradually enhance their ability to express the structural characteristics and motion states of bird flocks.

[0126] Step S3.2, Bird Target Identification and Node Classification. After completing the node feature update, the DGNN-based node identification module performs target identification operations on each node based on the updated node feature vectors to determine whether the node belongs to a bird target. Specifically, for nodes... Based on its updated feature vector The result of determining whether a target belongs to the bird category is calculated using a discriminant function, which can be expressed as:

[0127] ;

[0128] in, This represents the confidence value of the bird target corresponding to the node. This represents the node discrimination function.

[0129] In this embodiment, the node discrimination function A probabilistic output format is used to map node features to their confidence values ​​for belonging to the bird target category. Specifically, the node discrimination function can be expressed as:

[0130] ;

[0131] in, is the feature vector obtained by the node at time t after dynamic graph structural feature propagation and aggregation; W and b are the weight parameters and bias terms of the discriminant function, respectively; σ(·) represents the Sigmoid activation function, used to map the output to the interval [0,1]. This confidence value is used to characterize the probability that the node belongs to the bird target and serves as the basis for subsequent node determination and bird flock target aggregation.

[0132] In other alternative implementations, the node discrimination function It can also be implemented using equivalent probability output forms such as Softmax, but this embodiment does not limit it to this.

[0133] Based on the confidence level and the preset threshold τ c The comparison yields the target determination result for each node. This result includes at least: whether the node belongs to the bird target category, the corresponding confidence level, and the node's continuous existence in the time series. The confidence level is preset to a threshold τ. cThis is used to distinguish between bird target nodes and non-target nodes. The value is determined based on the statistical characteristics of the node confidence distribution, and its preferred range is [0.5, 0.8]. When the node confidence value is greater than or equal to a preset threshold, the corresponding node is identified as a bird target node; when the node confidence value is lower than the threshold, it is identified as a non-bird target node. This embodiment is not limited to a specific single value.

[0134] Through the above discrimination process, the bird target node and the environmental noise node (including raindrops, insects and ground reflections) can be effectively distinguished.

[0135] Step S3.3: Bird Flock Target Aggregation and Continuity Constraints. Based on the single-node discrimination results, bird target nodes with continuous relationships in both spatial and temporal dimensions are aggregated to form a set of candidate bird flock targets. Specifically, let's assume that at time... The set of nodes identified as bird targets is:

[0136] ;

[0137] in, The target discrimination threshold τ is used to distinguish between bird target nodes and non-target nodes, and its value is determined based on the statistical distribution characteristics of node discrimination confidence. In this embodiment, the preferred range of the target discrimination threshold τ is: τ∈[0.5, 0.8]. When the value is greater than or equal to the threshold τ, the corresponding node is identified as a bird target node; when... If the value is less than the threshold τ, it is determined to be a non-bird target node.

[0138] The threshold setting takes into account both the accuracy and robustness requirements of bird target recognition. Under complex weather conditions or with high background noise, the threshold can be appropriately increased to suppress false detections. When the bird target is small or has weak reflectivity, the threshold can be appropriately decreased to enhance detection sensitivity. This invention is not limited to a specific single value.

[0139] Based on this, the set Nodes that are consistent in spatial location and movement trend are aggregated to construct a set of candidate targets for bird flocks.

[0140] Furthermore, for the candidate target set of bird flocks, in order to improve the stability of the recognition results, a time continuity constraint is introduced. Nodes or node aggregates that are continuously identified as bird targets in multiple consecutive frames are given higher confidence, thereby suppressing the interference of occasional noise points on the detection results, optimizing the identified bird targets, and forming a bird target set.

[0141] Finally, the set of bird targets at each moment and their distribution in spatial and temporal dimensions are output as input for subsequent bird flock behavior analysis and risk assessment steps.

[0142] Based on the above scheme, in the bird target identification stage, the graph nodes identified as bird targets are initially aggregated at the node level to enhance the continuity of the target identification results in the spatial and temporal dimensions. Then, the interference of occasional noise nodes is suppressed by temporal continuity constraints, which can stabilize the target identification results and denoise the results, thus ensuring the accuracy of the final bird target identification.

[0143] Step S4: Aggregate the identified bird targets to construct a bird flock cluster set, combine it with the bird strike risk assessment model, calculate the potential risk value of the bird flock cluster to airport operations, and issue graded warnings based on the risk value.

[0144] Specifically, the bird target identification result output in step S3 is further improved from "whether bird targets exist" to "whether bird activities pose a safety risk to airport operations". That is, based on the spatiotemporal distribution results of bird targets obtained by dynamic graph structure learning (distribution characteristics in spatial and temporal dimensions), the spatial proximity, size density and movement trend of bird flocks are comprehensively analyzed to construct a unified risk assessment model, assess the potential risk level of bird flocks relative to the runway area, and generate graded early warning results applicable to airport operation management to assist airport operation scheduling and safety decision-making.

[0145] Step S4.1: Bird Cluster Construction and Size / Density Assessment. Based on the confirmed bird target set output in Step S3, bird targets with continuous relationships in spatial location and time are aggregated to construct bird cluster sets. Each bird cluster represents a set of multiple bird targets that are spatially continuous in their distribution at the same or adjacent moments after bird target identification is completed. It is used to characterize the size, spatial distribution range and overall behavioral characteristics of the bird flock, providing a basis for subsequent risk assessment calculations.

[0146] In this embodiment, to characterize the size of bird flocks, the bird flock size density index is defined as follows, based on the number of targets contained in the bird cluster and their spatial distribution range:

[0147] ;

[0148] in, Indicates the first The number of bird targets contained in a bird flock cluster This represents the projected area of ​​the bird flock on the horizontal plane. This is a preset lower bound for the area to prevent abnormal density calculations caused by excessively small flock sizes. This flock density index reflects the size of the flock and its potential impact range.

[0149] Step S4.2, Spatial Proximity and Motion Trend Assessment. After obtaining the bird flock aggregation, the overall spatial position of the flock and its motion relative to the airport flight area are further calculated. Specifically, let time be... Below, the centroid of the bird flock is located at... Based on this centroid location, calculate the minimum spatial distance from it to the airport runway centerline. This serves as an indicator of the spatial proximity of bird flocks. Simultaneously, to characterize the approach trend of bird flocks relative to the flight zone, the approach speed of the bird flock is defined based on the change in the centroid distance between adjacent time points:

[0150] ;

[0151] in, This represents the time interval between adjacent point cloud frames. This approach speed reflects whether the flock of birds is trending towards critical flight areas, and as a motion trend indicator, it is an important basis for subsequent risk assessment.

[0152] Step S4.3: Risk Function Construction and Graded Early Warning Output. Specifically, by integrating bird flock size and density indicators, spatial proximity indicators, and movement trend indicators, a bird flock risk assessment function is constructed to quantify the potential threat level of bird activity to airport operational safety. Its risk value is defined as:

[0153] ;

[0154] in, This represents the minimum spatial distance from the center of mass of the flock of birds to the center line of the runway. Indicators of bird flock density This represents a motion trend index obtained by normalizing the approximate velocity. These are the corresponding weighting coefficients.

[0155] Based on the bird flock risk value calculated using the above method, and compared with a preset threshold, the bird flock activity status can be divided into three warning levels: safe, concerning, and dangerous, and corresponding warning records can be generated. These warning records can be sent to the airport operation management system for use in bird deterrence equipment linkage, operational scheduling assistance, or safety alert dissemination, thereby assisting the airport in carrying out bird strike risk prevention and control work under complex weather conditions and low visibility conditions.

[0156] Preferably, to further verify the superiority of the method proposed in this embodiment, a verification test dataset was constructed using point cloud data collected by a VLP-32C lidar to verify the above-mentioned bird flock detection method based on dynamic graph neural networks. The density-based point cloud clustering method (DBSCAN), commonly used in the prior art, was selected as a comparative example and compared with this method. The method was evaluated in terms of the number of detected target clusters, noise filtering rate, adaptability to sparse point clouds, single-frame processing time, and detection consistency. The experimental results are shown in Table 1 below.

[0157] Table 1. Experimental comparison results between the traditional DBSCAN method and this method in airport lidar bird flock detection tasks.

[0158]

[0159] Traditional DBSCAN methods primarily rely on local density thresholds in point clouds for clustering. In situations with uneven point cloud density or sparse targets (such as airport scenarios), this can easily lead to the merging of bird targets with background noise into larger clusters, or misclassification of real bird targets consisting of only a few points as noise points. In contrast, the method proposed in this embodiment constructs a dynamic graph structure that integrates spatial proximity and temporal continuity, introducing graph edge topological constraints to model the associations of sparse nodes. This enables accurate identification of bird targets composed of a small number of laser points in complex backgrounds, significantly improving the stability and consistency of low-altitude small bird target detection. As shown in Table 1, with a similar processing time per frame, this method significantly outperforms the traditional DBSCAN method in terms of the number of detected target clusters, noise filtering capability, and adaptability to sparse point clouds. It achieves more stable and accurate bird target identification in complex airport environments, validating the effectiveness and engineering feasibility of this method in practical applications.

[0160] The above verification results show that this embodiment, by constructing a dynamic graph structure that integrates spatial proximity and temporal continuity, and by using a dynamic graph neural network to jointly model the spatiotemporal correlation between nodes, can achieve stable identification of bird targets under conditions of sparse point clouds, complex backgrounds, and noise interference, and maintain the continuity and consistency of detection results in the time series.

[0161] Based on this, by performing aggregate analysis on the identified bird targets, candidate bird flocks can be formed, and their size characteristics, spatial distribution, and movement trends can be evaluated, providing a reliable data foundation for subsequent bird strike risk assessment and graded early warning. The above verification results demonstrate that the method of this invention has good feasibility and effectiveness in practical applications.

[0162] Example 2

[0163] This embodiment proposes a lidar-based airport bird flock detection and early warning system based on DGNN, which specifically includes the following modules:

[0164] The data acquisition module includes lidar equipment deployed in the airport flight area and its surrounding areas, used to collect continuous three-dimensional point cloud data of the airport runway, taxiway and adjacent airspace, and to preprocess the collected point cloud data to form a continuous multi-frame structured point cloud sequence.

[0165] The dynamic graph structure construction module is used to establish connecting edges between nodes based on spatial proximity, temporal continuity, and operational similarity, using points or point cloud clusters in the sequence that meet preset spatial and intensity conditions as graph nodes, thus forming a dynamic graph structure that evolves over time.

[0166] The bird target recognition module is used to identify bird targets in nodes based on a dynamic graph structure and by jointly analyzing the features of graph nodes and their interrelationships using DGNN.

[0167] The early warning module is used to aggregate identified bird targets to construct bird clusters, combine them with a bird strike risk assessment model, calculate the potential risk value of bird clusters to airport operations, and issue graded early warnings based on the risk value.

[0168] Example 3

[0169] This embodiment provides an electronic device, including: a memory for storing executable instructions; and a processor for executing the executable instructions stored in the memory to implement the method provided in this embodiment.

[0170] Example 4

[0171] This embodiment also provides a computer-readable storage medium storing executable instructions, which, when executed by a processor, will cause the processor to execute the method described above in this embodiment.

[0172] Example 5

[0173] This embodiment provides a computer program product including executable instructions, which are computer instructions; the executable instructions are stored in a computer-readable storage medium. When the processor of an electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, the electronic device performs the method described in this embodiment.

[0174] The steps and methods involved in Embodiments 2 to 5 above correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0175] The modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0176] The above description is only a preferred embodiment of the present invention. Although the specific implementation of the present invention has been described in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or variations that can be made by those skilled in the art without creative effort are still within the scope of protection of the present invention.

Claims

1. A method for detecting and warning of bird flocks at airports using lidar based on DGNN, characterized in that, include: Based on lidar equipment deployed in the airport flight area and its surrounding areas, continuous three-dimensional point cloud data of airport runways, taxiways and adjacent airspace are collected, and the collected point cloud data is preprocessed to form a continuous multi-frame structured point cloud sequence. Using points or point cloud clusters in the sequence that satisfy preset spatial and intensity conditions as graph nodes, connections between nodes are established based on spatial proximity, temporal continuity, and operational similarity, forming a dynamic graph structure that evolves over time. The construction of the dynamic graph structure based on a continuous multi-frame structured point cloud sequence includes: In each frame of point cloud, points or point cloud clusters that meet the preset height range, reflection intensity threshold, and local density constraints are selected as graph nodes. The graph nodes are recorded as the spatial center position of the node, the node feature vector, and the node timestamp. The node feature vector includes the average reflection intensity of the point cloud corresponding to the node, the spatial distribution scale, and the local motion features estimated based on the changes in the node center position in adjacent frames. Based on the constructed set of nodes, connecting edges between nodes are established according to the spatial proximity and temporal correlation between nodes. For any two nodes, a connecting edge is established if at least one constraint condition is met. The edge weight is a weighted combination of spatial distance, temporal continuity and motion similarity between nodes. Among them, the constraints include: spatial proximity constraint, temporal continuity and motion consistency constraint, and reflection intensity change consistency constraint. The nodes and their connections at different times are dynamically updated to construct a dynamic graph structure that evolves over time. Based on the dynamic graph structure, bird targets in the nodes are identified by jointly analyzing the features of the graph nodes and their interrelationships using DGNN. The identified bird targets are aggregated to construct bird flock clusters. Combined with a bird strike risk assessment model, the potential risk value of these bird flock clusters to airport operations is calculated, and graded warnings are issued based on the risk value. The calculation of the potential risk value includes: Based on the identified set of bird targets, bird targets with continuous correlation in spatial location and time dimension are aggregated to construct a bird flock cluster set; Based on the number of targets contained in a bird flock and their spatial distribution range, calculate the bird flock size density index, the bird flock spatial proximity index, and the movement trend index. By combining bird flock size and density indicators, spatial proximity indicators, and movement trend indicators, a bird strike risk assessment model is constructed. The values ​​of the three indicators are weighted and calculated to obtain the final potential risk value of bird flocks to airport operations.

2. The method for airport bird flock detection and early warning based on DGNN lidar as described in claim 1, characterized in that, The 3D point cloud data of frame t includes the 3D coordinates, reflection intensity, and timestamp of each point in the current frame; the collected point cloud data is preprocessed, including noise filtering, density constraint removal, reflection intensity correction, ground point removal, and time synchronization resampling.

3. The method for airport bird flock detection and early warning based on DGNN lidar as described in claim 1, characterized in that, Based on a dynamic graph structure, bird targets within the nodes are identified by jointly analyzing the features and relationships of the graph nodes using DGNN, including: Based on a dynamic graph structure, the features of each node are combined with the features of its neighboring nodes, and the features of the current node are updated through the graph structure feature propagation mechanism. Based on the updated node features, a discriminant function is used to identify targets in each node, and bird targets in the nodes are identified. After bird flock target aggregation and continuity constraints, the set of bird targets at each time step and their distribution results in the spatial and temporal dimensions are output.

4. A DGNN-based lidar airport bird flock detection and early warning system, implemented based on the DGNN-based lidar airport bird flock detection and early warning method according to any one of claims 1-3, characterized in that, include: The data acquisition module includes lidar equipment deployed in the airport flight area and its surrounding areas, used to collect continuous three-dimensional point cloud data of the airport runway, taxiway and adjacent airspace, and to preprocess the collected point cloud data to form a continuous multi-frame structured point cloud sequence. The dynamic graph structure construction module is used to take points or point cloud clusters that meet preset spatial and intensity conditions in the sequence as graph nodes, and establish connecting edges between nodes based on spatial proximity, temporal continuity, and operational similarity to form a dynamic graph structure that evolves over time. The bird target recognition module is used to identify bird targets in nodes based on a dynamic graph structure and by jointly analyzing the features of graph nodes and their interrelationships using DGNN. The early warning module is used to aggregate identified bird targets to construct bird flock clusters, combine them with a bird strike risk assessment model, calculate the potential risk value of bird flock clusters to airport operations, and issue graded early warnings based on the risk value.

5. An electronic device, characterized in that, include: Memory, used to store executable instructions; The processor, when executing executable instructions stored in the memory, implements the DGNN-based lidar airport bird flock detection and early warning method according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The device stores executable instructions that, when executed by a processor, implement the DGNN-based lidar method for airport bird flock detection and early warning as described in any one of claims 1-3.

7. A computer program product, characterized in that, The computer program product includes executable instructions stored in a computer-readable storage medium; When the processor of the electronic device reads the executable instructions from the computer-readable storage medium and executes the executable instructions, it implements the DGNN-based lidar airport bird flock detection and early warning method according to any one of claims 1-3.

Images