Airport bird detection radar monitoring method

Abstract

The invention belongs to the technical field of airport aviation safety perception, and particularly relates to an airport bird detection radar monitoring method, which comprises the steps of preprocessing acquired point cloud data; based on the preprocessed point cloud data, point cloud multi-scale features are extracted through a PointNeXt model in sequence, and birds are identified; the method comprises the following steps: constructing a trajectory tracking model based on multi-frame point cloud multi-scale features, associating adjacent frame point cloud features by using a Transform cross-frame attention mechanism, and predicting a matched flight trajectory in combination with a bird movement prior model, and comprises the following steps: forming a cross-frame feature by the adjacent frame point cloud multi-scale features through Transform coding; on the basis of the cross-frame feature and the query vector, calculating a trajectory association feature and a new target feature; forming a corresponding track according to the track correlation feature and the new target feature, and obtaining a first prediction position of the next frame; obtaining a second prediction position of the next frame through a motion prior model based on the position information of the two adjacent frames; and obtaining a final monitoring and tracking trajectory through weight fusion of the first prediction position and the second prediction position.

Description

Technical Field

[0001] The present invention belongs to the technical field of airport aviation safety perception, and in particular relates to an airport bird-detecting radar detection method. Background Art

[0002] In aviation, especially airport safety, bird strikes have always been a significant threat to flight safety. With the continued expansion of the global air transport industry and the increasing frequency of flights, changes in the ecological environment surrounding airports have led to an increasing overlap between bird activity areas and aircraft take-off and landing areas. Currently, airport bird monitoring has become a key operational management priority for major airports worldwide. These airports not only invest significant human and material resources in bird activity monitoring, but also actively explore innovative technological approaches in the hope of building more efficient and accurate bird monitoring systems to ensure aviation safety.

[0003] In the field of radar bird detection, traditional radars rely on artificially designed features (such as RCS and Doppler frequency) to identify bird targets. However, their detection capabilities for small birds (such as sparrows) and those in complex backgrounds (such as vegetation and building echoes) are insufficient, making them prone to missed detections and false detections. Furthermore, radar monitoring data processing often uses traditional algorithms, which have high computational complexity and are unable to meet airports' needs for real-time monitoring and rapid response to bird targets. Although deep learning technology has begun to be applied to radar data processing in recent years, existing methods have poor adaptability to the sparsity and high noise characteristics of lidar point cloud data. Furthermore, in the multi-target tracking phase, complex bird flight paths (such as rapid turns and flocking movements) make it difficult to achieve stable trajectory association and continuous tracking. Consequently, existing bird detection radar monitoring technology relies on artificially designed features, is prone to missed detections of small birds, false detections in complex backgrounds, and suffers from low data processing efficiency. Therefore, more efficient and accurate radar bird detection technology solutions are urgently needed. Summary of the Invention

[0004] The present invention is proposed based on the above-mentioned requirements of the prior art. The technical problem to be solved by the present invention is to provide an airport bird detection radar monitoring method to improve the accuracy.

[0005] In order to solve the above problems, the technical solutions provided by the present invention include:

[0006] Provided is an airport bird-detecting radar monitoring method, comprising: preprocessing acquired point cloud data; extracting multi-scale features of the point cloud through a PointNeXt model based on the preprocessed point cloud data, and identifying birds; correlating point cloud features of adjacent frames based on the multi-scale features of the point cloud of multiple frames using a Transformer cross-frame attention mechanism, and predicting and matching bird flight trajectories in combination with a bird motion prior model, comprising: encoding the multi-scale features of the point cloud of adjacent frames through a Transformer to form cross-frame features, and forming a query vector; generating an autoregressive trajectory query vector based on recorded bird targets, and simultaneously generating a static object query vector to detect new bird targets; calculating trajectory association features and new target features based on the cross-frame features and the query vector; forming a trajectory based on the trajectory association features and the new target features to obtain a first predicted position of the next frame; obtaining a second predicted position through a motion prior model based on position information of two adjacent frames; and obtaining a final predicted tracking trajectory by weighted fusion of the first predicted position and the second predicted position.

[0007] Preferably, the preprocessing of point cloud data includes removing outliers through voxel filtering, eliminating background clutter through statistical filtering, and optimizing the lidar point cloud data into an ordered voxel grid or feature vector adapted to the deep learning model through data enhancement and format conversion.

[0008] Preferably, the method of extracting multi-scale features of point clouds based on pre-processed point cloud data by using a PointNeXt model and identifying birds further includes optimizing the detection results by using an adaptive threshold NMS algorithm and calculating the point cloud density of the detection area, which is expressed as: Where ρ is the point cloud density of the detection area, N is the number of point clouds in the area, and V is the volume of the area. The NMS suppression threshold is dynamically adjusted based on the regional point cloud density, expressed as: T = T base +β·(ρ-ρ avg ), where T is the dynamically adjusted non-maximum suppression threshold, T base is the basic threshold, β is the adjustment coefficient, ρ avg The average point cloud density; when the point cloud density is high, the threshold is automatically lowered to retain more overlapping detection frames, and when the point cloud density is low, the threshold is increased to reduce false detections.

[0009] Preferably, the calculation of the trajectory association feature and the new target feature based on the cross-frame feature and the query vector includes expressing the trajectory association feature as follows: Among them, Attn track is the trajectory association feature, is the attention weight of the autoregressive trajectory query vector and key K, d k is the dimension of the corresponding autoregressive trajectory query vector, is the autoregressive trajectory query vector, K is the key, and V is the value; the new target feature is expressed as: Among them, Attn obj is the new target feature, is the attention weight of the static object query vector and key K, d k is the dimension of the corresponding static object query vector, Q obj It is a static object query vector, K is the key and V is the value.

[0010] Preferably, forming a trajectory based on the trajectory association features to obtain the first predicted position of the next frame includes: mapping the trajectory association features into the bounding box coordinates and category label of the target in the t+1 frame according to a multi-layer perceptron, expressed as: (x t+1 ,y t+1 , z t+1 , c t+1 )=MLP bbox (ATTn track ), where (x t+1 ,y t+1 , z t+1 ) is the bounding box coordinate, c t+1 For the classification of birds; continuous tracks are formed through adjacent frames.

[0011] Preferably, the second predicted position is obtained by using a motion prior model based on the position information of two adjacent frames, including: predicting the second predicted position by using a uniform acceleration model based on the positions of two historical adjacent frames, wherein the second predicted position is expressed as (x' t+1 , y' t+1 , z' t+1 ), where x' t+1 =x t +v xt ·Δt+0.5a xt ·(Δt) 2 , y' t+1 =y t +v yt ·Δt+0.5a yt ·(Δt) 2 , z' t+1 =z t +v zt ·Δt+0.5a zt ·(Δt) 2 ; Where Δt is the frame interval time, v xt is the velocity in the x direction, a xt is the acceleration in the x direction, v yt is the velocity in the y direction, a yt is the acceleration in the y direction, v zt is the velocity in the z direction, a zt is the acceleration in the z direction.

[0012] Preferably, the voxelization filter divides the three-dimensional space into a uniform voxel grid, and the resolution of the voxel grid in the x, y, and z directions is v x 、v y 、v z , for any point P in the point cloud data i =(x i 、y i 、z i ), the index of the voxel to which it belongs (u i 、v i 、w i ) is calculated using the following formula: Among them, (x i 、y i 、z i ) is the coordinate of any point Pi on the x, y, and z axes, (x min 、y min 、z min ) is the minimum coordinate value of the point cloud data in the x, y, and z axis directions.

[0013] Preferably, the statistical filtering removes background clutter and calculates each point p i The Euclidean distance d to its k neighborhood points i , Assuming that the distance from a point to its neighboring points follows a Gaussian distribution, the mean μ and standard deviation σ are calculated, and the distance threshold T = μ + ασ is set, where α is 2-3. Points exceeding the threshold are judged as background clutter and removed.

[0014] Preferably, the PointNeXt model includes a basic feature extraction unit, through To achieve the extraction of local geometric features, where R is the maximum pooling operation, j:(i,j)∈N is the neighborhood of point i, h θ is a multi-layer perceptron, is the original feature, is the relative coordinate information of the point cloud, is the position coordinate of point j in the lth layer, is the position coordinate of point i in the lth layer, r l is the normalization factor, is the original eigenvector of point i in the l+1th layer.

[0015] Preferably, the final predicted position is obtained by weighted fusion of the first predicted position and the second predicted position, which is expressed as: final =λx t+1 +(1-λ)x′ t+1 ,y final =λy t+1 +(1-λ)y′t+1 , z final =λz t+1 +(1-λ)z′ t+1 , where (x final ,y final , z final ) is the final predicted position, x t+1 is the x-coordinate of the first predicted position, x′ t+1 is the x coordinate of the second predicted position, y t+1 is the y coordinate of the first predicted position, y′ t+1 is the y coordinate of the second predicted position, z t+1 is the z coordinate of the first predicted position, z′ t+1 is the z coordinate of the second predicted position, and λ is the weight; a predicted tracking trajectory is formed according to the final predicted position.

[0016] Compared with the existing technology, the present invention significantly improves the quality of point cloud data by combining voxel filtering with statistical filtering; adopts the PointNeXt model to replace the traditional detection algorithm, and uses the hierarchical Transformer encoder and adaptive NMS to accurately identify bird targets of different sizes; in terms of tracking, the trajectory tracking model proposed by the present invention breaks through the limitations of traditional methods in tracking complex bird trajectories, and uses the cross-frame attention mechanism and motion prior model to achieve real-time and continuous tracking. The overall technical solution achieves comprehensive improvements in detection accuracy, processing efficiency and tracking stability. In summary, the present invention provides an airport bird detection radar monitoring method. This method breaks through the bottlenecks of traditional radar bird detection in accuracy, efficiency and complex trajectory processing through cloud preprocessing technology, PointNeXt precise detection model and predictive stable tracking mechanism, and significantly improves the airport bird target monitoring and prevention capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0018] Figure 1 The figure is a flowchart of the steps of an airport bird detection radar monitoring method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. 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 creative efforts shall fall within the scope of protection of the present invention.

[0020] In the description of the embodiments of the present invention, it should be noted that, unless otherwise specified or limited, the term "connected" should be understood in a broad sense. For example, it can mean a fixed connection, a detachable connection, or an integral connection. It can be a mechanical connection, an electrical connection, a direct connection, or an indirect connection through an intermediate medium. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0021] The terms "top," "bottom," "above," "below," and "on" used throughout the description refer to relative positions of components of a device, such as the relative positions of top and bottom substrates within a device. It will be understood that devices are multifunctional regardless of their orientation in space.

[0022] To facilitate understanding of the embodiments of the present invention, specific embodiments will be further explained below with reference to the accompanying drawings. The embodiments do not limit the embodiments of the present invention.

[0023] This embodiment provides an airport bird detection radar monitoring method, such as Figure 1 shown.

[0024] The airport bird detection radar monitoring method comprises:

[0025] Get point cloud data.

[0026] Preprocessing point cloud data includes removing outliers through voxel filtering, removing background clutter through statistical filtering, and optimizing lidar point cloud data into an ordered voxel grid or feature vector suitable for deep learning models through data enhancement and format conversion.

[0027] Voxel Grid Filter divides the three-dimensional space into a uniform voxel grid, thereby achieving efficient removal of outliers.

[0028] Specifically, let the resolution of the voxel grid in the x, y, and z directions be v x 、v y 、v z , for any point P in the point cloud data i =(x i 、yi 、z i ), the index of the voxel to which it belongs (u i 、v i 、w i ) is calculated using the following formula:

[0029]

[0030] Among them, (x i 、y i 、z i ) is the coordinate of any point Pi on the x, y, and z axes, (x min 、y min 、z min ) is the minimum coordinate value of the point cloud data in the x, y, and z axes. The point cloud data within each voxel is compressed to a single representative point, such as the centroid.

[0031] In airport bird spotting scenarios, the raw point cloud data collected by LiDAR often contains isolated noise points caused by equipment jitter and electromagnetic interference, which can interfere with subsequent target identification. Therefore, outliers are removed through voxel filtering. This operation not only reduces the amount of data and the computational complexity of subsequent processing, but also effectively removes isolated outliers in space by eliminating voxels containing a small number of outliers, laying the foundation for accurate bird detection. For example, in the raw point cloud data collected near an airport runway, occasional scattered noise points are quickly eliminated after voxel filtering, preserving the point cloud information of the actual bird targets.

[0032] Statistical Outlier Removal is used to remove background clutter. This method calculates the value of each point p i The Euclidean distance d to its k neighborhood points i , the calculation formula is as follows:

[0033]

[0034] Among them, p ij are the j area points corresponding to point pi.

[0035] Assume that the distance from a point to its neighboring points follows a Gaussian distribution, calculate the mean μ and standard deviation σ, set the distance threshold T = μ + ασ (α is usually 2-3), and determine the points exceeding the threshold as background clutter and remove them.

[0036] Airports are complex environments, and static objects like trees and buildings generate a large amount of point clouds, creating background clutter that can seriously interfere with bird detection. For example, the densely distributed point cloud of trees surrounding the airport presents distinct characteristics distinct from those of bird point clouds. Statistical filtering effectively separates this background clutter, highlighting the point cloud features of bird targets and significantly improving the accuracy of subsequent bird recognition algorithms.

[0037] In addition, a starting point is selected from the filtered point cloud data P, and subsequent sampling points are selected using the farthest point sampling method (FPS). This process ensures that the point cloud data of key parts of the target is preserved while preserving the overall geometric structure of the point cloud. At the same time, the sampled point cloud is randomly rotated and scaled. Random rotation is achieved by rotating the matrix around the x, y, and z axes, and the scaling operation transforms the coordinates using the scaling factor s, where the scaling factor is randomly selected between [0.8 and 1.2].

[0038] By simulating the point cloud morphology of birds in different flight postures and at different radar distances, the diversity of training data is greatly increased, enabling the deep learning model to learn the characteristics of bird targets at various angles and scales, reducing missed detections and false detections in practical applications. This improves the deep learning model's adaptability to different scenarios and enhances the point cloud data through the above processing.

[0039] The processed point cloud data is encoded using the PointCLIP model to form a high-dimensional feature vector. This vector not only contains the geometric shape information of the bird target, but also incorporates semantic features such as wing shape and flight posture. This provides more discriminative input for subsequent target recognition and tracking algorithms based on the Transformer architecture, helping the model to accurately distinguish birds from other low-altitude objects such as drones and kites. This format conversion process ensures that the processed point cloud data meets the input requirements of deep learning models.

[0040] Based on point cloud preprocessing data, the PointNeXt model is used to extract multi-scale features of point clouds through a hierarchical Transformer encoder, combined with adaptive threshold NMS, to accurately detect birds of different sizes in airport environments.

[0041] Initial point cloud features are extracted through the PointNeXt model infrastructure.

[0042] The PointNeXt model adopts a U-Net-like architecture, whose core components are the collection abstraction (SA) block and the inverted residual MLP (InvResMLP) block, which play a key role in the point cloud feature extraction process of birds.

[0043] Among them, the SA block is used as the basic feature extraction unit to extract local geometric features through the following formula:

[0044]

[0045] Among them, R is the maximum pooling operation, j:(i,j)∈N is the neighborhood of point i, h θ is a multi-layer perceptron, is the original feature, is the relative coordinate information of the point cloud, is the position coordinate of point j in the lth layer, is the position coordinate of point i in the lth layer, r l is the normalization factor, is the original eigenvector of point i in the l+1th layer.

[0046] Specifically, in each operation, the maximum pooling operation represented by R can filter out the most representative features from the neighborhood {j:(i,j)∈N} of point i, effectively avoiding the interference of redundant information. For example, when detecting the point cloud of bird wings, the maximum pooling operation can highlight the key geometric feature points on the edge of the wing and discard those points that contribute less to the overall shape description. The shared multi-layer perceptron h θ , the relative coordinate information of the point cloud With the original features The combination of these two methods enables the model to not only perceive the location information of the points, but also capture the spatial relationship between points, thereby accurately extracting the local geometric structure of the bird target.

[0047] In actual airport operations, radar point cloud data after preprocessing still exhibits unstructured characteristics and uneven distribution. For example, in the area surrounding an airport runway, point cloud data generated by trees and buildings tends to be densely and regularly distributed, while point cloud data for bird targets is relatively sparse and irregularly shaped. The above process effectively extracts the local geometric structure of bird targets.

[0048] The InvResMLP block addresses this issue by introducing residual connections, summing the outputs of shallow and deep networks, and optimizing the gradient backpropagation path. Furthermore, it employs a separable MLPs architecture, decomposing the fully connected layers into neighborhood and point-feature multilayer perceptrons. This significantly reduces the number of model parameters and computational complexity, while improving the ability to express multi-scale features. This solves the problem of vanishing gradients and difficulty learning deep features in traditional point cloud feature extraction modules as the number of network layers increases.

[0049] Multi-scale feature fusion is achieved through a hierarchical Transformer encoder.

[0050] This example introduces the Transformer attention mechanism into the field of point cloud processing. In each level of attention mechanism module processing, the model calculates the association between points in the local area of ​​the point cloud data through the attention formula, which is as follows:

[0051]

[0052] The model generates Q, K, and V matrices based on the input features, obtains attention weights through matrix operations, and performs weighted summation on V to focus on key features. For example, when detecting gliding birds, it is possible to focus on extracting features such as wings and body contours. The hierarchical Transformer encoder achieves accurate detection by fusing features at different levels: the low-level network captures details such as feather texture and claw shape (such as peacock feather texture), while the high-level network analyzes overall information such as body contour and flight posture (such as the circling trajectory of an eagle). This micro-to-macro feature fusion mechanism enables the PointNeXt model to effectively identify birds of different sizes in airport environments.

[0053] Traditional point cloud processing algorithms often struggle to effectively capture long-range dependencies within point cloud data, making it particularly challenging for bird detection. This is because birds in flight maintain complex spatial relationships between their body parts, such as the relative position of their wings and body, and the coordinated movement of their head and tail. These relationships are crucial for accurate bird identification. The PointNeXt model employed in this paper innovatively introduces the Transformer's attention mechanism into point cloud processing, successfully addressing this challenge through a multi-level Point Attention module.

[0054] Adaptive threshold NMS algorithm is used to optimize the detection results.

[0055] Calculate the point cloud density of the detection area, expressed as:

[0056]

[0057] Among them, ρ is the point cloud density of the detection area, N is the number of point clouds in the area, and V is the volume of the area.

[0058] The NMS suppression threshold is dynamically adjusted based on the regional point cloud density, expressed as:

[0059] T=T base +β·(ρ-ρ avg )

[0060] Among them, T is the non-maximum suppression threshold after dynamic adjustment, T base is the basic threshold, β is the adjustment coefficient, ρ avg is the average point cloud density.

[0061] When the point cloud density is high, the threshold is automatically lowered to retain more overlapping detection frames. When the point cloud density is low, the threshold is increased to reduce false detections, thereby improving detection accuracy.

[0062] In actual airport operations, bird activity is unevenly distributed. In wetland and grassland habitats surrounding airports, flocks of birds result in high target density in point cloud data and extensive detection frame overlap. In areas like runways, however, bird targets are sparse, resulting in low point cloud density and large detection frame spacing. Traditional NMS algorithms use a fixed threshold to process all detection frames, which presents drawbacks in complex airport environments: missed detections are common in densely populated areas, while false detections are common in sparsely populated areas. This process ensures comprehensive detection in complex environments and improves bird recognition accuracy.

[0063] Based on multi-frame point cloud target monitoring, a trajectory tracking model is proposed. The Transformer cross-frame attention mechanism is used to associate the point cloud features of consecutive frames. Combined with the bird motion prior model, real-time and stable tracking of complex bird flight trajectories is achieved.

[0064] The aforementioned PointNeXt model already implements feature extraction and target detection from preprocessed radar point cloud data. Based on these extracted point cloud features, the Transformer encoder-decoder architecture is employed to transform multi-frame point cloud tracking into a collective prediction problem. In actual bird surveillance scenarios at airports, lidar collects multi-frame point cloud data containing bird targets in real time. This data contains a wealth of environmental information and target features. The processing flow of the trajectory tracking model proposed in this paper is designed precisely to accurately extract and track bird targets from this complex data.

[0065] Based on extracting multi-scale features F of point clouds from adjacent frames t and F t+1 , and the trajectory identification and motion state of the t-1 frame are processed by the Transformer encoder to obtain the encoded cross-frame features, and the query vector is generated by the decoder.

[0066] In an airport environment, the flight of birds is a continuous dynamic process. The point cloud data of two adjacent frames (frame t and frame t+1) contain key information about the motion changes of the bird target. Specifically, the Transformer encoder converts F t and F t+1 Encoded as cross-frame features Enc(F t ) and Enc(F t+1 ) and captures global environmental information. Exemplarily, the global environmental information includes airport runway layout and tree locations.

[0067] The Transformer encoder also processes two types of query vectors simultaneously: static object queries and autoregressive trajectory queries. Static object queries are used to detect newly entered bird targets. For example, when a bird flies from the airport perimeter into the monitoring area, this query vector can promptly capture the appearance of the new target. Autoregressive trajectory queries are responsible for tracking existing bird targets and continuously recording their movement trajectories. The decoder uses a self-attention mechanism to jointly infer all query vectors, exploring the spatial relationships between targets. Through the encoder-decoder attention mechanism, the query vectors capture global feature information, accurately predicting the target's bounding box and category, achieving precise localization and classification of bird targets. A multi-layer perceptron (MLP) is then used to map the decoder's output embedding vector into bounding box coordinates, category labels, and trajectory identifiers. This step enables the model to transform abstract feature information into intuitive results that can be used for practical monitoring and early warning. It can localize, classify, and associate bird targets with their trajectories, providing direct data support for airport bird strike prevention efforts.

[0068] The decoder generates an autoregressive trajectory query vector based on the bird target tracked in the current frame (frame t). This vector is used to track the position changes of existing targets in subsequent frames; at the same time, a static object query vector Q is generated. obj It is used to detect new targets and ensure that birds that have just entered the monitoring area are not missed. obj Expressed as:

[0069] Q obj =MLP INT ()

[0070] During the model training phase, static object query vectors are randomly sampled to generate a set of fixed-dimensional vectors. After model training, they can be generated using a learnable parameter layer (such as a multi-layer perceptron (MLP)) with position embeddings or scene prior features as input.

[0071] For the regression trajectory query vector, if t>1, it is generated based on the motion state of the previous frame trajectory. The motion state includes position and velocity, which is expressed as:

[0072]

[0073] represents the autoregressive trajectory query vector of the t-th frame, (x t-1 ,y t-1 , z t-1 ) is the position at the t-1 frame, v t-1 is the speed at frame t-1, MLP track() is a multi-layer perceptron neural network.

[0074] If t = 1, Initialized by the first detection result, illustratively, the center coordinates of the detection box are used as the initial query vector.

[0075] The trajectory association features and new target features are calculated based on the encoded cross-frame features and the query vector.

[0076] The encoded cross-frame features form the concatenated key K and value V. Specifically, the key K and value V are expressed as:

[0077] K=[Enc(F t );Enc(F t+1 )]

[0078] V=K

[0079] The new target feature is obtained by the static object query vector and the attention weight with the key K to identify the new target that is not associated with the historical trajectory in the t or t+1 frame, for example, the first appearance of the bird. The new target feature is represented as:

[0080]

[0081] Among them, Attn obj is the new target feature, is the attention weight of the static object query vector and key K, d k Query the vector dimension for the corresponding static object.

[0082] The trajectory-related features are obtained by the autoregressive trajectory query vector and the attention weight with the key K, so as to find the features matching the historical trajectory in the t-th frame and the t+1-th frame and update the target position. For example, it is possible to obtain the target position from F t The position in (x t ,y t ) is associated with F t+1 (x t+1 ,y t+1 ). The trajectory association feature is expressed as:

[0083]

[0084] Among them, Attn track is the trajectory association feature, is the attention weight of the autoregressive trajectory query vector and key K, d k Query the vector dimension for the corresponding autoregressive trajectory.

[0085] Cross-frame feature fusion uses the encoded features Ft, Ft+1 of the t-th frame and the t+1-th frame as the key K and value V. For the autoregressive trajectory query vector By calculating cross-frame attention It can capture the position change of the target between two frames and update the target trajectory information, such as accurately tracking the turning and acceleration of birds during flight; for static objects, the query vector Q obj ,By calculating the attention with the features of two frames, the newly appeared bird targets are detected,and the potential bird strike risks are discovered in time.

[0086] Based on the trajectory association features and the new target features, a corresponding trajectory is formed to obtain the first predicted position of the next frame.

[0087] According to the multi-layer perceptron, the trajectory association features are mapped to the bounding box coordinates and category labels of the target in the t+1 frame, which is expressed as:

[0088] (x t+1 ,y t+1 , z t+1 , c t+1 )=MLP bbox (ATTn track )

[0089] Continuous trajectories are formed through adjacent frames, corresponding to the classification of birds.

[0090] If a new target is detected, a new track ID is assigned and its detection result is used as the initial track point, forming the starting point of the new target's track.

[0091] Based on the formed trajectory and combined with the kinematic formula, the predicted position information of the corresponding target in the next frame is obtained.

[0092] The second predicted position of the next frame is obtained through the position information of two adjacent frames using a motion prior model.

[0093] Preferably, a priori model is constructed based on historical data, wherein the historical data includes speed, acceleration, and steering angle, and the priori model is a motion model and kinematic formula including the above parameters.

[0094] For example, based on the positions of two adjacent historical frames, the position of the next frame is predicted by the uniform acceleration model, that is, the second predicted position, which is expressed as (x' t+1 , y' t+1 , z' t+1 ), where x' t+1 =x t +v t ·Δt+0.5a t ·(Δt) 2 , y' t+1 and z' t+1 Similarly, Δt is the frame interval time, which is a tPrevious frame acceleration.

[0095] The final predicted position is obtained by weighted fusion of the first predicted position and the second predicted position.

[0096] Specifically, it is expressed as:

[0097] x final =λx t+1 +(1-λ)x′ t+1

[0098] y final =λy t+1 +(1-λ)y′ t+1

[0099] z final =λz t+1 +(1-λ)z′ t+1

[0100] Among them, (x final ,y final , z final ) is the final predicted position, and a predicted tracking trajectory is formed based on the final predicted position.

[0101] When the detection confidence is high (such as c t+1 >0.9), λ=0.8 (trusting the measured data first); when the confidence is low or the trajectory fluctuates greatly, λ=0.5 (relying on the prior smooth trajectory).

[0102] Finally, taking into account the complexity of bird flight trajectories, such as sudden turns, rapid ascent and descent, etc., the present invention also integrates a motion prior model on the basis of the trajectory tracking model to improve trajectory stability, specifically including the following two aspects: 1. Based on the analysis of a large amount of bird flight data, a motion model and kinematic formula (which can be approximated as a uniform acceleration kinematic formula) containing parameters such as speed, acceleration, and steering angle are established; 2. The prediction results of the motion prior model are weightedly combined with the output of the Transformer model. Through the above technical solution, the method proposed by the present invention based on the combination of cross-frame features and query vectors realizes efficient tracking of bird targets in the airport environment, provides strong technical support for airport bird strike prevention, effectively reduces the risk of bird strike incidents, and ensures aviation flight safety.

[0103] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An airport bird detection radar monitoring method, characterized in that: include: Preprocess the acquired point cloud data; Based on the pre-processed point cloud data, the PointNeXt model is used to extract multi-scale features of the point cloud and identify birds. Based on the multi-scale features of multi-frame point clouds, a trajectory tracking model is constructed. The Transformer cross-frame attention mechanism is used to associate the point cloud features of adjacent frames. Combined with the bird motion prior model, the model predicts and matches the bird's flight trajectory. The following steps are performed: the multi-scale features of the point clouds of adjacent frames are encoded by the Transformer to form cross-frame features, and the decoder is used to form a query vector. An autoregressive trajectory query vector is generated based on the recorded bird targets, and a static object query vector is generated to detect new bird targets. The trajectory association features and new target features are calculated based on the cross-frame features and the query vector. The corresponding trajectory is formed according to the trajectory association features and the new target features, and the first predicted position of the next frame is obtained; the second predicted position of the next frame is obtained through the motion prior model based on the position information of the two adjacent frames; and the final predicted tracking trajectory is obtained by weighted fusion of the first predicted position and the second predicted position.

2. The airport bird detection radar monitoring method according to claim 1, characterized in that: The preprocessing of point cloud data includes removing outliers through voxel filtering, eliminating background clutter through statistical filtering, and optimizing the lidar point cloud data into an ordered voxel grid or feature vector adapted to the deep learning model through data enhancement and format conversion.

3. The airport bird detection radar monitoring method according to claim 1, characterized in that: The method extracts multi-scale features of the point cloud based on the pre-processed point cloud data through the PointNeXt model and identifies birds, and further includes optimizing the detection results through the adaptive threshold NMS algorithm and calculating the point cloud density of the detection area, which is expressed as: Among them, ρ is the point cloud density of the detection area, N is the number of point clouds in the area, and V is the volume of the area; The NMS suppression threshold is dynamically adjusted based on the regional point cloud density, expressed as: T=T base +β·(ρ-ρ avg ) Among them, T is the non-maximum suppression threshold after dynamic adjustment, T base is the basic threshold, β is the adjustment coefficient, ρ avg is the average point cloud density; When the point cloud density is high, the threshold is automatically lowered to retain more overlapping detection frames. When the point cloud density is low, the threshold is increased to reduce false detections.

4. The airport bird detection radar monitoring method according to claim 1, characterized in that: The trajectory association feature and the new target feature are calculated based on the cross-frame feature and the query vector, including the trajectory association feature being expressed as: Among them, Attn track is the trajectory association feature, is the attention weight of the autoregressive trajectory query vector and key K, d k is the dimension of the corresponding autoregressive trajectory query vector, is the autoregressive trajectory query vector, K is the key, and V is the value; The new target feature is expressed as: Among them, Attn obj is the new target feature, is the attention weight of the static object query vector and key K, d k is the dimension of the corresponding static object query vector, Q obj It is a static object query vector, K is the key and V is the value.

5. The airport bird detection radar monitoring method according to claim 1, characterized in that: The step of forming a trajectory based on the trajectory association features to obtain the first predicted position of the next frame includes mapping the trajectory association features into the bounding box coordinates and category label of the target in the t+1 frame based on a multi-layer perceptron, which is expressed as: (x t+1 ,y t+1 ,z t+1 ,c t+1 )=MLP bbox (ATTn track ) Among them, (x t+1 ,y t+1 , z t+1 ) is the bounding box coordinate, c t+1 For the classification of birds; continuous tracks are formed through adjacent frames.

6. The airport bird detection radar monitoring method according to claim 1, characterized in that: The method of obtaining a second predicted position based on the position information of two adjacent frames through a motion prior model includes: predicting the second predicted position through a uniform acceleration model based on the positions of two historical adjacent frames, wherein the second predicted position is expressed as (x' t+1 , y' t+1 , z' t+1 ), where x' t+1 =x t +v xt ·Δt+0.5a xt ·(Δt) 2 , y' t+1 =y t +v yt ·Δt+0.5a yt ·(Δt) 2 , z' t+1 =z t +v zt ·Δt+0.5a zt ·(Δt) 2 ; Where Δt is the frame interval time, v xt is the velocity in the x direction, a xt is the acceleration in the x direction, v yt is the velocity in the y direction, a yt is the acceleration in the y direction, v zt is the velocity in the z direction, a zt is the acceleration in the z direction.

7. The airport bird detection radar monitoring method according to claim 2, characterized in that: The voxelization filter divides the three-dimensional space into a uniform voxel grid. Assume that the resolution of the voxel grid in the x, y, and z directions is v x 、v y 、v z , for any point P in the point cloud data i =(x i 、y i 、z i ), the index of the voxel to which it belongs (u i 、v i 、w i ) is calculated using the following formula: Among them, (x i 、y i 、z i ) is the coordinate of any point Pi on the x, y, and z axes, (x min 、y min 、z min ) is the minimum coordinate value of the point cloud data in the x, y, and z axis directions.

8. The airport bird detection radar monitoring method according to claim 2, characterized in that: The statistical filtering removes background clutter and calculates each point p i The Euclidean distance d to its k neighborhood points i , Assuming that the distance from a point to its neighboring points follows a Gaussian distribution, the mean μ and standard deviation σ are calculated, and the distance threshold T = μ + ασ is set, where α is 2-3. Points exceeding the threshold are judged as background clutter and removed.

9. The airport bird detection radar monitoring method according to claim 2, characterized in that: The PointNeXt model includes a basic feature extraction unit, through To achieve the extraction of local geometric features, where R is the maximum pooling operation, j:(i,j)∈N is the neighborhood of point i, h θ is a multi-layer perceptron, is the original feature, is the relative coordinate information of the point cloud, is the position coordinate of point j in the lth layer, is the position coordinate of point i in the lth layer, r l is the normalization factor, is the original eigenvector of point i in the l+1th layer.

10. The airport bird detection radar monitoring method according to claim 2, characterized in that: The final predicted tracking trajectory is obtained by weight fusion of the first predicted position and the second predicted position, which is expressed as: final =λx t+1 +(1-λ)x′ t+1 ,y final =λy t+1 +(1-λ)y′ t+1 , z final =λz t+1 +(1-λ)z′ t+1 , Among them, (x final ,y final , z final ) is the final predicted position, x t+1 is the x-coordinate of the first predicted position, x′t +1 is the x coordinate of the second predicted position, y t+1 is the y coordinate of the first predicted position, y′t +1 is the y coordinate of the second predicted position, z t+1 is the z coordinate of the first predicted position, z′t +1 is the z coordinate of the second predicted position, and λ is the weight; A predicted tracking trajectory is formed according to the final predicted position.

Images