Multi-scale unmanned aerial vehicle trajectory prediction method, device, equipment and medium

Abstract

This invention discloses a multi-scale UAV trajectory prediction method, apparatus, device, and medium. The method includes: acquiring multi-sensor fused trajectory data, including the UAV's position, velocity, acceleration, covariance matrix, and motion model; for confirmed trajectory data, extracting and encoding physical state, sensor quality, and motion pattern features to generate a trajectory embedding representation containing physical constraint priors; performing cross-track attention processing on the embedding representations of multiple trajectories to generate a swarm perception enhanced representation; combining the covariance matrix with physical feature value attention processing to generate a refined temporal representation; parallel decoding based on the refined temporal representation to generate trajectory prediction results at multiple time scales; and performing physical constraint post-processing on the prediction results to output a multi-scale predicted trajectory that satisfies kinematic constraints. This invention achieves high-precision, low-latency multi-scale UAV trajectory prediction.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to a multi-scale UAV trajectory prediction method, apparatus, device, and medium. Background Technology

[0002] With the gradual opening of low-altitude airspace and the rapid popularization of commercial applications of drones, cities and sensitive areas face the challenge of controlling large-scale, multi-type drones flying simultaneously. The threat of intrusion by non-cooperative targets (i.e., drones without ADS-B response) is becoming increasingly prominent, and traditional air traffic control models based on flight plans are difficult to adapt to the dynamic control needs of low, slow, and small targets. Accurately predicting the future flight trajectory of drones is the core technological foundation for achieving early warning, conflict avoidance, and intent recognition. Therefore, how to achieve high-precision, multi-scale prediction of drone trajectories in complex airspace environments has become a pressing technical problem to be solved in this field.

[0003] In existing technologies, early UAV trajectory prediction primarily relied on multi-sensor fusion and statistical models. Specifically, by fusing data from various heterogeneous sensors such as radar, ADS-B, RemoteID, and TDOA, stable tracking of multiple targets was achieved using Kalman filters (supporting various motion models such as CV, CA, CT, and IMM) and data association algorithms (such as GNN and JPDA). Building upon this, Kalman extrapolation based on assumptions of uniform or uniformly accelerated motion became the most commonly used real-time trajectory prediction method. However, this type of method exhibits significant prediction errors in UAV maneuvering flight scenarios, and these errors increase dramatically with the extension of the prediction time domain, making it difficult to meet the prediction accuracy requirements of highly dynamic scenarios.

[0004] In recent years, data-driven trajectory prediction methods have gradually become the mainstream research approach. Recurrent neural networks (such as LSTM and GRU) have been introduced into the field of trajectory prediction, capable of learning the motion patterns of targets from historical trajectory data, but they suffer from error accumulation in long-term predictions. Subsequently, the Transformer architecture based on self-attention mechanisms has gradually become an important direction in trajectory prediction research due to its ability to model long sequences. However, existing Transformer-based single-target trajectory prediction methods typically perform future position regression independently on a single flight path, failing to fully utilize the group interaction information between multiple targets in the same airspace, and lacking systematic adaptation for UAV airspace management scenarios.

[0005] To address the aforementioned issues, some research has begun exploring technical approaches that integrate physical dynamics constraints with deep learning. For example, kinematic constraints (velocity and acceleration boundaries) are explicitly embedded into neural networks to improve the physical feasibility of prediction results. However, existing fusion schemes still have the following shortcomings: First, they lack adaptability to large-scale concurrent multi-target scenarios, making it difficult to meet real-time requirements while ensuring prediction accuracy. Second, they fail to effectively integrate prediction needs across multiple time scales; collision warning, path prediction, and intent recognition often require multiple independent systems to work together, resulting in timing inconsistencies and high engineering complexity. Third, existing schemes do not adequately consider engineering factors such as sensor quality and trajectory status, leading to poor prediction stability when sensor data quality fluctuates or target state changes. Summary of the Invention

[0006] This invention provides a multi-scale UAV trajectory prediction method, apparatus, device, and medium, aiming to solve the technical problem in the prior art that UAV trajectory prediction is difficult to balance the needs of multi-target interactive perception and multi-scale prediction while ensuring physical feasibility.

[0007] In a first aspect, embodiments of the present invention provide a multi-scale unmanned aerial vehicle (UAV) trajectory prediction method, comprising: Acquire multi-sensor fusion trajectory data, which includes the UAV's position, velocity, acceleration, covariance matrix, and motion model at the current and historical times. The trajectory data carries trajectory status identifiers, which include confirmation status, trial status, and smoothing status. For the confirmed state track, based on the track data, physical state features, sensor quality features and motion pattern features are extracted and feature encoding is performed to generate a track embedding representation that includes physical constraint priors. The embedded representations of multiple tracks are subjected to cross-track attention processing to generate a group perception-enhanced representation; Based on the group perception enhancement representation and combined with the covariance matrix, physical feature value attention processing is performed to generate a refined temporal representation; Based on the refined temporal representation, parallel decoding generates multiple trajectory prediction results at different time scales; The trajectory prediction results at multiple different time scales are subjected to physical constraint post-processing to output multi-scale predicted trajectories that satisfy kinematic constraints.

[0008] Secondly, embodiments of the present invention also provide a multi-scale unmanned aerial vehicle (UAV) trajectory prediction device, comprising: The data acquisition module is used to acquire multi-sensor fusion trajectory data. The trajectory data includes the position, velocity, acceleration, covariance matrix and motion model of the UAV at the current time and historical time. The trajectory data carries trajectory status identifiers, which include confirmation status, trial status and smoothing status. The feature extraction and embedding module is used to extract physical state features, sensor quality features and motion pattern features based on the track data for the confirmed state track, and perform feature encoding to generate a track embedding representation that includes physical constraint priors. A cross-track attention module is used to perform cross-track attention processing on the track embedding representations of multiple tracks to generate a group perception enhanced representation; The physical feature value attention module is used to perform physical feature value attention processing based on the group perception enhancement representation and combined with the covariance matrix to generate a refined temporal representation; A multi-scale prediction generation module is used to decode and generate multiple trajectory prediction results at different time scales in parallel based on the refined temporal representation. The physical constraint post-processing module is used to perform physical constraint post-processing on the trajectory prediction results of the multiple different time scales, and output multi-scale predicted trajectories that satisfy kinematic constraints.

[0009] Thirdly, embodiments of the present invention also provide a computer device, which includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described multi-scale UAV trajectory prediction method.

[0010] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the above-described multi-scale UAV trajectory prediction method.

[0011] This invention encodes kinematic constraints into physical priors and embeds them into feature representations. It then combines a cross-track attention mechanism to extract group interaction features across multiple tracks and uses physical feature value attention to refine high-uncertainty directions based on the covariance matrix. Based on this, it employs parallel decoding to output track prediction results at multiple time scales at once. Furthermore, it corrects the kinematic feasibility of prediction points through physical constraint post-processing. This allows a single model to support multi-scale prediction needs while ensuring the physical feasibility of the prediction results. The invention also improves prediction accuracy and computational efficiency through cross-track information sharing and uncertainty focusing mechanisms. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the system architecture provided for an embodiment of the present invention; Figure 2 This is a schematic diagram of the data preprocessing workflow provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the physical sensing feature encoding and prediction architecture provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the encoder architecture and dual attention mechanism provided in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the complete encoder-decoder process and physically constrained attention provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of the multi-scale rolling prediction execution process provided in an embodiment of the present invention; Figure 7A A schematic diagram of the encoder stage process for multi-scale trajectory prediction provided in an embodiment of the present invention; Figure 7B This is a schematic diagram of the decoder stage process for multi-scale trajectory prediction provided in an embodiment of the present invention; Figure 8A This is a schematic diagram of the spatial grouping process of the cross-track attention mechanism provided in an embodiment of the present invention; Figure 8B This is a schematic diagram illustrating the process of calculating attention weights for the cross-track attention mechanism provided in an embodiment of the present invention; Figure 9 This is a schematic diagram of the physical feature value attention process provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the complete end-to-end process provided in an embodiment of the present invention; Figure 11 A schematic diagram of cloud infrastructure provided for an embodiment of the present invention; Figure 12 This is a schematic diagram of system deployment provided for an embodiment of the present invention. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0016] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0017] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0018] As used in this specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if [described condition or event] is detected" may be interpreted, depending on the context, as "once determined," "in response to determination," "once [described condition or event] is detected," or "in response to detection of [described condition or event]."

[0019] This invention provides a multi-scale UAV trajectory prediction method, including steps S1-S6: S1. Acquire multi-sensor fusion trajectory data, the trajectory data including the UAV's position, velocity, acceleration, covariance matrix and motion model at the current time and historical time, the trajectory data carrying trajectory status identifiers, the trajectory status identifiers including confirmed status, trial status and smoothing status.

[0020] In this step, such as Figure 1As shown, the multi-sensor network 102 includes five types of sensor nodes: ADS-B, Radar, RemoteID, 5G-A, and TDOA. Deployed within the monitored airspace, it is responsible for collecting raw measurement data from the UAV, including position, velocity, and signal characteristics. The real-time fusion engine 103 is connected to the multi-sensor network 102, receiving asynchronous observation data from multiple sources. It performs multi-target tracking fusion using a Kalman filter and data association algorithms, outputting a standardized fused trajectory stream. The Kalman filter supports four motion models: CV, CA, CT, and IMM. The data association algorithms include GNN, JPDA, and AI association methods. Each fused trajectory carries position, velocity, covariance matrix, sensor source identifier, identity confidence level, and trajectory age information in the ENU coordinate system. The message bus 104 adopts a Kafka architecture, with the theme "processed-drone-points," decoupling the output of the real-time fusion engine 103 and pushing it to the multi-level trajectory prediction engine 105, supporting high-throughput streaming processing of over 1000 concurrent trajectories. By fusing multi-source heterogeneous sensors and using the Kafka message bus, the integrity and real-time performance of the flight path data were ensured, providing a reliable data foundation for subsequent high-concurrency prediction.

[0021] like Figure 6 As shown, the end-to-end workflow of this system is triggered by a 1Hz timer 671, and in step 672, the track input 670 of the current time window is read. The track input includes: encoder input composed of the fused states of the most recent 10 time windows of the currently active track (after coordinate difference and normalization processing, shape [N_tracks,10,6]), physical feature auxiliary input formed by splicing physical state feature vectors and sensor quality embeddings (shape [N_tracks,48]), and target time lists of L1 (future 1 to 15 seconds), L2 (future 20 to 60 seconds), and L3 (future 75 to 300 seconds). Through the 1Hz time window mechanism of the strictly aligned fusion engine, time-consistent rolling prediction is achieved.

[0022] like Figure 10 As shown, step 1000 is the trajectory state feature determination step. This step determines the required features from the track stream output by the fusion engine 103, including: physical state features with ENU position difference and velocity as the core, sensor quality features with SNR, GDOP and covariance trace as the core, motion pattern features with CV / CA / CT / IMM model type, maneuver intensity and turning rate as the core, and time series features such as history length, update frequency and number of missed detections; each processing window involves more than 1,000 concurrent active tracks.

[0023] S2. For the confirmed state track, based on the track data, extract physical state features, sensor quality features and motion pattern features, and perform feature encoding to generate a track embedding representation that includes physical constraint priors.

[0024] In one specific embodiment, the step of extracting physical state features, sensor quality features, and motion pattern features based on the track data for the confirmed state track, and performing feature encoding to generate a track embedding representation containing prior physical constraints, includes: S21. Convert the absolute coordinates in the trajectory data into a relative displacement sequence, and combine it with the velocity sequence to construct an input feature vector.

[0025] In this step, such as Figure 2 As shown, the raw track status input includes three types of information: continuous numerical status, sensor quality indicators, and track metadata. Continuous numerical status includes position (x, y, z) and velocity (v) in the ENU coordinate system. x v y v z ), acceleration (a x a y The data includes: turning rate (ω) and covariance matrix P_xy; sensor quality indicators include SNR (dB), GDOP, RCS (dBsm) and sensor type; track metadata includes track ID, track status (TENTATIVE / CONFIRMED / COASTING), identity confidence, track age and currently active motion model; the time history is the state sequence of the most recent 10 time windows.

[0026] The coordinate standardization and differential coding step 222 converts the absolute ENU coordinates into a relative displacement sequence, which serves as the core input to the model. The displacement difference calculation formula is as follows: .

[0027] Where H=10 is the length of the historical sequence, and the duration corresponds to 10 time windows. The final input vector is 6-dimensional, represented as follows: .

[0028] The displacement components were normalized using MinMax with a range of ±150m; the velocity components were normalized using Z-Score, based on the global statistic σ_v≈5m / s. By encoding and normalizing relative displacement, the static offset of absolute coordinates was eliminated, allowing the model to focus more on changes in motion patterns and improving generalization ability.

[0029] S22. Encode the sensor type, signal-to-noise ratio, geometric precision factor, and covariance trace to generate a sensor quality embedding.

[0030] In this step, sensor quality feature encoding step 223 converts category features such as sensor type into numerical vectors through learning embeddings, preserving the hierarchical relationship of accuracy between sensors (e.g., semantic information that ADS-B accuracy is better than TDOA). The total dimension of the quality embedding is 32 dimensions, specifically composed of: 16-dimensional one-hot embedding of sensor type, 4-dimensional SNR segmented embedding (<10dB / 10-20dB / >20dB), 4-dimensional GDOP segmented embedding (<5 / 5-15 / >15), 4-dimensional covariance trace log encoding, and 4-dimensional identity confidence segmented embedding. Sensor quality embedding enables the model to perceive the reliability of different sensors, automatically reducing the dependence on corresponding features when there is low-quality data, thus improving the robustness of prediction.

[0031] S23. Encode the motion model type, maneuver intensity, turning rate and new Mahalanobis distance to generate a physical state feature vector.

[0032] In this step, the physical state feature extraction step 224 extracts physical motion features from the Kalman state, including: motion model type (CV=0, CA=1, CT=2, IMM=3) using One-Hot encoding; a 1-dimensional scalar of maneuver intensity as a scalar feature for maneuver detection; a 1-dimensional absolute value of the turning rate, reflecting the severity of heading changes; a 1-dimensional scalar of the innovative Mahalanobis distance, reflecting the reliability of the current state estimate; and three 3-dimensional eigenvalues ​​of the position covariance matrix P_{3×3}, reflecting the main direction of uncertainty, totaling 16-dimensional physical feature vectors. These physical features provide the model with rich dynamic priors, enabling the deep learning network to better understand the motion nature of the UAV.

[0033] S24. Encode the preset kinematic constraint parameters into constraint embeddings.

[0034] In this step, the upper limit for velocity is set at 30 m / s², and the upper limit for acceleration is set at 5 m / s². 2 Physical constraint parameters, such as the upper limit of angular velocity (45° / s), are encoded into learnable soft constraint embeddings, enabling the neural network to learn dynamic characteristics within the physically feasible domain. Through constraint embedding, the model is always guided by the kinematic boundaries during training and inference, fundamentally avoiding the problem of physically infeasible predictions arising from purely data-driven methods.

[0035] S25. The input feature vector, sensor quality embedding, physical state feature vector and constraint embedding are fused and encoded to generate the track embedding representation.

[0036] In this step, such as Figure 3As shown, the architecture is based on a physical constraint Transformer encoder-decoder. The input layer receives the historical state sequence 330-1 (shape [B,10,6]), physical constraints and mass features 331-1 (a 48-dimensional auxiliary vector formed by concatenating the 32-dimensional sensor mass embedding and the 16-dimensional physical state features), and decoder timestamps 330-2 and 331-2. After passing through a linear projection layer and a learnable position embedding, a physical perception embedding 332 with a dimension of d_model=256 is generated, which serves as the input to the Transformer. This embedding simultaneously encodes the historical trajectory, sensor confidence, physical state, and kinematic constraints, providing a rich feature representation for the subsequent attention mechanism and significantly improving the ability to model complex maneuver patterns.

[0037] like Figure 4 As shown, encoder input 440 receives four types of information: historical state sequence (shape [batch, H=10, 6]), physical quality features (shape [batch, 48]), decoder timestamp query (shape [batch, T_decode, d_model]), and physical constraint parameter vector.

[0038] S3. Perform cross-track attention processing on the track embedding representations of multiple tracks to generate a group perception enhanced representation.

[0039] In one specific embodiment, the step of performing cross-track attention processing on the track embedding representations of multiple tracks to generate a group-aware enhanced representation includes: S31. Based on the spatial location of each track, multiple tracks are divided into multiple spatial groups.

[0040] In this step, such as Figure 8A As shown, the local spatial grouping sub-process is 2.0km 2 For each grid cell (approximately 1414m on each side), the grid coordinates of all active tracks in the current window are calculated based on the latest ENU position, and each track is assigned to its corresponding grid cell. Grids with at least two tracks within each cell are filtered out. For grids with more than 64 tracks per cell, they are sorted by distance from the grid center and truncated to 64 tracks. The output is a list of groups organized by spatial proximity, with a global time complexity of O(N·k). The local spatial grouping strategy reduces the cross-track attention complexity from the global O(N·k) to O(N·k). 2 The value is reduced to a local O(k) 2 In typical scenarios with more than 1,000 concurrent flight paths, the computational efficiency is improved by more than two orders of magnitude.

[0041] like Figure 4As shown, the first stage is cross-track attention 441, which aims to capture the correlation dynamics between different tracks within the same airspace, including patterns such as formation, pursuit, and avoidance. The reshaping step 444 reshapes the input data from [N_track,H,d_model] to [1,N_track,H×d_model], making the track quantity dimension the sequence dimension, thereby calculating attention between tracks.

[0042] S32. Within each spatial group, calculate the attention weights between each track and other tracks to obtain the cross-track weight matrix.

[0043] In this step, such as Figure 8B As shown, the intra-group cross-track attention calculation is as follows: The feature tensors of the k tracks (≤64) within the group are reshaped from [k,H,d_model] to [1,k,H×d_model]; the query matrix Q, key matrix K, and value matrix V are obtained through linear mapping; the pairwise distance matrices of the latest positions of all tracks within the group are calculated, and the distance prior bias is constructed. .

[0044] Give higher initial attention weights to track pairs that are spatially closer; calculate the attention score as follows: .

[0045] The cross-track weight matrix (shape [1,k,k]) is obtained after Softmax normalization. The higher value of the element attn[i,j] in the cross-track weight matrix indicates that the historical state of track j is assigned a higher weight when predicting track i. The physical meaning corresponds to the leader-follower relationship in formation flying and the mutual perception relationship in the approach collision avoidance scenario.

[0046] like Figure 4 As shown, multi-head attention step 445 performs multi-head self-attention with nhead=8. The attention scores are initialized using distance prior bias, and track pairs that are spatially closer are assigned higher initial weights. Output steps 446 and 442 generate a cross-track weight matrix h_cross∈R^{N×N}, which reflects the degree of influence of each track on the prediction of the others.

[0047] S33. Based on the cross-track weight matrix, the track embedding representations of each track are weighted and fused to generate the group perception enhancement representation.

[0048] In this step, after weighted aggregation of the value matrix V, the tensor shape is restored to [k, H, d_model], outputting a swarm-aware enhanced representation and a cross-track association weight matrix. The output dimension is [N_tracks, H, d_model], and the historical representation of each track is fed into subsequent modules after swarm-aware enhancement. Compared to independent single-track prediction, the cross-track attention mechanism enables the model to capture swarm dynamics patterns such as formation flying and cooperative avoidance, effectively reducing the average prediction error of member tracks in formation scenarios.

[0049] like Figure 4 As shown, steps 446 and 442 generate a cross-track weight matrix h_cross∈R^{N×N}, reflecting the degree of influence of each track on the prediction of the others, and multiply it with the feature vectors of each track to complete the group perception enhancement. To meet the efficiency requirements of more than 1000 concurrent tracks, the cross-track attention adopts a local spatial grouping strategy, dividing the monitoring area into areas covering approximately 2km. 2 In a spatial lattice, attention is calculated only between paths within the same lattice, with a typical number of paths per lattice not exceeding 64, resulting in a global O(N) time complexity. 2 The complexity is reduced to local O(k) 2 (k) <N)。

[0050] S4. Based on the group perception enhancement representation and combined with the covariance matrix, perform physical feature value attention processing to generate a refined temporal representation.

[0051] In one specific embodiment, the step of performing physical feature value attention processing based on the swarm-aware enhanced representation and combined with the covariance matrix to generate a refined temporal representation includes: S41. Extract the position covariance submatrix from the covariance matrix.

[0052] In this step, such as Figure 9 As shown, the input preparation step extracts the position covariance submatrix P_pos = P[…;3,…;3] (shape [B,3,3]), which describes the spatial distribution of uncertainty in target position estimation. The position covariance submatrix reflects the uncertainty of UAV position estimation in various directions, providing a physical basis for attention focusing.

[0053] S42. Perform eigenvalue decomposition on the position covariance submatrix to obtain eigenvalues ​​and corresponding eigenvectors.

[0054] In this step, SVD is used to ensure numerical stability. P_pos is decomposed using Ppos = V / Lambda V^T to obtain eigenvalues ​​(shape [B,3]) and eigenvector matrices (shape [B,3,3]) arranged in ascending order. Eigenvalue decomposition separates uncertainty into primary and secondary directions, enabling the model to identify the dimension with the highest prediction difficulty.

[0055] S43. Dynamically determine the main direction based on the eigenvalues, and construct a low-dimensional projection space based on the eigenvectors corresponding to the main direction.

[0056] In this step, the dynamic threshold is calculated as: τ = λ_max - median(λ), retaining the principal directions that satisfy λ_i > τ, with the number of principal directions being p = max(count,1), ensuring that at least one direction is retained; the principal feature vector matrix (taking the last p columns) J ​​= eigenvectors[…;…;p:] (shape [B,3,p]) is extracted and expanded to the feature space through a learnable linear transformation to obtain J_feat = W_expand(J) (shape [B,d=64,p]). Typical p values ​​are approximately 8 to 16. By dynamically determining the principal directions, the model automatically focuses on the dimension with the highest uncertainty in state estimation. When the target's height along a certain direction is uncertain (such as a smooth state track after crossing an occluded area), the corresponding eigenvalue is large, and attention automatically focuses on the variation pattern of that direction in the historical sequence.

[0057] like Figure 5 As shown, the detailed sub-steps of the physical eigenvalue attention 562 are as follows: First, calculate the query Q, K, and value V (all with shapes [B, H, 64]); second, calculate the covariance matrix from the feature space where K is located and extract the eigenvector matrix J ∈ R^{64×p}; then, dynamically determine the number of principal directions p to retain, and the calculation rule is as follows: Typical values ​​are approximately 8 to 16; subsequently, scaled dot product attention is applied in the low-dimensional feature space, calculated as follows: Finally, the output is projected back to the original dimension d_model=256, and the shape is restored to [B,H,256], which is used as the refined encoder output for the decoder.

[0058] S44. Project the group perception enhancement representation onto the low-dimensional projection space, perform attention calculation, and map the calculation result back to the original feature space to generate the refined temporal representation.

[0059] In this step, sparse attention is calculated as follows: First, query Q = W_q(x), key K = W_k(x), and value V = W_v(x) (all with shapes [B,H,d]); then, Q and K are projected onto the low-dimensional space along the principal feature direction, and the following calculations are performed: , Shape [B, H, p], where p << d, significantly reducing the computational load; the calculation of sparse attention scores and weights is as follows: ; Weighted aggregation is performed on the original value matrix V and a residual connection is obtained: ; Finally, a refined temporal representation (shape [B, H, 256]) is generated. The physical eigenvalue attention mechanism compresses the attention calculation dimension from d = 64 to p ≤ 4, reducing the computational load by about 20 times; at the same time, the direction of the main eigenvector corresponds to the axis of maximum uncertainty of the target motion, giving the attention mechanism an interpretable physical meaning, and the secondary eigenvalues (corresponding to the measurement noise direction) are discarded, effectively reducing the jitter noise of the predicted trajectory.

[0060] In this step, as Figure 3 shown, the physical weighted scalar product 334 multiplies the cross-track attention output and the physical state features element-wise, and the calculation formula is: h_phys = h_cross ⊙ σ(W_phys·f_phys) where ⊙ represents element-wise multiplication (Hadamard product), σ is the activation function (usually sigmoid, compressing the value to (0, 1) as a gating mechanism), W_phys is the learnable weight matrix, f_phys is the input physical feature vector, h_cross is the feature vector output by the cross-attention module, and h_phys is the output feature after physical gating fusion. Through the physical gating mechanism, the model can dynamically adjust the contribution weight of the cross-track attention according to the current physical state, making the prediction result more in line with the kinematic law.

[0061] In this step, the physical eigenvalue attention 335 uses the eigenvalue decomposition of the trajectory state covariance matrix as the first core optimization of the model, automatically focusing on the dimension with the highest state estimation uncertainty. Specifically, it includes: extracting the eigenvalues λ_i from the Kalman covariance matrix P and arranging them in descending order; selecting the directions corresponding to the top p main eigenvalues to construct a low-dimensional feature space, where p is dynamically determined by max(λ) - median(λ); the attention calculation is performed in this low-dimensional space, naturally concentrating the prediction energy on the high-uncertainty direction. This mechanism enables the model to automatically identify and focus on the state dimension with the highest prediction difficulty, and allocate more computational resources in the uncertainty direction.

[0062] In this step, the masked physical eigenvalue attention 337 combines the physical constraint masking mechanism in the decoding stage as the second core optimization of the model to prevent the prediction result from violating the kinematic constraints. This module receives the refined features from step 335 and the timestamp embedding of the target prediction time domain, ensures the prediction directionality through causal masking, and for violations of v_max = 30m / s or a_max = 5m / s 2A negative infinity masking value is applied to the prediction direction of the constraint, filtering out infeasible predictions in advance at the attention level. By masking the constraints at the attention layer, the generation of physically infeasible predictions is prevented from the source, further enhancing the kinematic feasibility of the prediction results.

[0063] like Figure 4 As shown, the second stage is the physical probability-based attention 443, which aims to filter key motion information by combining physical priors. This stage adopts a dual-input fusion structure: in path A, the enhanced representation 447 from cross-track attention is processed by a 1D convolutional network 449 with 3 kernels to extract local temporal patterns, outputting a shape [batch, H, d_model / 2]; in path B, the physical quality features 448 are expanded into temporal features by a fully connected network 450, outputting a shape [batch, H, 2]. The physical probability attention 451 concatenates the two features and applies sparse attention based on covariance feature values, utilizing state uncertainty to focus on important dynamic dimensions. The output 452 is a physically calibrated refined temporal representation with a shape [batch, H, d_model], which is passed to the decoder.

[0064] like Figure 5 As shown, the encoder path includes the following processing steps on the left. Input 555 has a shape of [B, H=10, d_model=256], where B is the number of concurrent tracks within the batch. Reshaping step 556 converts the input to [1, B, H × d_model] to prepare for cross-track attention computation. Cross-track multi-head attention 557 applies local spatial grouping cross-track attention, outputting a group association weight matrix of shape [1, B, H]. Transpose step 558 transposes the matrix back to [B, H, 1] and broadcasts it to [B, H, d_model] for weighted fusion with the original features. Feedforward network and residual step 559 uses a two-layer feedforward network (FFN) with d_ff=1024, combined with residual connections and LayerNorm, replacing BatchNorm to adapt to variable-length batches and online inference scenarios. Downsampling step 561, combined with residual connections, compresses the feature dimension to [B, H, 64] to extract core dynamic features and reduce subsequent computation.

[0065] S5. Based on the refined temporal representation, generate multiple trajectory prediction results at different time scales through parallel decoding.

[0066] In one specific embodiment, the step of generating multiple trajectory prediction results at different time scales through parallel decoding based on the refined temporal representation includes: S51, Generate corresponding UAV time stamp embeddings for the first prediction scale, the second prediction scale and the third prediction scale respectively.

[0067] In this step, such as Figure 3 and Figure 5 As shown, target time-stamp embeddings are generated for L1 short-term (0 to 15 seconds), L2 medium-term (15 to 60 seconds), and L3 long-term (60 to 300 seconds), respectively. These embeddings encode the time difference Δt between the target time and the current time, informing the decoder "how far to predict." The L1 prediction range is 0-15s with a 1s sampling interval, outputting 15 prediction points; the L2 prediction range is 15-60s with a 5s sampling interval, outputting 9 prediction points; and the L3 prediction range is 60-300s with a 15s sampling interval, outputting 16 prediction points. This multi-scale timestamp embedding allows the decoder to perceive the time-scale differences across different prediction ranges, thus adaptively adjusting its prediction behavior.

[0068] S52. Using the refined time sequence representation as the key and value, and embedding the UAV time timestamps of the first prediction scale, the second prediction scale, and the third prediction scale as the query, respectively, the trajectory prediction results of the corresponding scale are generated in parallel through the cross-attention mechanism.

[0069] The trajectory prediction results at the first prediction scale include short-term position and velocity predictions, the trajectory prediction results at the second prediction scale include medium-term position and velocity predictions and maneuver intent classifications, and the trajectory prediction results at the third prediction scale include long-term position and velocity predictions and mission intent classifications.

[0070] In this step, three independent prediction heads decode in parallel from the same encoded representation: the L1 short-term prediction head outputs 15 points × 1 second of position, velocity, and 3 × 3 covariance predictions, with shapes [B,15,6] and [B,15,3,3]; the L2 medium-term prediction head outputs 9 points × 5 seconds of trajectory predictions and 5 types of maneuver intent classifications (straight flight, turning, climbing, descending, hovering); the L3 long-term prediction head outputs 16 points × 15 seconds of trajectory predictions and 5 types of mission intent classifications (patrol, reconnaissance, infiltration, retreat, unknown). Parallel multi-scale decoding avoids the error accumulation of stepwise recursion, and a single forward propagation can simultaneously output the prediction results and intent classifications of the three scales, realizing unified modeling of collision warning, path prediction, and intent recognition operations.

[0071] like Figure 3 As shown, the parallel multi-head decoder 338 consists of three independent prediction heads that decode in parallel from the same encoded representation: the L1 short-term prediction head outputs 15 points × 1s of position, velocity, and 3 × 3 covariance predictions in the shapes [B,15,6] and [B,15,3,3]; the L2 medium-term prediction head outputs 9 points × 5s of trajectory prediction and maneuver type intent classification; and the L3 long-term prediction head outputs 16 points × 15s of trajectory prediction and mission intent classification.

[0072] like Figure 5As shown, the right side of the decoder path includes the following processing steps. Three-scale timestamp embedding (564 to 565) generates timestamp embeddings for the target prediction time at L1, L2, and L3, respectively, encoding the time difference Δt between the target time and the current time. Input fusion step 566 uses the encoder output 562 as the key and value, and the three-scale timestamp embedding as the query, fusing historical dynamic information and target temporal information through a cross-attention layer. Feedforward layer 567 further refines the prediction representations at each scale. Three-scale parallel output prediction 568 includes: L1 short-term prediction head output shape [B,15,6], containing position mean and velocity, with added covariance prediction [B,15,3,3]; L2 medium-term prediction head output shape [B,9,6], with added 5 types of maneuver intent classification logits, including straight flight, turning, climbing, descending, and hovering; L3 long-term prediction head output shape [B,16,6], with added 5 types of mission intent classification logits, including patrol, reconnaissance, infiltration, retreat, and unknown.

[0073] like Figure 7A and 7B As shown, the initialization parameters of the algorithm are as follows: historical sequence length H=10, input dimension input_DIM=6, Transformer hidden layer dimension d_enc=256, number of multi-head attention heads nhead=8, number of encoder stacking layers num_layers=3, feedforward network hidden dimension d_ff=1024, Dropout ratio 0.2, L1 prediction steps T_L1=15 (corresponding to a time span of 15 seconds), L2 prediction steps T_L2=9 (corresponding to a time span of 45 seconds), L3 prediction steps T_L3=16 (corresponding to a time span of 240 seconds), maximum drone flight speed limit MAX_SPEED=30.0m / s, maximum linear acceleration constraint MAX_ACCELERATION=5.0m / s. 2 The maximum jerk constraint is MAX_JERK = 10.0 m / s². 3 The maximum yaw rate constraint is MAX_TURN_RATE = 45.0° / s.

[0074] Encoder stage Figure 7A The processing flow for (cross-track attention to physical feature value attention) is as follows: The embedding layer linearly projects the track history sequence and superimposes the learnable position embedding, outputting a shape [B,H,256]; subsequently, it is processed by stacking encoder blocks of num_layers=3 layers, with each layer sequentially performing local spatial grouping (in 2km increments). 2The algorithm consists of a grid cell (no more than 64 tracks per group), multi-head attention across tracks within the group (including distance prior bias), residual fusion, and physical gating weighting: g = σ(W_phys·f_phys), x = x⊙g; dual-path feature extraction (path A outputs 128 dimensions via Conv1D, path B outputs 128 dimensions via fully connected expansion), physical feature value attention is applied after concatenation, and a feedforward network is connected to the residual; the encoder finally outputs shape [B,H,64] after time dimension downsampling (factor 4), and generates key K and value V for the decoder.

[0075] Decoder stage Figure 7B The processing flow (parallel prediction at three scales to post-processing with physical constraints) is as follows: target time stamp embeddings are generated for L1, L2, and L3 respectively, with shapes [B,15,256], [B,9,256], and [B,16,256] respectively; the three prediction heads perform cross-attention fusion and feedforward network processing in parallel; the L1 prediction head outputs position [B,15,3], velocity [B,15,3], and covariance [B,15,3,3], the L2 prediction head outputs position [B,9,3], velocity [B,9,3], and 5 types of maneuver intention logits, and the L3 prediction head outputs position [B,16,3], velocity [B,16,3], and 5 types of task intention logits; the outputs at each scale are processed by physical constraint projection to ensure that the velocity, acceleration, and turning rate satisfy the kinematic constraints.

[0076] The training loss function of the model includes a physical penalty term, and its expression is: Where L_NLL is the negative log-likelihood loss based on prediction covariance, L_CE is the cross-entropy loss for intent classification, and L_phys is the physical violation soft constraint penalty term (activated when speeding or over-acceleration occurs). By using a weighted multi-task training loss, trajectory prediction accuracy, intent recognition accuracy, and physical feasibility are optimized in a unified manner, achieving a synergistic improvement among the three.

[0077] S6. Perform physical constraint post-processing on the trajectory prediction results of the multiple different time scales, and output a multi-scale predicted trajectory that satisfies kinematic constraints.

[0078] In one specific embodiment, the step of performing physical constraint post-processing on the trajectory prediction results at multiple different time scales to output a multi-scale predicted trajectory that satisfies kinematic constraints includes: S61. For each prediction point in the multiple trajectory prediction results at different time scales, check whether its speed, acceleration, and turning rate meet the preset kinematic constraints.

[0079] In this embodiment, the preset kinematic constraints include: a maximum velocity v_max = 30 m / s and a maximum acceleration a_max = 5 m / s². 2 The maximum turning rate ω_max = 45° / s. For each prediction point, the following checks are performed: velocity continuity: |v_{k+1} - v_k| / Δt ≤ a_max, velocity amplitude: ||v|| ≤ v_max, and turning rate: |Δθ / Δt| ≤ ω_max. These three layers of constraint checks ensure that the predicted trajectory remains within physically achievable limits, eliminating infeasible predictions such as overspeeding and sharp turns common in purely data-driven methods.

[0080] like Figure 3 As shown, the physical constraint post-processing step 338-post performs kinematic feasibility projection on the decoded output to ensure that the velocity difference between adjacent predicted points does not exceed a_max·Δt and the turning rate does not exceed ω_max=45° / s. The final output 339 contains the fused trajectory prediction results at three scales and is pushed to the alarm generation module and the track status database 106.

[0081] like Figure 6 As shown, post-physical constraint verification step 673 performs a fast physical feasibility verification on the model output. The verification conditions include the velocity continuity constraint: |v_{k+1} - v_k| / Δt ≤ a_max=5m / s 2 Speed ​​limit constraint: ||v|| ≤ v_max=30m / s; Turning rate constraint: |Δθ / Δt| ≤ ω_max=45° / s.

[0082] S62. If the predicted point does not satisfy the kinematic constraints, then the predicted point is kinematically projected and corrected to satisfy the kinematic constraints.

[0083] In this step, the velocity constraint projection formula is: The position is recalculated based on the corrected velocity. The number of corrections is recorded for model performance monitoring. The L1 covariance propagates along the prediction step, while the L2 and L3 uncertainties increase with time by σ. 2 (t)=σ0 2 +q 2 ·Δt 2 Growth. Through projection correction, the final output predicted trajectory strictly satisfies kinematic feasibility, improving the reliability of alarm decisions. Optimization techniques include LayerNorm, L2 regularization, early stopping (patience=20 rounds), and cosine annealing learning rate scheduling.

[0084] S63. Output the corrected multi-scale predicted track.

[0085] In this step, the prediction results are pushed to the alarm module and the downstream decision-making system; the prediction error is used as an online learning signal for incremental model updates, forming a closed loop of "perception-prediction-feedback". For example... Figure 6 As shown, in steps 675 to 676, the storage and alarm triggering steps write the three-scale prediction results into the track state database 606 and associate them with the corresponding track ID and time window key; the L1 prediction immediately triggers a collision and restricted area intrusion warning assessment (response delay less than 5ms); the L2 and L3 predictions are pushed to the intent analysis module; in step 676, the true fusion state of the current time window is updated to the historical input of the next round of prediction, realizing time-series rolling that is strictly aligned with the time window mechanism of the fusion engine; at the same time, the error between the L1 prediction result of the previous window and the true state of the current window is asynchronously pushed to the training queue to support incremental model updates. The closed-loop feedback mechanism enables the model to continuously adapt to changes in the flight behavior distribution in the current airspace and maintain the stability of prediction performance when the flight mode undergoes regular drift.

[0086] like Figure 10 As shown, step 1008 is the automated response execution step. This step performs one or more automated response operations based on the three-scale prediction results of step 1006: Based on L1 prediction, the real-time warning triggering delay is less than 5ms, performing target entry into the no-fly zone warning, target collision risk assessment, and smooth state trajectory observation gap filling extrapolation; based on L2 prediction, the flight management triggering delay is less than 500ms, performing flight path and no-fly zone boundary conflict prediction, dynamic threat level assessment based on maneuver intent, and multi-target coordinated control strategy generation; based on L3 prediction, the situation assessment is executed asynchronously, including non-cooperative target mission intent inference and alarm reporting, flight mode anomaly detection, and long-term surveillance resource scheduling optimization; online model updates use L1 prediction error as the online learning signal, asynchronously executing incremental parameter updates to continuously adapt to the current airspace flight behavior distribution, forming a real-time closed loop of perception-prediction-response with steps 1000 to 1006.

[0087] In one specific embodiment, the multi-scale UAV trajectory prediction method further includes: A11. For the exploratory state track, only the first prediction scale prediction among the multiple track prediction results generated by the parallel decoding is executed, and it is output as the short-term prediction result.

[0088] In this step, due to insufficient historical information or low confidence levels, the tentative track only performs L1 short-term prediction (0-15 seconds), omitting medium- and long-term predictions to reduce computational load and avoid unreliable long-term forecasts. Figure 2As shown, the batch processing strategy schedules tracks in groups based on their state: tracks in the CONFIRMED state are processed first and L1, L2, and L3 full-scale predictions are performed; tracks in the COASTING state are processed with a one-window delay; and tracks in the TENTATIVE state only undergo L1 short-term predictions. This hierarchical scheduling strategy effectively reduces the overall computational load of the system while ensuring the prediction quality of critical targets.

[0089] A12. For the smoothed state track, extrapolation prediction is performed using a Kalman filter based on a uniform motion model. In this step, a smoothed state (COASTING) track indicates that the target is temporarily lost or sensor data is missing. In this case, it is downgraded to Kalman linear extrapolation based on a CV (uniform velocity) model, and the depth model is only used for the CONFIRMED state track. Figure 6 As shown, when the GPU is unavailable or under excessive load, the COASTING and TENTATIVE state tracks are degraded to Kalman linear extrapolation based on the CV model, while only the CONFIRMED state tracks continue to use the deep model to ensure the prediction quality of critical targets. Through hierarchical processing based on track state, the system can still guarantee the real-time performance and reliability of core services in resource-constrained scenarios.

[0090] like Figure 10 As shown, step 1002 is the physical sensing embedding creation step. This step uses the physical sensing embedding technology component 114 to encode the features output from step 1000, setting the velocity upper limit to 30 m / s² and the acceleration upper limit to 5 m / s². 2 Physical constraint parameters, such as the maximum turning rate of 45° / s, are encoded into learnable constraint embedding vectors; categorical features such as sensor type and motion model are transformed into numerical vectors that retain the hierarchical relationship of accuracy; the eigenvalues ​​of the covariance matrix are encoded into uncertainty embeddings that reflect the confidence of state estimation in each direction; this step transforms the physical domain knowledge in the fusion algorithm into a form that the neural network can understand, providing a strong inductive bias.

[0091] like Figure 10 As shown, step 1004 is the spatiotemporal attention processing step. This step uses the spatiotemporal attention technology component 116 to process the multi-track history embedding sequence, which includes two sub-processes: the cross-track attention sub-process extracts group dynamics patterns such as formation and avoidance, and generates a track representation with enhanced group perception; the physical feature value attention sub-process uses covariance feature values ​​to focus on directions with high uncertainty, filters key temporal dynamic information, and realizes sparse attention guided by physical constraints; this step generates a compact track representation that simultaneously encodes individual motion history and group interaction patterns.

[0092] like Figure 10As shown, step 1006 is the multi-scale trajectory prediction generation step. This step uses the three-scale parallel decoding head of the multi-scale trajectory prediction generator 118 to generate prediction results at once, and ensures kinematic feasibility through physical constraint projection: L1 short-term prediction corresponds to a time range of 0 to 15 seconds, a sampling interval of 1 second, and outputs 15 progressive position and velocity prediction points with covariance; L2 medium-term prediction corresponds to a time range of 15 to 60 seconds, a sampling interval of 5 seconds, and outputs 9 flight path prediction points and 5 types of maneuver intent classifications: straight flight, turning, climbing, descending, and hovering; L3 long-term prediction corresponds to a time range of 60 to 300 seconds, a sampling interval of 15 seconds, and outputs 16 strategic path prediction points and 5 types of mission intent classifications: patrol, reconnaissance, infiltration, retreat, and unknown; the inference delay for a single batch of 1000 trajectories is less than 20ms. In summary, the technical effects of this embodiment of the invention are as follows: (1) Real-time performance: Based on the collaborative optimization of local spatial grouping cross-track attention and physical feature value dimensionality reduction attention, the model inference latency of a single batch of 1000 concurrent tracks is less than 20 ms, which meets the real-time processing frequency requirement of 1 Hz; the L1 warning trigger latency is less than 5 ms, which meets the millisecond-level response requirement of collision warning.

[0093] (2) Prediction accuracy effect: The cross-track attention mechanism effectively reduces the average prediction error of member tracks in formation scenarios compared to independent single track prediction; physical feature value attention reduces jitter noise in the predicted trajectory by discarding secondary feature values ​​(corresponding to the direction of measurement noise).

[0094] (3) Physical feasibility effect: The full-process physical constraint mechanism (embedded coding, attention masking, and kinematic projection post-processing three-layer guarantee) ensures that the output trajectory meets the UAV kinematic constraints in the three dimensions of speed, acceleration, and turning rate, eliminating the problem of physical infeasibility prediction caused by pure data-driven methods and improving the reliability of alarm decision.

[0095] (4) Multi-scale business coverage effect: The single inference synchronously outputs L1 position and covariance prediction, L2 path prediction and 5 types of maneuver intention classification, L3 strategic path prediction and 5 types of task intention classification, covering three types of business such as collision warning, restricted area control and threat assessment with a single model. Compared with the deployment scheme of multiple independent systems, it eliminates the problem of cross-system timing inconsistency and reduces engineering complexity.

[0096] (5) Large-scale concurrent processing effect: The local spatial grouping strategy reduces the cross-track attention complexity from global O(N) to O(N) 2 Reduced to local O(k) 2 The theoretical computational cost is reduced by approximately (N / k). 2 In typical scenarios with more than 1,000 concurrent flight paths within the monitoring area, the computational efficiency is significantly better than that of the global attention scheme.

[0097] (6) Online adaptive effect: The asynchronous incremental training mechanism with L1 prediction error as the online learning signal enables the model to continuously adapt to the changes in flight behavior distribution in the current airspace, forming a closed loop with the entire process of perception-prediction-response, and maintaining the stability of prediction performance when flight patterns drift regularly (such as a surge in drone density during holidays).

[0098] In a specific embodiment, the present invention also provides a multi-scale unmanned aerial vehicle (UAV) trajectory prediction device, comprising: The data acquisition module is used to acquire multi-sensor fusion trajectory data. The trajectory data includes the position, velocity, acceleration, covariance matrix and motion model of the UAV at the current time and historical time. The trajectory data carries trajectory status identifiers, which include confirmation status, trial status and smoothing status. The feature extraction and embedding module is used to extract physical state features, sensor quality features and motion pattern features based on the track data for the confirmed state track, and perform feature encoding to generate a track embedding representation that includes physical constraint priors. A cross-track attention module is used to perform cross-track attention processing on the track embedding representations of multiple tracks to generate a group perception enhanced representation; The physical feature value attention module is used to perform physical feature value attention processing based on the group perception enhancement representation and combined with the covariance matrix to generate a refined temporal representation; A multi-scale prediction generation module is used to decode and generate multiple trajectory prediction results at different time scales in parallel based on the refined temporal representation. The physical constraint post-processing module is used to perform physical constraint post-processing on the trajectory prediction results of the multiple different time scales, and output multi-scale predicted trajectories that satisfy kinematic constraints.

[0099] In one specific embodiment, the cross-track attention module is specifically used for: Based on the spatial location of each track, multiple tracks are divided into multiple spatial groups; Within each spatial group, the attention weights between each track and other tracks are calculated to obtain the cross-track weight matrix; Based on the cross-track weight matrix, the track embedding representations of each track are weighted and fused to generate the group perception enhancement representation.

[0100] In one specific embodiment, the physical feature value attention module is specifically used for: Extract the position covariance submatrix from the covariance matrix; The position covariance submatrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues ​​and corresponding eigenvectors. The principal direction is dynamically determined based on the eigenvalues, and a low-dimensional projection space is constructed based on the eigenvectors corresponding to the principal direction. The group perception enhancement representation is projected onto the low-dimensional projection space, attention is calculated, and the calculation result is mapped back to the original feature space to generate the refined temporal representation.

[0101] In one specific embodiment, the multi-scale prediction generation module is specifically used for: Generate corresponding UAV time timestamp embeddings for the first prediction scale, the second prediction scale, and the third prediction scale, respectively; Using the refined time sequence representation as keys and values, and embedding the UAV time timestamps of the first, second, and third prediction scales as queries, the corresponding scales of trajectory prediction results are generated in parallel through a cross-attention mechanism. The trajectory prediction results at the first prediction scale include short-term position and velocity predictions, the trajectory prediction results at the second prediction scale include medium-term position and velocity predictions and maneuver intent classifications, and the trajectory prediction results at the third prediction scale include long-term position and velocity predictions and mission intent classifications.

[0102] In one specific embodiment, the physical constraint post-processing module is specifically used for: For each prediction point in the multiple trajectory prediction results at different time scales, check whether its speed, acceleration, and turning rate meet the preset kinematic constraints. If the predicted point does not satisfy the kinematic constraints, then the predicted point is kinematically projected and corrected to satisfy the kinematic constraints. Output the corrected multi-scale predicted track.

[0103] In one specific embodiment, the feature extraction and embedding module is specifically used for: The absolute coordinates in the trajectory data are converted into a relative displacement sequence, and combined with the velocity sequence to construct an input feature vector; The sensor type, signal-to-noise ratio, geometric precision factor, and covariance trace are encoded to generate a sensor quality embedding. The motion model type, maneuver intensity, turning rate, and new Mahalanobis distance are encoded to generate a physical state feature vector; Encode the preset kinematic constraint parameters into constraint embeddings; The input feature vector, sensor quality embedding, physical state feature vector, and constraint embedding are fused and encoded to generate the track embedding representation.

[0104] In one specific embodiment, the multi-scale UAV trajectory prediction device further includes: The short-term prediction module is used to perform only the first prediction scale prediction among the multiple trajectory prediction results generated by the parallel decoding at different time scales for the trial state trajectory, and output it as the short-term prediction result. The extrapolation prediction module is used to perform extrapolation prediction based on a uniform motion model using a Kalman filter for the smooth state trajectory.

[0105] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned multi-scale UAV trajectory prediction device and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0106] like Figure 11 As shown, the cloud infrastructure 1100 consists of multiple layers of resources, including physical servers, virtual machines, and containers, supporting the high availability, elasticity, and distributed deployment of the UAV multi-level trajectory prediction information processing system 100. The cloud infrastructure 1100 includes: physical servers 1105, serving as the underlying computing and storage resource pool; virtualization infrastructure 1104, supporting unified scheduling and isolation of virtual machines (VMs) and containers, with typical implementations including KVM, VMware, Docker, and Kubernetes; multiple virtual machine / container clusters 1102-1, 1102-2…1102-L, each hosting different service modules of the system 100, such as data acquisition, fusion engine, trajectory prediction service, database, alarm and visualization, etc.; each service module runs as a microservice in an independent container or virtual machine, supporting elastic scaling and fault isolation; the cloud platform can also integrate acceleration resources such as GPUs / TPUs to meet the needs of deep learning inference. Each virtual machine / container cluster 1102 is uniformly managed through the virtualization infrastructure 1104, and the underlying physical servers 1105 can be horizontally scaled. The services are interconnected through a high-bandwidth, low-latency network, supporting real-time processing of large-scale concurrent drone targets.

[0107] like Figure 12As shown, the processing platform 1200 consists of multiple physical or virtual processing devices 1202-1, 1202-2, 1202-3…1202-K. These devices are interconnected via a network 1204 to collaboratively complete the various functions of the UAV multi-level trajectory prediction system 100. Each processing device 1202 can be a physical server, virtual machine, container node, or edge computing node, internally containing hardware resources such as a processor 1210 (e.g., GPU, CPU, TPU, FPGA), memory 1212 (RAM / ROM / SSD, etc.), and a network interface 1214. Each processing device can deploy different system modules according to actual business needs, such as data acquisition, fusion engine, trajectory prediction, database, alarm and visualization, etc. The system supports horizontal scaling and heterogeneous resource collaboration to adapt to the UAV management needs of different scales and scenarios. The network 1204 can be a local area network, wide area network, Internet, 5G / satellite private network, etc., ensuring high-speed data communication and secure isolation between processing devices. System 100 can be flexibly deployed in a single data center, across multiple data centers, or in an edge-cloud collaborative architecture to achieve high availability, elasticity, and disaster recovery.

[0108] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0109] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Since these modifications and variations fall within the scope of the claims and their equivalents, this invention also intends to include these modifications and variations.

[0110] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-scale UAV trajectory prediction method, characterized in that, include: Acquire multi-sensor fusion trajectory data, which includes the UAV's position, velocity, acceleration, covariance matrix, and motion model at the current and historical times. The trajectory data carries trajectory status identifiers, which include confirmation status, trial status, and smoothing status. For the confirmed state track, based on the track data, physical state features, sensor quality features and motion pattern features are extracted and feature encoding is performed to generate a track embedding representation that includes physical constraint priors. The embedded representations of multiple tracks are subjected to cross-track attention processing to generate a group perception-enhanced representation; Based on the group perception enhancement representation and combined with the covariance matrix, physical feature value attention processing is performed to generate a refined temporal representation; Based on the refined temporal representation, parallel decoding generates multiple trajectory prediction results at different time scales; The trajectory prediction results at multiple different time scales are subjected to physical constraint post-processing to output multi-scale predicted trajectories that satisfy kinematic constraints.

2. The multi-scale UAV trajectory prediction method according to claim 1, characterized in that, The step of performing cross-track attention processing on the track embedding representations of multiple tracks to generate a group-aware enhanced representation includes: Based on the spatial location of each track, multiple tracks are divided into multiple spatial groups; Within each spatial group, the attention weights between each track and other tracks are calculated to obtain the cross-track weight matrix; Based on the cross-track weight matrix, the track embedding representations of each track are weighted and fused to generate the group perception enhancement representation.

3. The multi-scale UAV trajectory prediction method according to claim 1, characterized in that, The step of performing physical feature value attention processing based on the group-aware enhanced representation and combined with the covariance matrix to generate a refined temporal representation includes: Extract the position covariance submatrix from the covariance matrix; The position covariance submatrix is ​​decomposed into eigenvalues ​​to obtain eigenvalues ​​and corresponding eigenvectors. The principal direction is dynamically determined based on the eigenvalues, and a low-dimensional projection space is constructed based on the eigenvectors corresponding to the principal direction. The group perception enhancement representation is projected onto the low-dimensional projection space, attention is calculated, and the calculation result is mapped back to the original feature space to generate the refined temporal representation.

4. The multi-scale UAV trajectory prediction method according to claim 1, characterized in that, The step of generating multiple trajectory prediction results at different time scales through parallel decoding based on the refined temporal representation includes: Generate corresponding UAV time timestamp embeddings for the first prediction scale, the second prediction scale, and the third prediction scale, respectively; Using the refined time sequence representation as keys and values, and embedding the UAV time timestamps of the first, second, and third prediction scales as queries, the corresponding scales of trajectory prediction results are generated in parallel through a cross-attention mechanism. The trajectory prediction results at the first prediction scale include short-term position and velocity predictions, the trajectory prediction results at the second prediction scale include medium-term position and velocity predictions and maneuver intent classifications, and the trajectory prediction results at the third prediction scale include long-term position and velocity predictions and mission intent classifications.

5. The multi-scale UAV trajectory prediction method according to claim 1, characterized in that, The step of performing physical constraint post-processing on the trajectory prediction results at multiple different time scales to output multi-scale predicted trajectories that satisfy kinematic constraints includes: For each prediction point in the multiple trajectory prediction results at different time scales, check whether its speed, acceleration, and turning rate meet the preset kinematic constraints. If the predicted point does not satisfy the kinematic constraints, then the predicted point is kinematically projected and corrected to satisfy the kinematic constraints. Output the corrected multi-scale predicted track.

6. The multi-scale UAV trajectory prediction method according to claim 1, characterized in that, The step of extracting physical state features, sensor quality features, and motion pattern features based on the confirmed state track data, and performing feature encoding to generate a track embedding representation containing prior physical constraints, includes: The absolute coordinates in the trajectory data are converted into a relative displacement sequence, and combined with the velocity sequence to construct an input feature vector; The sensor type, signal-to-noise ratio, geometric precision factor, and covariance trace are encoded to generate a sensor quality embedding. The motion model type, maneuver intensity, turning rate, and new Mahalanobis distance are encoded to generate a physical state feature vector; Encode the preset kinematic constraint parameters into constraint embeddings; The input feature vector, sensor quality embedding, physical state feature vector, and constraint embedding are fused and encoded to generate the track embedding representation.

7. The multi-scale UAV trajectory prediction method according to claim 1, characterized in that, Also includes: For the exploratory state track, only the first prediction scale prediction among the multiple track prediction results generated by the parallel decoding is executed, and it is output as the short-term prediction result. For the smooth state trajectory, extrapolation prediction is performed using a Kalman filter based on a uniform motion model.

8. A multi-scale unmanned aerial vehicle (UAV) trajectory prediction device, characterized in that, include: The data acquisition module is used to acquire multi-sensor fusion trajectory data. The trajectory data includes the position, velocity, acceleration, covariance matrix and motion model of the UAV at the current time and historical time. The trajectory data carries trajectory status identifiers, which include confirmation status, trial status and smoothing status. The feature extraction and embedding module is used to extract physical state features, sensor quality features and motion pattern features based on the track data for the confirmed state track, and perform feature encoding to generate a track embedding representation that includes physical constraint priors. A cross-track attention module is used to perform cross-track attention processing on the track embedding representations of multiple tracks to generate a group perception enhanced representation; The physical feature value attention module is used to perform physical feature value attention processing based on the group perception enhancement representation and combined with the covariance matrix to generate a refined temporal representation; A multi-scale prediction generation module is used to decode and generate multiple trajectory prediction results at different time scales in parallel based on the refined temporal representation. The physical constraint post-processing module is used to perform physical constraint post-processing on the trajectory prediction results of the multiple different time scales, and output multi-scale predicted trajectories that satisfy kinematic constraints.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the multi-scale UAV trajectory prediction method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, can implement the multi-scale UAV trajectory prediction method as described in any one of claims 1 to 7.

Images