Low-altitude target trajectory tracking model and training method

Abstract

The invention discloses a low-altitude target trajectory tracking model and a training method, and aims to solve the technical problems of large trajectory tracking error caused by low-altitude complex channel interference, insufficient positioning precision under multi-node cooperation, poor model generalization ability caused by sensor deployment density change and the like. By integrating spatial-temporal feature extraction, attention mechanism fusion and an adaptive filtering post-processing module and combining a transfer learning strategy of high-altitude scene pre-training and low-altitude scene fine adjustment, the positioning robustness of a low-altitude target is improved, the deployment cost is remarkably reduced, and cross-scene rapid adaptation and efficient modeling are realized.

Description

Technical Field

[0001] The invention relates to a low-altitude target trajectory tracking model and a training method. Background Art

[0002] As a strategic fulcrum for new urban three-dimensional transportation and smart logistics, the development of the low-altitude economy faces unique safety and regulatory technical challenges. In complex airspace environments, aircraft must perform three-dimensional maneuvers between obstacles such as building clusters and undulating terrain, making electromagnetic signals susceptible to multipath and obstruction interference. At the same time, scenarios such as drone logistics and emergency rescue place stringent demands on dynamic responsiveness, requiring millisecond-level obstacle avoidance decisions and real-time path replanning. It is worth noting that while the large-scale deployment of low-cost sensor networks is a key prerequisite for the sustainable development of the low-altitude economy, hardware performance limitations necessitate algorithmic innovation to compensate for the loss of accuracy caused by declining data quality.

[0003] The current technological system faces multiple bottlenecks. The established satellite navigation and airborne positioning systems for traditional high-altitude missions are not directly adaptable to the monitoring needs of low-altitude targets (such as drones and eVTOLs), which are small, highly maneuverable, and require operation in complex electromagnetic environments. The accuracy-cost trade-off is particularly prominent. Traditional high-precision radar systems, while advantageous, are expensive, limiting network coverage density. Low-cost sensors, while capable of wide-area deployment, suffer from significant noise interference due to hardware limitations. A more severe challenge lies in data closed-loop verification. Because new operational models such as eVTOL flight corridors and drone logistics networks have yet to achieve large-scale operation, real-world trajectory data is severely scarce. This vicious cycle of "system first, data later" severely constrains the training and validation of intelligent monitoring algorithms. These intertwined technical challenges create a dilemma: the maturity of intelligent monitoring systems requires a large amount of real-world data, and the accumulation of this data relies on a robust monitoring system. Summary of the Invention

[0004] This paper proposes a low-altitude target trajectory tracking model and training method, aiming to address technical challenges such as large trajectory tracking errors caused by complex low-altitude channel interference, insufficient positioning accuracy under multi-node collaboration, and poor model generalization due to variations in sensor deployment density. By integrating spatiotemporal feature extraction, attention mechanism fusion, and adaptive filtering post-processing modules, combined with a transfer learning strategy that combines high-altitude scene pre-training with low-altitude scene fine-tuning, the model improves the robustness of low-altitude target positioning while significantly reducing deployment costs, enabling rapid adaptation and efficient modeling across scenarios.

[0005] In the first aspect, a low-altitude target trajectory tracking model includes: a data preprocessing module, which is configured to: receive spatiotemporal data from multiple ground radar nodes, preprocess the spatiotemporal data, and generate a standardized input tensor; a spatiotemporal encoder, which is configured to: perform a convolution operation on the input tensor through a one-dimensional convolutional network to generate a spatial feature representation that describes the spatial correlation features between multiple nodes F t ; Input the time series data of each node into the long short-term memory network, extract the time dynamic features, and generate the hidden state representation of each node ; Set the hidden state of all nodes Splicing to form a full-time coding representation that describes the overall characteristics of the node in the time series H t ; Attention fusion module, which is configured to: based on spatial feature representation F t , calculate the query-key-value vector across nodes, dynamically weight and aggregate the spatial information of key nodes, and generate spatial attention results s t ; Based on full-time coding representation H t , process the node sequence across time, focus on the key node behavior in the time series, and generate time attention results v t ; The spatial attention result s t and temporal attention results v t Splicing to generate context feature vector c t ; and LSTM decoder, which is configured to generate the target trajectory in an autoregressive manner, with the initial input initialized based on the mean of historical data; each step input is composed of the predicted trajectory of the previous moment and the current context feature vector c t Splicing, update the hidden state through the long short-term memory network; generate the trajectory prediction result at the current moment, and correct the prediction accuracy through the residual to output the complete trajectory sequence.

[0006] In some examples, the spatiotemporal data includes the spatial location of the node, signal arrival time, and received signal strength.

[0007] In some examples, the received signal strength is adaptively normalized to eliminate dimensionality differences between different radars to generate a standardized input tensor.

[0008] In some examples, a UKF module is used to perform state prediction and observation update on the trajectory output by the LSTM decoder, and the trajectory is smoothed in an iterative optimization manner to generate a smoothed trajectory sequence.

[0009] In the second aspect, a training method for the low-altitude target trajectory tracking model includes: a pre-training stage and a fine-tuning stage; in the pre-training stage, the model is trained using the ADS-B message trajectory of civil aviation passenger aircraft and synchronous ground radar observation data to learn the cross-node and cross-time spatiotemporal correlation modeling capabilities under the conditions of large air distance, weak signal, and sparse nodes, so that the model has the ability to resist channel interference, sequence stable prediction and spatial multi-node fusion, and completes the initialization of the model weights and the acquisition of attention weight distribution.

[0010] In some examples, the fine-tuning stage uses partial low-altitude actual flight data or enhanced synthetic data, and only fine-tunes the LSTM decoder and normalization parameters by freezing the spatiotemporal encoder network to adapt to the characteristics of low-altitude scenes.

[0011] In some examples, the adjustment strategy of the fine-tuning stage includes: matching the RSSI dynamic range of the low-altitude scene to optimize the normalization strategy.

[0012] In some examples, the adjustment strategies in the fine-tuning phase include: adjusting the receptive field through position bias or mask of the attention mechanism to adapt to changes in radar deployment density; or shortening the LSTM decoding step size to improve the prediction accuracy of low-altitude target track curvature.

[0013] In a third aspect, a computer system includes a processor and a memory. After the low-altitude target trajectory tracking model is trained, it is stored in the memory and configured to be executed by the processor.

[0014] In a fourth aspect, a computer-readable storage medium is provided for storing non-transitory computer-readable instructions, which are capable of running the low-altitude target trajectory tracking model when the non-transitory computer-readable instructions are executed by a computer. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of a low-altitude target trajectory tracking model in one embodiment of the present invention.

[0016] Figure 2 2 is a schematic diagram of the low-altitude target trajectory tracking model training phase in one embodiment of the present invention. DETAILED DESCRIPTION

[0017] like Figure 1As shown, the present invention constructs a "verification-first framework (i.e., a low-altitude target trajectory tracking model)." By jointly modeling high-altitude ADS-B trajectories and ground radar observation data, it completes feasibility verification and optimized design of the perception structure and decoding module under complex high-altitude channel conditions. Through prior training on high-altitude scenarios, this framework achieves stable convergence of the algorithm structure and training mechanism, laying the foundation for subsequent model migration. On this basis, a "pre-training + fine-tuning" strategy is adopted: first, the network weights trained at high altitude are used as the initial parameters for low-altitude scenarios. Subsequently, only the decoder parameters and normalization module are fine-tuned based on low-altitude actual flight data or enhanced synthetic data, allowing for rapid adaptation to changes in low-altitude radar deployment density and differences in channel characteristics.

[0018] This migration mechanism has dual advantages: on the one hand, the physical characteristics of low-altitude channels (such as close distance and strong signals) are superior to those of high-altitude scenarios, which provides a natural rationality for model migration; on the other hand, it meets the core requirements of the future low-altitude economy for the "independent perception capability" of ground auxiliary facilities systems, that is, cross-scenario deployment can be achieved through parameter adaptation without reconstructing the algorithm architecture.

[0019] The low-altitude target trajectory tracking model utilizes an end-to-end deep learning modeling framework. By integrating four core modules: spatiotemporal feature extraction, attention mechanism fusion, trajectory prediction, and adaptive filtering post-processing, it efficiently mines the spatiotemporal correlation information of distributed radar nodes, thereby achieving low-altitude target localization with high robustness and low-cost control. The following details the functions and design details of each module of the model.

[0020] 1. Data preprocessing module At each time step The system receives data from three ground radar sensors, each sensor node i Collect the following 5-dimensional features: node spatial position ;Signal arrival time Received signal strength . Therefore, the input tensor at a single moment is: , the entire observation sequence is , TOA and RSSI are adaptively normalized according to the node signal-to-noise ratio (SNR), and the dimension remains unchanged after normalization :

[0021] Where: Meaning: Radar nodes at time The normalized received signal strength is used to eliminate dimensional differences and enhance the model’s compatibility with multi-node heterogeneous data. Meaning: Radar nodes at time The raw received signal strength (in dBm) is directly measured by the radar hardware. Meaning: The average RSSI value of a node in a sliding time window (for example, k steps before and after) can be calculated by sliding mean; its function is to reflect the signal baseline level of the node under specific SNR conditions. Meaning: The standard deviation of the RSSI value within the same time window, indicating the volatility of the node data.

[0022] 2. Space-time encoder (1) Spatial feature extraction (CNN) For every moment , the input is The spatial correlation features between the three sensors are extracted by one-dimensional convolution. In one embodiment, the convolution kernel size is 1, the step size is 1, and the number of output channels is ,Right now: .

[0023] (2) Temporal Modeling (LSTM) The time series of each sensor Enter the unidirectional LSTM network. Assume the hidden state dimension is , then the encoding of each node is expressed as: .

[0024] After all nodes are spliced together, a full-time coding representation is formed:

[0025] 3. Attention Fusion In order to enhance the model's ability to focus on key nodes and key moments, spatial attention and temporal attention mechanisms are designed.

[0026] (1) Spatial attention (across nodes): Let For attention query, key, and value vectors:

[0027] (2) Temporal attention (across time): Similarly, for each node , aggregate its time series context:

[0028] Node-level outputs are averaged:

[0029] (3) Decoding context splicing:

[0030] 4. LSTM Decoder (Trajectory Prediction) A single-layer LSTM decoder is used to generate target trajectories in an autoregressive manner.

[0031] Define the initial input trajectory point : Can be initialized by historical mean.

[0032] The input of each step is the concatenation of the prediction result of the previous step and the context feature:

[0033] Update hidden state through LSTM

[0034] Predict the current trajectory point:

[0035] Add residual correction:

[0036] Generate trajectory sequence:

[0037] 5. UKF module In order to further improve the trajectory smoothness and continuity, the unscented Kalman filter is introduced to post-process the decoder output results.

[0038] Define state variables:

[0039] Status prediction:

[0040] Noise covariance update:

[0041] Observation Update:

[0042] The final output is a smoothed trajectory sequence .

[0043] The model combines end-to-end sequence learning with physical constraint filtering, and has strong generalization capabilities and deployment flexibility. It is suitable for various mission scenarios such as real-time tracking of low-altitude targets, trajectory recovery, and intrusion detection.

[0044] like Figure 2 To address the lack of real-world trajectory data in the early stages of the low-altitude economy and maximize the engineering value of validated high-altitude models, this paper draws on the "pretrain-finetune" approach to propose a transferable model expansion mechanism. The core concept is to first perform large-scale pretraining on mid- and high-altitude target data to acquire robust perception and modeling capabilities. The model is then transferred to low-altitude scenarios, achieving adaptive adaptation to changes in radar density and signal patterns through parameter fine-tuning while maintaining the same structure.

[0045] (1) Pre-training stage (high-altitude model training) Data source: ADS-B message trajectories from civil airliners and synchronous ground radar observations (TOA, RSSI).

[0046] Training goal: To learn the ability to model spatiotemporal correlations across nodes and time under conditions of large air distances, weak signals, and sparse nodes.

[0047] Output capability: The model has the ability to resist channel interference, stable sequence prediction, and spatial multi-node fusion.

[0048] Parameter initialization: Complete the model weights (such as network parameters W , bias b ) and obtain a reasonable attention weight distribution.

[0049] (3) Fine-tuning stage (low-altitude scene adaptation) Data conditions: Use some low-altitude actual flight data (such as drone patrol data) or enhanced synthetic data.

[0050] Fine-tuning method: Freeze the spatiotemporal encoder network and only fine-tune the LSTM decoder and normalization parameters.

[0051] Adjustment strategy: a) Normalization strategy: match the RSSI dynamic range of low-altitude scenes; b) Radar deployment density change: through The position bias or mask mechanism in attention adjusts the receptive field; c) track curvature adaptation, appropriately shortening the LSTM decoding step size, and improving the low-altitude target trajectory prediction accuracy.

[0052] The design of this invention is based on the following core logic: low-altitude targets are closer to ground-based radars, and propagation losses are significantly reduced, resulting in a higher signal-to-noise ratio (SNR) and measurement accuracy; at the same time, by increasing the number of low-altitude sensors, the node density can be flexibly controlled to improve spatial observation resolution; in addition, the model structure that has been verified in high-altitude scenarios (such as the combination of spatiotemporal attention mechanism and long-short-term memory network) has good generalization ability, and there is no need to reconstruct the model structure, only parameters need to be adjusted to quickly adapt to low-altitude target tracking needs.

[0053] The present invention also provides an embodiment of a computer system. The computer system includes a processor and a memory. The memory is configured to store non-transitory computer-readable instructions (e.g., one or more computer program modules). The processor is configured to execute the non-transitory computer-readable instructions. When the processor executes the non-transitory computer-readable instructions, one or more steps of the low-altitude target trajectory tracking model and training method are performed. The memory and processor may be interconnected via a bus system and / or other connection mechanisms.

[0054] For example, a processor can be a central processing unit (CPU), a graphics processing unit (GPU), or other forms of processing units with data processing capabilities and / or program execution capabilities. The processor can be a general-purpose processor or a special-purpose processor that can control other components in the computer to perform desired functions.

[0055] For example, the memory may include any combination of one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Non-volatile memory may include, for example, read-only memory (ROM), a hard disk, an erasable programmable read-only memory (EPROM), a compact disc read-only memory (CD-ROM), a USB memory, a flash memory, etc. One or more computer program modules may be stored on the computer-readable storage medium, and the processor may execute the one or more computer program modules to implement various functions of the computer.

[0056] The present invention also provides a computer-readable storage medium for storing non-transitory computer-readable instructions. When the non-transitory computer-readable instructions are executed by a computer, one or more steps in the above-mentioned low-altitude target trajectory tracking model and training method can be implemented. When the low-altitude target trajectory tracking model of the present invention is implemented in the form of software and sold or used as an independent product, it can be stored in a computer-readable storage medium. For relevant descriptions of the storage medium, please refer to the corresponding description of the memory in the computer system above and will not be repeated here.

Claims

1. A low-altitude target trajectory tracking model, characterized in that: include: A data preprocessing module is configured to: receive spatiotemporal data from multiple ground radar nodes, preprocess the spatiotemporal data, and generate a standardized input tensor; The spatiotemporal encoder is configured to perform a convolution operation on the input tensor through a one-dimensional convolutional network to generate a spatial feature representation that describes the spatial correlation features between multiple nodes. F t ; Input the time series data of each node into the long short-term memory network, extract the time dynamic features, and generate the hidden state representation of each node ; Set the hidden state of all nodes Splicing to form a full-time coding representation that describes the overall characteristics of the node in the time series H t ; Attention fusion module, which is configured to: F t , calculate the query-key-value vector across nodes, dynamically weight and aggregate the spatial information of key nodes, and generate spatial attention results s t ; Based on full-time coding representation H t , process the node sequence across time, focus on the key node behavior in the time series, and generate time attention results v t ; The spatial attention result s t and temporal attention results v t Splicing to generate context feature vector c t ; The LSTM decoder is configured to generate the target trajectory in an autoregressive manner, with the initial input initialized based on the mean of historical data; each step input is composed of the predicted trajectory of the previous moment and the current context feature vector c t Splicing, update the hidden state through the long short-term memory network; generate the trajectory prediction result at the current moment, and correct the prediction accuracy through the residual to output the complete trajectory sequence.

2. The low-altitude target trajectory tracking model according to claim 1, characterized in that: Spatiotemporal data includes the spatial location of the node, signal arrival time, and received signal strength.

3. The low-altitude target trajectory tracking model according to claim 2, characterized in that: Adaptively normalize the received signal strength to eliminate the dimensionality differences between different radars and generate a standardized input tensor.

4. The low-altitude target trajectory tracking model according to claim 1, characterized in that: The UKF module is used to perform state prediction and observation update on the trajectory output by the LSTM decoder, and the trajectory is smoothed by iterative optimization to generate a smoothed trajectory sequence.

5. A training method for a low-altitude target trajectory tracking model according to any one of claims 1 to 4, characterized in that: include: Pre-training stage and fine-tuning stage; in the pre-training stage, the model is trained using ADS-B message trajectories of civil airliners and synchronous ground radar observation data to learn cross-node and cross-time spatiotemporal correlation modeling capabilities under conditions of large air distance, weak signal, and sparse nodes, so that the model has the ability to resist channel interference, stable sequence prediction, and spatial multi-node fusion, and completes the initialization of the model weights and the acquisition of attention weight distribution.

6. The method according to claim 5, characterized in that The fine-tuning stage uses some low-altitude actual flight data or enhanced synthetic data, and only fine-tunes the LSTM decoder and normalization parameters by freezing the spatiotemporal encoder network to adapt to the characteristics of low-altitude scenarios.

7. The method according to claim 5, characterized in that The adjustment strategy in the fine-tuning stage includes: matching the RSSI dynamic range of the low-altitude scene to optimize the normalization strategy.

8. The method according to claim 5, characterized in that The adjustment strategies in the fine-tuning stage include: adjusting the receptive field through position bias or mask of the attention mechanism to adapt to changes in radar deployment density; or shortening the LSTM decoding step size to improve the prediction accuracy of low-altitude target track curvature.

9. A computer system, characterized in that: The method comprises a processor and a memory, wherein the low-altitude target trajectory tracking model according to any one of claims 1 to 4 is stored in the memory after being trained and is configured to be executed by the processor.

10. A computer-readable storage medium for storing non-transitory computer-readable instructions, characterized in that: When the non-transitory computer-readable instructions are executed by a computer, the low-altitude target trajectory tracking model according to any one of claims 1 to 4 can be run.

Images