A radar fully automatic tracking method

Abstract

The present application relates to radar signal processing and target tracking technical field, specifically to a kind of radar full-automatic tracking method, comprising the following steps: a, data acquisition and pre-processing assisted by deep learning: synchronous acquisition radar, laser radar, microwave radiometer and environmental data, suppress clutter and extract target features by U-Net network;B, dynamic adaptation of target model driven by reinforcement learning: construct multi-level model library, realize model autonomous switching based on DQN reinforcement learning algorithm, and online iteration model parameters using PSO algorithm, the present application realizes data preprocessing enhancement: through the clutter suppression of U-Net network, the target detection probability is improved to more than 98%, and high-quality data basis is provided for subsequent tracking;Model adaptive capacity is improved: the adaptive capacity of unknown maneuvering mode is improved by 60% by the model switching strategy driven by reinforcement learning, and tracking error is reduced by 35%.

Description

Technical Field

[0001] This invention relates to the field of radar signal processing and target tracking technology, and in particular to a fully automatic radar tracking method. Background Technology

[0002] Existing radar tracking methods suffer from numerous technical bottlenecks: limited noise and clutter suppression during data preprocessing leads to decreased tracking accuracy; target model switching relies on empirical thresholds, resulting in poor adaptability to unknown maneuvering patterns; decision adjustments based solely on a single tracking metric are ill-suited to handle dynamic changes in complex environments; and high trajectory association error rates exist in multi-target tracking, leading to insufficient tracking stability in high-density target scenarios. Therefore, a novel tracking methodology is urgently needed to enhance radar's fully automated tracking capabilities in complex environments. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a fully automatic radar tracking method.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a fully automatic radar tracking method, comprising the following steps: a. Deep learning-assisted data acquisition and preprocessing: Simultaneously acquire radar, lidar, microwave radiometer and environmental data, and use U-Net network to suppress clutter and extract target features; b. Dynamic adaptation of target models driven by reinforcement learning: Construct a multi-level model library, realize autonomous model switching based on DQN reinforcement learning algorithm, and use PSO algorithm to iterate model parameters online; c. Intelligent decision-making and control with multiple indicators linked: By classifying the comprehensive evaluation value CEV, different radar parameter adjustment strategies are executed to form a closed-loop feedback regulation; d. Multi-target tracking and association through spatiotemporal trajectory clustering: Extract target trajectory features and cluster them, construct an association matrix based on the Hungarian algorithm, and combine it with a spatiotemporal conflict resolution mechanism to achieve multi-target tracking.

[0005] Preferably, in step a, the clutter suppression ratio of the U-Net network is ≥40dB, the lidar point cloud density is ≥200 points / ㎡, and the data acquisition cycle is ≤100ms.

[0006] Preferably, the reward function for reinforcement learning in step b is R = 0.6×(1 - |prediction position error| / target distance) + 0.4×(1 - model switching frequency / tracking cycle), and the model switching response time is ≤50ms.

[0007] Preferably, in step c, the comprehensive evaluation value CEV = 0.5×(1 - TER) + 0.3×(SNR / 20) + 0.2×(1 - PL), and the tracking enhancement mode is activated when CEV<0.5.

[0008] Preferably, in step d, the DBSCAN algorithm is used to cluster the target trajectory, with the correlation matrix weights being 0.3 for position distance, 0.4 for trajectory feature similarity, and 0.3 for velocity vector difference, thus reducing the cluster radius by 30% in high-density target scenarios.

[0009] Preferably, the radar data includes range, azimuth, elevation, and radial velocity.

[0010] Preferably, the PSO algorithm iterates the model parameters every 150ms, and the iteration step size is increased by 2 times when the target is maneuvering rapidly.

[0011] Preferably, the closed-loop feedback adjustment adjusts the clutter suppression threshold of the U-Net network every 120ms, and triggers a system self-test when CEV<0.5 for 5 consecutive cycles.

[0012] Preferably, when resolving multi-target conflicts, spatial conflicts are resolved using a fine-resolution mode with beamwidth reduced to 0.5°×0.5°, and identity conflicts are resolved through three-dimensional contour feature matching (matching degree > 0.85).

[0013] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Enhanced data preprocessing: By suppressing clutter through the U-Net network, the target detection probability is increased to over 98%, providing a high-quality data foundation for subsequent tracking; Improved model adaptability: The reinforcement learning-driven model switching strategy improves adaptability to unknown maneuvering modes by 60% and reduces tracking error by 35%; Improved scientific decision-making: The multi-indicator linkage decision-making mechanism enhances the tracking stability of radar in complex environments (the proportion of time with CEV≥0.7) to 92%; Multi-target tracking performance optimization: Combining spatiotemporal trajectory clustering with improved association algorithms, achieving performance improvement at 150 targets / km. 2 In high-density scenarios, the association error rate is reduced to below 3%, and the tracking continuity is significantly enhanced. Attached Figure Description

[0014] Figure 1 This is a schematic diagram illustrating the steps of a fully automatic radar tracking method proposed in this invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0016] Please see Figure 1 This invention provides a technical solution: a fully automatic radar tracking method, comprising the following steps: Deep learning-assisted data acquisition and preprocessing steps: Multi-source data synchronous acquisition: Radar data: Using the active phased array radar (operating frequency band X band), the target's distance (accuracy ±1m), azimuth (accuracy ±0.1°), elevation (accuracy ±0.1°) and radial velocity (accuracy ±0.2m / s) are acquired at a period of 80ms. Auxiliary sensing data: LiDAR (point cloud density 200 points / ㎡) acquires target 3D contour data, microwave radiometer acquires target radiation characteristic parameters, and the sampling period for both is 100ms; Environmental data: Electromagnetic environment monitoring instruments (measurement range 1MHz-18GHz) and wind speed sensors (measurement range 0-60m / s) are deployed to record interference signal strength and wind speed information in real time.

[0017] Deep preprocessing algorithm: Clutter suppression based on U-Net architecture: Convert radar echo data into two-dimensional images, segment target signals and clutter regions through a trained U-Net network, and improve the clutter suppression ratio to over 40dB; Feature enhancement processing: The lidar point cloud data is voxelized to extract the three-dimensional morphological features of the target (such as length, width, height, and surface area), and a target feature vector library is constructed by combining microwave radiation characteristic parameters.

[0018] Reinforcement learning-driven dynamic adaptation steps for target models: Multi-level model library construction: Basic motion models: including uniform circular motion (CVM) models (suitable for targets with a turning radius ≥ 500m) and uniformly accelerated curvilinear motion (AVCM) models (suitable for targets with acceleration ≤ 5m / s²). 2 (Targets moving along curves); special maneuver models: jump maneuver model (for targets that suddenly change direction), spiral maneuver model (for hovering targets), model parameters are initialized offline through historical data training.

[0019] Reinforcement learning model switching: State space definition: includes parameters such as the target's current acceleration, angular velocity, position deviation, and intensity of environmental disturbances; Action space definition: model switching instructions (such as switching from CVM to AVCM), model parameter adjustment amounts (such as noise variance correction values); Reward function design: R = 0.6×(1 - |Prediction position error| / Target distance) + 0.4×(1 - Model switching frequency / Tracking cycle). The agent is trained through deep reinforcement learning (DQN algorithm) to achieve autonomous optimization switching of the model. The model switching response time is ≤50ms.

[0020] Online model parameter iteration: The particle swarm optimization (PSO) algorithm is used to minimize the tracking error, and the model parameters are iteratively updated every 150ms. When a target is detected to enter a new motion mode (such as changing from linear motion to spiral motion), a rapid parameter update mechanism is triggered, and the iteration step size is increased by 2 times, lasting for 3 cycles.

[0021] Intelligent decision-making and control steps involving multiple indicators: Comprehensive tracking quality assessment: Construct a multi-indicator assessment system, including tracking error rate (TER) = tracking error / target distance (threshold ≤ 5%), signal-to-noise ratio (SNR) (threshold ≥ 10dB), and target loss probability (PL) (threshold ≤ 0.05); Calculate the comprehensive assessment value CEV = 0.5×(1 - TER) + 0.3×(SNR / 20) + 0.2×(1 - PL), CEV classification: excellent (≥ 0.9), good (0.7-0.9), average (0.5-0.7), poor (< 0.5). Dynamic control strategy: When CEV is excellent: the radar operates in low-power mode, with a transmit power of 80W and a pulse repetition frequency of 2kHz; when CEV is good: adaptive power adjustment is activated, and the transmit power is dynamically adjusted between 100-120W according to SNR, while the beam scanning range is reduced by 20%; when CEV is average: interference suppression mode is enabled, adaptive frequency hopping is activated (frequency hopping bandwidth 500MHz), and coherent accumulation time is increased to 1ms; when CEV is poor: the tracking enhancement mode is switched to, the transmit power is increased to 250W, beam staring tracking is activated (beam pointing accuracy ±0.05°), and auxiliary sensors are used for data enhancement.

[0022] Closed-loop feedback adjustment: Every 120ms, the comprehensive evaluation value CEV is fed back to the data preprocessing module to adjust the clutter suppression threshold of the U-Net network (the lower the CEV, the stricter the threshold); when CEV < 0.5 for 5 consecutive cycles, the system self-test mechanism is triggered to automatically detect and correct the radar hardware status and sensor calibration parameters.

[0023] Multi-target tracking and association steps of spatiotemporal trajectory clustering: Trajectory feature extraction and clustering: Extract features from the historical trajectory (20 frames in length) of each target, including average velocity, trajectory curvature, acceleration change rate, etc.; Density-based clustering (DBSCAN) is used to cluster multi-target trajectories, grouping targets with similar motion characteristics into one class. The cluster radius is dynamically adjusted based on the target density (target density > 50 targets / km). 2 (At that time, the radius is reduced by 30%).

[0024] An improved multi-target association algorithm: The association matrix is ​​constructed based on the Hungarian algorithm, which comprehensively considers the target location distance (weight 0.3), trajectory feature similarity (weight 0.4), and velocity vector difference (weight 0.3); A time sliding window mechanism is introduced: the association results of the past 5 frames are cumulatively analyzed, and the association is confirmed to be correct when the cumulative confidence of a certain association pair is > 0.8.

[0025] Conflict resolution mechanism: Spatial conflict resolution: When the predicted intersection probability of the trajectories of two targets is > 70%, the fine resolution mode is activated, and the radar beamwidth is reduced to 0.5°×0.5°; Identity conflict resolution: The target's 3D contour features extracted by LiDAR are used for matching. When the matching degree is > 0.85, the target's identity is confirmed, thus resolving the trajectory confusion problem.

[0026] Taking a ground-based air defense radar system (operating in the S-band) as an example, the implementation steps are as follows: Hardware configuration: Radar system: Active phased array radar, peak power 50kW, beam scanning range azimuth 0-360°, elevation 0-90°, maximum detection range 500km; Auxiliary equipment: 16-line lidar (point cloud frame rate 10Hz), microwave radiometer (operating frequency band 10-30GHz). Processing unit: An embedded processing platform equipped with a GPU (20 TFLOPS computing power) and an FPGA (800 Gbps parallel processing capability).

[0027] Software implementation: Data preprocessing: A U-Net clutter suppression model was trained based on the PyTorch framework. The training dataset contained 100,000 sets of radar echo data containing clutter. Model switching: The DQN reinforcement learning model is implemented using TensorFlow, with an experience replay pool capacity of 1 million records and a learning rate of 0.001. Decision control: Based on the PLC logic controller, parameter adjustment of multiple indicators is realized, with a response time of ≤30ms; Multi-target processing: C++ implementation of DBSCAN clustering and Hungarian algorithm, processing time ≤50ms / frame (100 targets).

[0028] Test results: Single target tracking: for a velocity of 500 m / s² and an acceleration of 10 m / s². 2 For maneuvering targets, the tracking error is ≤3m; Multi-target tracking: In a dense scene with 100 targets, continuous tracking for 1 hour resulted in a trajectory association error rate of 2.3%. Anti-interference performance: In an environment where the interference signal strength is 20dB higher than the target signal, the CEV remains above 0.75.

[0029] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A fully automatic radar tracking method, characterized in that, Includes the following steps: a. Deep learning-assisted data acquisition and preprocessing: Simultaneously acquire radar, lidar, microwave radiometer and environmental data, and use U-Net network to suppress clutter and extract target features; b. Dynamic adaptation of target models driven by reinforcement learning: Construct a multi-level model library, realize autonomous model switching based on DQN reinforcement learning algorithm, and use PSO algorithm to iterate model parameters online; c. Intelligent decision-making and control with multiple indicators linked: By classifying the comprehensive evaluation value CEV, different radar parameter adjustment strategies are executed to form a closed-loop feedback regulation; d. Multi-target tracking and association through spatiotemporal trajectory clustering: Extract target trajectory features and cluster them, construct an association matrix based on the Hungarian algorithm, and combine it with a spatiotemporal conflict resolution mechanism to achieve multi-target tracking.

2. The fully automatic radar tracking method according to claim 1, characterized in that: In step a, the clutter suppression ratio of the U-Net network is ≥40dB, the point cloud density of the lidar is ≥200 points / ㎡, and the data acquisition cycle is ≤100ms.

3. The fully automatic radar tracking method according to claim 1, characterized in that: In step b, the reward function for reinforcement learning is R = 0.6×(1 - |prediction position error| / target distance) + 0.4×(1 - model switching frequency / tracking cycle), and the model switching response time is ≤50ms.

4. The fully automatic radar tracking method according to claim 1, characterized in that: In step c, the comprehensive evaluation value CEV = 0.5×(1 - TER) + 0.3×(SNR / 20) + 0.2×(1 - PL). When CEV < 0.5, the tracking enhancement mode is activated.

5. The fully automatic radar tracking method according to claim 1, characterized in that: In step d, the DBSCAN algorithm is used to cluster the target trajectory. The weights of the association matrix are position distance 0.3, trajectory feature similarity 0.4, and velocity vector difference 0.

3. In high-density target scenes, the clustering radius is reduced by 30%.

6. The fully automatic radar tracking method according to claim 1, characterized in that: The radar data includes range, azimuth, elevation, and radial velocity.

7. The fully automatic radar tracking method according to claim 1, characterized in that: The PSO algorithm iterates the model parameters every 150ms, and the iteration step size is doubled when the target is maneuvering rapidly.

8. The fully automatic radar tracking method according to claim 1, characterized in that: The closed-loop feedback adjustment adjusts the clutter suppression threshold of the U-Net network every 120ms, and triggers a system self-test when CEV<0.5 for 5 consecutive cycles.

9. The fully automatic radar tracking method according to claim 1, characterized in that: When resolving multi-target conflicts, spatial conflicts are resolved using a fine-resolution mode with beamwidth reduced to 0.5°×0.5°, while identity conflicts are resolved through three-dimensional contour feature matching (matching degree > 0.85).

Images