Unmanned aerial vehicle autonomous tracking method and system based on millimeter wave radar

Abstract

The invention discloses an unmanned aerial vehicle autonomous tracking method and system based on a millimeter wave radar, and belongs to the technical field of unmanned aerial vehicle autonomous tracking. According to the method, a radar-motion coupling model and an adaptive trajectory prediction algorithm are constructed, and after radar signal propagation distortion is dynamically compensated, a layered processing architecture is adopted to realize target separation and trajectory optimization; the front end solves a point cloud fuzzy problem through an adaptive Euclidean clustering algorithm, and the rear end completes trajectory prediction and control through an improved extended Kalman filtering and model prediction control composite architecture. The system mainly comprises a dynamic noise compensation module, a self-adaptive clustering module and a composite control module. According to the method, the problems of low tracking precision, large response delay and target adhesion of the unmanned aerial vehicle in a complex dynamic scene are solved, static target positioning errors and dynamic tracking errors are reduced, and high robustness and real-time performance are improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous drone tracking technology, specifically a method and system for autonomous drone tracking based on millimeter-wave radar, which is particularly suitable for military reconnaissance, disaster relief, and logistics delivery scenarios in complex and dynamic environments. Background Technology

[0002] With the widespread application of multi-rotor UAVs in military reconnaissance, disaster relief, and smart logistics, dynamic target tracking in complex urban environments faces severe challenges. Densely built-up areas cause significant attenuation of satellite positioning signals, and complex electromagnetic environments interfere with wireless communication links, leading to distortion of point cloud data acquired by millimeter-wave radar. The velocity-range coupling effect of moving targets causes signal spectrum broadening, and point cloud blurring significantly reduces target separation accuracy. As mission duration increases, positioning errors accumulate, tracking accuracy continues to decline, severely impacting the reliability of UAVs in critical missions.

[0003] Millimeter-wave radar autonomous control systems face three major technical bottlenecks in urban environments: point cloud blurring caused by moving targets significantly reduces the signal-to-noise ratio, and fixed-parameter filtering struggles to adapt to dynamic changes; traditional clustering algorithms cannot address the nonlinear decay of point cloud density with distance, making long-range target identification difficult; and the single control architecture exhibits significant response delays, making it difficult to meet the real-time tracking requirements of high-speed maneuvering targets. These technical deficiencies limit the practical application effectiveness of millimeter-wave radar in complex urban scenarios.

[0004] In existing technologies, multiple-input multiple-output waveform diversity techniques have failed to solve the problem of dynamic noise coupling; the parameter decoupling scheme at the theoretical level of frequency-controlled array multiple-input multiple-output radar leads to a surge in hardware complexity; and the combination of traditional density clustering and Kalman filtering methods lacks stability in dynamic scenes. A more fundamental deficiency lies in the lack of an adaptive mechanism for the dynamic urban environment, making it unable to maintain stable tracking under multiple constraints such as building obstruction, electromagnetic interference, and high-speed movement. Summary of the Invention

[0005] This application provides a dynamic target tracking method and system for unmanned aerial vehicles (UAVs) based on millimeter-wave radar. By establishing a speed-distance bivariate noise compensation model, the problem of millimeter-wave radar signal distortion in motion scenarios is solved. An adaptive Euclidean clustering algorithm is used to achieve accurate separation of dynamic targets. The stability of trajectory tracking is optimized by combining a composite control architecture, and a closed-loop adaptive strategy adjustment mechanism is constructed.

[0006] This invention relates to an autonomous tracking method for unmanned aerial vehicles based on millimeter-wave radar, comprising the following steps:

[0007] S100: Real-time acquisition of 3D point cloud data of the target UAV and surrounding obstacles via millimeter-wave radar acquisition template. A dynamic coordinate system transformation model is constructed based on the UAV's velocity and acceleration parameters. Position calibration is performed using calculated spatial offset. To address signal distortion caused by motion, a power noise function with velocity and distance as dual variables is established. And through the parameters of the adaptive Butterworth filter The cutoff frequency is dynamically adjusted to suppress the spectral broadening caused by Doppler frequency shift;

[0008] The formula for calculating spatial offset is:

[0009] ;

[0010] In the formula, This indicates the target's position relative to the radar sensor itself;

[0011] Indicates the absolute position of the target in the global reference frame;

[0012] Let be the rotation matrix, representing the rotation angle of the radar coordinate system relative to the world coordinate system. ;

[0013] These represent the radar's position offset in the x and y directions of the world coordinate system, respectively.

[0014] The constructed noise power function is wherein the noise power is proportional to the 1.5th power of the velocity and proportional to the distance; the noise power function

[0015] In the formula, It is the noise variance value determined by the target velocity v and range R in millimeter-wave radar measurements;

[0016] v represents the target's relative radial velocity;

[0017] R represents the straight-line distance between the target and the radar;

[0018] The parameters of the Tevos filter ;

[0019] In the formula, The dynamic cutoff frequency representing the target velocity v and distance R;

[0020] This refers to the system sampling frequency;

[0021] This represents the noise standard deviation function related to speed and distance.

[0022] S200 uses an adaptive Euclidean clustering algorithm and an adaptive clustering module to process point cloud data. It calculates point cloud density and relative distance through kernel density estimation, and dynamically adjusts the neighborhood radius according to the distance to the target to accurately extract the target's location information.

[0023] Specifically, adaptive Euclidean clustering is performed on the processed point cloud, and the local density of each data point is calculated using a Gaussian kernel function. and relative distance The calculation formula is:

[0024] ;

[0025] In the formula, Point The local density estimate;

[0026] Point The minimum distance to a higher density point;

[0027] Represents the Euclidean distance between two points, where This indicates the point cloud coordinates of the target detected by the radar;

[0028] h represents the bandwidth parameter of the kernel function;

[0029] Based on target distance The neighborhood radius is dynamically adjusted; when the target is relatively close, the neighborhood radius is set to 0.3 times the maximum distance threshold.

[0030] When the target is at a relatively far distance, set the neighborhood radius to 0.6 times the maximum distance threshold;

[0031] For clusters with a density difference of less than 20% and a distance of less than 60% of the dynamic threshold, a density similarity cluster merging operation is performed to avoid close-range clustering of targets and over-segmentation of distant targets.

[0032] The cluster merging formula is:

[0033] ;

[0034] In the formula, d represents the effective radius of the cluster;

[0035] , These are the dynamic distance thresholds for the two clusters, respectively;

[0036] Indicates the density difference threshold;

[0037] This represents the core density of the first and second clusters.

[0038] S300 uses a composite control architecture and modules to predict and optimize the trajectory based on the target's position information: the outer loop proportional-integral-derivative controller generates the basic trajectory, and the inner loop model predictive controller optimizes the attitude control to enable the UAV to track stably.

[0039] The use of a composite control architecture for trajectory prediction and optimization specifically includes:

[0040] Construct a six-degree-of-freedom state equation based on the target location information:

[0041] ;

[0042] In the formula, The position of the target in the inertial coordinate system. Linear velocity, Yaw angle Angular velocity;

[0043] Design the velocity-dependent noise covariance matrix, where the lateral velocity noise figure is twice that of the longitudinal velocity.

[0044] The outer loop proportional-integral-derivative controller is configured with a proportional coefficient of 1.2, an integral coefficient of 0.05, and a derivative coefficient of 0.3.

[0045] Extended Kalman filter is used to estimate the target motion state, and the nonlinear system is linearized by Jacobian matrix. The predicted trajectory is input into the outer loop PID controller to generate the basic control quantity. At the same time, the inner loop model predictive controller optimizes the attitude angle command with the objective function to achieve composite control with fast response delay.

[0046] The inner-loop model predicts the control objective function, which also includes:

[0047] The prediction time domain is set to 5 control periods;

[0048] Configure a trajectory tracking weight matrix to enhance position tracking permissions;

[0049] Configure the control input weight matrix to limit sudden changes in control commands;

[0050] The objective function expression is:

[0051] ) ;

[0052] It is the noise covariance matrix of the time-varying process;

[0053] () is a constructor for diagonal matrices;

[0054] , These represent the target linear velocity and the target angular velocity, respectively.

[0055] The S400 drone communicates with the ground station and monitors the distribution of environmental obstacles and changes in the target's motion state in real time on the ROS / Gazebo simulation platform. The control parameters and clustering thresholds are dynamically adjusted through the strategy update module.

[0056] When the distance to an obstacle is detected to be less than the safety threshold, the trajectory replanning mechanism is triggered;

[0057] Obstacles and target point clouds are separated in real time using an adaptive Euclidean clustering algorithm;

[0058] When response latency is reduced, continuous turning tracking of dynamic targets is supported.

[0059] The S500 and UAV update the target position information based on the point cloud information of dynamic targets and obstacles acquired and processed in real time by millimeter-wave radar, and continue to predict the tracking trajectory. Based on the closed-loop feedback results, the tracking trajectory is continuously optimized to achieve adaptive tracking strategy adjustment for dynamic targets.

[0060] The S600, based on an adjusted tracking strategy, enables UAVs to perform real-time, continuous, and stable tracking of dynamic targets.

[0061] This application also provides a dynamic target tracking system for multi-rotor unmanned aerial vehicles based on millimeter-wave radar, including:

[0062] The millimeter-wave radar acquisition module is used to acquire point cloud data of the target UAV and obstacles. The UAV equipped with millimeter-wave radar acquires real-time 3D point cloud data of the target and surrounding obstacles. The real-time motion state parameters of the UAV, including velocity, acceleration and attitude angle, are analyzed by the dynamic offset compensation unit. The spatial position offset is calculated based on the dynamic coordinate transformation model to eliminate the cumulative positioning drift error caused by the movement of the aircraft.

[0063] The dynamic noise compensation module is used to construct a bivariate noise model related to speed and distance. It calls the speed-distance bivariate noise model to construct a nonlinear noise function that correlates the target's speed and distance variables. Through an adaptive Butterworth filter, the frequency response characteristics are dynamically adjusted to effectively suppress the Doppler spectrum broadening effect caused by the target's high-speed motion, thus solving the problem of point cloud data blurring and distortion. The filter parameters are dynamically updated with the motion state, ensuring the integrity of the point cloud topology is maintained even in high-speed maneuvering scenarios above 5 m / s.

[0064] An adaptive clustering module is used to perform distance-dependent neighborhood radius adjustment. This module includes a target separation module, which calculates the local density distribution characteristics of point cloud data based on the kernel density estimation algorithm, identifies the initial cluster core region through peak detection, and selects effective cluster center points by combining the relative distance judgment criterion.

[0065] A multi-scale clustering strategy is adopted to dynamically configure the neighborhood search radius. In the near-field region where the target distance is less than 5 meters, a compact radius parameter is used to prevent the target point cloud from sticking together. In the far-field region where the distance is greater than 20 meters, the radius parameter is expanded to compensate for the point cloud density attenuation and avoid over-segmentation.

[0066] It also introduces density similarity constraint rules to automatically merge discrete clusters that are spatially adjacent and have matching density distribution characteristics, which significantly improves the complete recognition rate of broken targets behind occluders.

[0067] The composite control module is used to realize the cascade optimization of extended Kalman filtering and model predictive control. The trajectory prediction unit of this module uses the extended Kalman filtering algorithm to solve the target's six-degree-of-freedom pose state, and performs local linearization of the nonlinear motion equations through the Jacobian matrix to estimate the target's position, velocity and turning angle in three-dimensional space in real time.

[0068] The attitude optimization unit deploys a model predictive controller and generates the optimal attitude control command through a look-ahead rolling optimization mechanism: a multi-objective function containing trajectory tracking error and actuator energy consumption is constructed in the prediction time domain, and the optimal attitude angular rate command set is solved through quadratic programming;

[0069] The outer-loop proportional-integral-derivative controller generates basic propulsion commands based on trajectory prediction results. Its output is coupled with the torque compensation commands of the inner-loop model predictive controller to achieve coordinated optimization of heading accuracy and attitude stability.

[0070] The simulation verification module is used to test tracking performance on the ROS / Gazebo platform. It constructs a dynamic test environment on the robot operating system and physical simulation platform to simulate complex multi-dimensional scenarios such as urban building occlusion, electromagnetic signal interference, and sudden weather disturbances. The specific method is as follows:

[0071] The target movement speed range is set to cover 0-10 m / s, the detection distance range spans from 0.8 m to 100 m, programmable dynamic obstacles are deployed to simulate vehicle and aircraft conflicts, and a low-latency communication link is established between the UAV and the ground control station to transmit radar point cloud processing results and control command execution status back in real time.

[0072] The tracking decision module, including the strategy update module, monitors the dynamics of environmental variables and changes in target motion characteristics in real time, including target acceleration mutation threshold triggering, real-time calculation of environmental occlusion density, and analysis of relative velocity vectors of obstacles.

[0073] The system parameters are dynamically adjusted based on the monitoring results: when the tracking error exceeds the set threshold, the proportional control coefficient is automatically increased to enhance the response speed; when the intensity of environmental signal interference increases, the clustering neighborhood radius is adaptively shrunk to improve the target separation robustness; and the noise covariance matrix configuration of the trajectory prediction algorithm is continuously optimized to adapt to the target maneuvering mode switching.

[0074] The closed-loop tracking module drives the brushless motor actuator according to optimized control commands and adopts a feedforward-feedback composite control structure to eliminate the influence of aerodynamic disturbances. The feedforward branch predicts the inertial compensation amount based on the target kinematic model, and the feedback branch extracts the acceleration signal through the differential tracker for phase lead correction, so as to realize the smooth transition of control commands under high-speed turning conditions. Combined with the target position feedback collected in real time by millimeter-wave radar, a closed-loop tracking control system is formed to ensure that the dynamic target is always in the effective detection field of the sensor.

[0075] Existing UAV tracking technologies rely on multi-sensor fusion schemes, which suffer from high hardware complexity, signal attenuation under urban electromagnetic interference, and trajectory deviation caused by dynamic noise. This application leverages the strong penetration and all-weather capability of millimeter-wave radar, combined with a dynamic noise compensation model to reduce signal distortion, utilizes an adaptive clustering algorithm to improve target separation success rate, and reduces tracking errors through a composite control architecture. While maintaining a simplified hardware structure, it significantly improves positioning accuracy and system robustness in complex environments, providing reliable technical support for scenarios such as military reconnaissance and disaster relief. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating the UAV autonomous tracking method based on millimeter-wave radar according to the present invention.

[0077] Figure 2 This is a flowchart illustrating the dynamic noise compensation process in the method of this invention.

[0078] Figure 3 This is a flowchart illustrating the adaptive Euclidean clustering algorithm in the method of this invention;

[0079] Figure 4 This is a flowchart illustrating the composite control architecture in the method of this invention;

[0080] Figure 5 This is a schematic diagram illustrating the tracking of a drone using the method of this invention in a scenario of continuous turning.

[0081] Figure 6 This is a diagram illustrating the target tracking trajectory of a drone in a complex scenario using the method of this invention. Detailed Implementation

[0082] To better understand the technical solution of the present invention, the content of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0083] The core of this invention's autonomous UAV tracking method lies in utilizing point cloud data provided by millimeter-wave radar to accurately extract target positions through noise models and data processing algorithms. Combined with a composite control architecture, this enables the UAV to achieve real-time, continuous, and stable autonomous tracking of dynamic targets in complex environments. The system level provides the functional modules required to implement this method.

[0084] Example

[0085] A method for autonomous tracking of unmanned aerial vehicles based on millimeter-wave radar, referring to Figure 1 The method includes the following steps:

[0086] S100: Through the millimeter-wave radar acquisition module, point cloud data of the target UAV and surrounding obstacles are collected. Through the dynamic noise compensation module, a bivariate noise model related to speed and distance is constructed to suppress the problems of radar signal distortion and point cloud blur caused by motion.

[0087] S200 uses an adaptive Euclidean clustering algorithm and an adaptive clustering module to process point cloud data. It calculates point cloud density and relative distance through kernel density estimation, and dynamically adjusts the neighborhood radius according to the distance to the target to accurately extract the target's location information.

[0088] S300 uses a composite control architecture and modules to predict and optimize the trajectory based on the target's position information: the outer loop proportional-integral-derivative controller generates the basic trajectory, and the inner loop model predictive controller optimizes the attitude control to enable the UAV to track stably.

[0089] The S400 and UAV are connected to the ground station for real-time communication, and dynamic targets are tracked in a multi-obstacle environment using the ROS / Gazebo simulation platform.

[0090] The S500 and UAV update the target position information based on the point cloud information of dynamic targets and obstacles acquired and processed in real time by millimeter-wave radar, and continue to predict the tracking trajectory, thereby achieving adaptive tracking strategy adjustment.

[0091] The S600, based on an adjusted tracking strategy, enables UAVs to perform real-time, continuous, and stable tracking of dynamic targets.

[0092] Furthermore, S100, millimeter-wave radar data acquisition and dynamic noise modeling compensation, is the foundation of the entire tracking method. The key is to acquire high-quality point cloud data and suppress dynamic errors to solve the problem of radar signal distortion caused by moving targets.

[0093] The process of dynamic noise compensation, such as Figure 2As shown, firstly, the raw signal is collected. The millimeter-wave radar sensor mounted on the UAV starts working, continuously transmitting and receiving echo signals using the 77GHz high-frequency millimeter wave band. The radar simultaneously collects point cloud data of the tracking target and its surrounding environmental obstacles. These raw radar point cloud data will be affected by the movement of the UAV itself and the relative movement of the target, resulting in signal distortion and point cloud blurring.

[0094] To address the issues of signal distortion and point cloud blurring, this invention constructs a velocity- and distance-dependent bivariate noise model. The core idea of ​​this model is that the noise power is not constant, but depends on two variables: the radial velocity V of the target relative to the radar and the distance R of the target.

[0095] Specifically, the mathematical expression of this model defines the noise power as a function of speed. It grows proportionally and increases linearly with distance R, and this modeling method more accurately characterizes the non-stationary noise characteristics brought about by motion.

[0096] To implement this model and perform real-time noise suppression, this invention employs an adaptive Butterworth filter, whose key parameters, including the cutoff frequency and time constant, are dynamically adjusted based on the noise power calculated in real-time using V and R.

[0097] When the target is detected to be moving at high speed or the distance increases, causing the noise power to increase, the filter will automatically raise the cutoff frequency to more effectively filter out high-frequency noise and retain the useful signal.

[0098] Noise modeling and processing were performed in the local coordinate system of the millimeter-wave radar. However, since the tracking UAV itself is also in motion, its dynamic offset introduces additional errors. Therefore, coordinate transformation is required after noise processing.

[0099] The system calculates and tracks the UAV's own motion offset in real time. Combined with an accurate rotation matrix, it transforms the target point cloud coordinates in the noise-compensated radar coordinate system to a unified world coordinate system. This compensates for the viewpoint deviation caused by the UAV's motion and provides accurate target position data in the global coordinate system for subsequent steps.

[0100] Furthermore, the adaptive Euclidean clustering process in S200 can extract the target location and, based on the noise-compensated point cloud data in the world coordinate system provided by S100, the core task of this step is to accurately separate and locate the point cloud of the target UAV from the complex point cloud background containing environmental obstacles.

[0101] Traditional Euclidean clustering algorithms with fixed parameters struggle to adapt to changes in point cloud density caused by variations in target distance. Therefore, this invention employs an innovative "adaptive Euclidean clustering algorithm" to address the issues of clustering of nearby targets and over-segmentation of distant targets. Figure 3 As shown, the algorithm first uses kernel density estimation technology to calculate the local density around each data point in the radar point cloud data and its relative distance to other data points.

[0102] The key to adaptive characteristics lies in the setting of the neighborhood radius. The neighborhood radius determines the maximum distance threshold at which two points are considered members of the same cluster (or group), and it dynamically adjusts the neighborhood radius used during clustering based on different target distance ranges. When preliminary analysis indicates that the target is within a relatively close range, the system uses a smaller scaling factor to calculate the neighborhood radius. This is because the point cloud is usually denser at close range, and using a smaller neighborhood radius can accurately separate nearby targets from obstacles, avoiding false merging caused by over-segmentation. Conversely, when the target is determined to be within a relatively far range, the system uses a larger scaling factor. This is because the point cloud becomes sparser at far range, and using a larger neighborhood radius ensures that the same target can be correctly aggregated into a single cluster, avoiding target loss or fragmented segmentation.

[0103] After completing the initial clustering, the algorithm also performs cluster merging. It calculates the center position, average density, and other features of each potential cluster, and merges adjacent clusters with a density difference of less than 20% and a distance between their centers of less than a dynamic threshold. This step helps eliminate redundant small clusters caused by minor fluctuations or point cloud segmentation algorithms, ultimately ensuring the identification of a main cluster that accurately represents the target and whose location information is accurate, and outputs the three-dimensional spatial coordinates of its center point as the target's current location information.

[0104] Furthermore, the S300 uses a composite control architecture for trajectory prediction and optimization. Once the target's accurate location information is obtained, the core of this step is to generate control commands for the tracking UAV, enabling it to plan a flight trajectory that can effectively track dynamic targets and maintain its own stability, while optimizing attitude control to accurately follow the trajectory.

[0105] This invention innovatively employs a composite control architecture, such as... Figure 4 As shown, the architecture consists of a cascaded outer-loop PID controller and an inner-loop MPC controller. The outer-loop controller adopts the classic proportional-integral-derivative controller form. This outer-loop PID controller calculates based on the error vector between the target's current position and the tracking UAV's current position. At the same time, the output of the outer-loop PID controller generates a basic, desired tracking trajectory, which mainly focuses on macroscopic path planning and target position tracking.

[0106] The inner loop controller adopts the form of a model predictive controller. The core of the model predictive controller is to build a model that describes the dynamic behavior of the system and use the model to continuously optimize the control input over a period of time in the future.

[0107] In this embodiment, a key six-dimensional state vector is defined, which includes six types of parameters: the three-dimensional position, three-dimensional velocity, attitude angle and corresponding angular velocity of the tracking UAV in the world coordinate system.

[0108] In the state prediction or covariance calculation of MPC, considering the characteristics of actual motion noise, a velocity-related noise covariance matrix is ​​designed, in which it is specifically stipulated that the noise figure of the lateral velocity Vy is twice the noise figure of the longitudinal velocity Vx.

[0109] The objective function of the model predictive controller can simultaneously optimize two key aspects: trajectory tracking accuracy and control stability.

[0110] To achieve multi-objective optimization, a trajectory tracking weight matrix is ​​configured in the function to enhance the priority and accuracy of position tracking; a control input weight matrix is ​​also configured to limit abrupt changes in control commands and ensure smooth and stable flight.

[0111] The basic desired trajectory generated by the outer-loop PID controller is input into the inner-loop MPC controller as a reference command for its tracking. The inner-loop MPC controller then calculates the final optimized attitude control command based on the current state, prediction model, and optimization objective function. These commands are then transmitted to the underlying flight control system of the tracking UAV.

[0112] This composite structure leverages the robustness and simplicity of PID in macroscopic tracking, while taking advantage of MPC's strengths in precise control, constraint handling, and future state prediction, ultimately enabling the tracking drone to stably and accurately follow the generated optimized trajectory.

[0113] Furthermore, for real-time tracking and simulation verification in the S400 multi-obstacle environment, see [link / reference]. Figure 5 and Figure 6 This step involves actually performing the tracking task, emphasizing robustness in multi-obstacle environments and verification capabilities on a simulation platform. The tracking drone establishes a stable communication connection with the ground control station, transmitting status information in real time and receiving necessary monitoring commands.

[0114] like Figure 5 As shown, in the X-axis direction, the actual tracking trajectory, the target trajectory, and the estimated tracking trajectory are highly consistent on the time axis, with the error approaching zero in the later stages. In the Y-axis direction, although there are deviations caused by obstacles, close and continuous tracking can still be maintained subsequently according to the tracking strategy. In the Z-axis direction, although there are slight fluctuations in altitude in the middle stage, the tracking strategy can quickly self-adjust and maintain stability.

[0115] like Figure 6 As shown, overall, it can maintain stable tracking. Although there are obstacles and sharp turns during the tracking process, it can quickly adjust itself and maintain stable tracking under the action of tracking strategy and composite control.

[0116] To effectively verify the performance of the algorithm in complex environments, the entire tracking system was integrated and tested in the mature open-source robot simulation platform ROS / Gazebo.

[0117] In the Gazebo environment, a meticulously constructed 3D scene containing multiple static and dynamic obstacles was created, simulating the target drone performing arbitrary and complex maneuvers. The tracking drone acquires point cloud data of the dynamic target and surrounding obstacles in real time based on a millimeter-wave radar module plugin. The noise modeling compensation module and adaptive clustering module described in steps S100 and S200 perform high-speed processing on the acquired point cloud, continuously outputting updated target position information and identified obstacle position information. Real-time updates are the core of the adaptive mechanism. Based on the latest processing results provided by the millimeter-wave radar in real time, including the dynamic target position and obstacle position, the tracking drone system continuously updates its understanding of the target state and uses the composite control architecture described in step S300 to re-predict the tracking trajectory in the near future based on the latest information. The system continuously monitors the distance to known obstacles in the environment. When any obstacle is detected to be less than the predetermined pathpoint of the tracking drone's trajectory, the system immediately triggers a "trajectory replanning mechanism". This mechanism utilizes obstacle point cloud data in the world coordinate system constructed in step S100, and combines the optimized prediction capability of the model prediction controller in step S300 or calls an independent path planner to quickly calculate a new avoidance path and smoothly transition to the new path to ensure flight safety.

[0118] This invention significantly reduces the overall system latency from radar perception to control response through efficient algorithm design and data processing. This enables the system to effectively track target UAVs performing rapid maneuvers such as continuous turns, maintaining continuous tracking capability even in multi-obstacle simulation environments. During simulation, the system continuously records key performance indicators such as the accuracy error of static target localization and the position tracking error when tracking dynamic targets to verify the effectiveness of the method.

[0119] Furthermore, the S500 adaptive tracking strategy adjustment and real-time continuous stable tracking form the closed-loop link in the tracking cycle, demonstrating the method's adaptability. Based on the latest information perceived, processed, and fused in real-time by the millimeter-wave radar in the S400, including the latest precise location of dynamic targets, the distribution of identified environmental obstacles and their relative relationship with preset safety zones, the tracking UAV system continuously performs situational assessments and strategy adjustments. The adaptive tracking strategy adjustment fine-tunes the risk tolerance and conservatism of its motion control and path planning based on the target's real-time motion state, relative distance, and environmental complexity. More importantly, through mechanisms such as rolling optimization of MPC in the S300, real-time trajectory replanning triggered by obstacle collision avoidance in the S400, online dynamic adjustment of noise model parameters in the S100, and adaptive selection of neighborhood radius coefficients in the S200 clustering algorithm, the system achieves comprehensive and multi-layered adaptive capabilities. This continuous adaptive adjustment ensures that the entire tracking strategy is always optimized for the current scenario.

[0120] Ultimately, based on the continuously optimized tracking strategy described above, the tracking drone, through precise attitude control commands output by the composite control architecture, drives its flight platform to follow the target's trajectory in real time, continuously, and stably. Even when the target is maneuvering highly or environmental obstacles are dynamically changing, the system can maintain the stability of the tracking lock to the maximum extent through rapid perception, accurate calculation, and adaptive adjustment.

Claims

1. An autonomous tracking method for unmanned aerial vehicles based on millimeter-wave radar, characterized in that, The method includes the following steps: S100: Through the millimeter-wave radar acquisition module, point cloud data of the target UAV and surrounding obstacles are collected. Through the dynamic noise compensation module, a bivariate noise model related to speed and distance is constructed to suppress the problems of radar signal distortion and point cloud blur caused by motion. S200 uses an adaptive Euclidean clustering algorithm and an adaptive clustering module to process point cloud data. It calculates point cloud density and relative distance through kernel density estimation, and dynamically adjusts the neighborhood radius according to the distance to the target to accurately extract the target's location information. S300 uses a composite control architecture and modules to predict and optimize the trajectory based on the target's position information: the outer loop proportional-integral-derivative controller generates the basic trajectory, and the inner loop model predictive controller optimizes the attitude control to enable the UAV to track stably. The S400 and UAV are connected to the ground station for real-time communication, and dynamic targets are tracked in a multi-obstacle environment using the ROS / Gazebo simulation platform. The S500 and UAV update the target position information based on the point cloud information of dynamic targets and obstacles acquired and processed in real time by millimeter-wave radar, and continue to predict the tracking trajectory, thereby achieving adaptive tracking strategy adjustment. The S600, based on an adjusted tracking strategy, enables UAVs to perform real-time, continuous, and stable tracking of dynamic targets.

2. The method as described in claim 1, characterized in that, The construction of the speed-distance related bivariate noise model described in S100 also includes: The intermediate frequency is extracted by Fourier transform, and the target distance and radial velocity are calculated. Constructing the noise power function The noise power is proportional to the 1.5th power of the speed and proportional to the distance. Through the parameters of the adaptive Butterworth filter Dynamically adjust the cutoff frequency and time constant; The system calculates the spatial offset of the UAV in real time, compensates for dynamic offset errors through a coordinate transformation model, and transforms the radar coordinates to the world coordinate system using a rotation matrix.

3. The method as described in claim 2, characterized in that: The noise power function In the formula, It is the noise variance value determined by the target velocity v and range R in millimeter-wave radar measurements; v represents the target's relative radial velocity; R represents the straight-line distance between the target and the radar; The parameters of the Tevos filter ; In the formula, The dynamic cutoff frequency representing the target velocity v and distance R; This refers to the system sampling frequency; This represents the noise standard deviation function related to speed and distance. The formula for calculating the spatial offset is: ； In the formula, This indicates the target's position relative to the radar sensor itself; Indicates the absolute position of the target in the global reference frame; Let be the rotation matrix, representing the rotation angle of the radar coordinate system relative to the world coordinate system. ; These represent the radar's position offset in the x and y directions in the world coordinate system, respectively.

4. The method as described in claim 1, characterized in that, S200 further includes the use of adaptive Euclidean clustering algorithm to process point cloud data: The local density and relative distance of each data point are calculated using the Gaussian kernel function; When the target is relatively close, set the neighborhood radius to 0.3 times the maximum distance threshold; When the target is at a relatively far distance, set the neighborhood radius to 0.6 times the maximum distance threshold; Perform a merge operation on clusters with a density difference of less than 20% and a distance of less than the dynamic threshold of 60%.

5. The method as described in claim 4, characterized in that: The local density of each data point is calculated using a Gaussian kernel function. and relative distance , The calculation formula is: ； In the formula, Point The local density estimate; Point The minimum distance to a higher density point; Represents the Euclidean distance between two points, where This indicates the point cloud coordinates of the target detected by the radar; h represents the bandwidth parameter of the kernel function; The cluster merging formula is as follows: ； In the formula, d represents the effective radius of the cluster; , These are the dynamic distance thresholds for the two clusters, respectively; Indicates the density difference threshold; This represents the core density of the first and second clusters.

6. The method as described in claim 1, characterized in that, The use of a composite control architecture for trajectory prediction and optimization as described in S300 also includes: Construct a six-degree-of-freedom state equation based on the target location information: ； In the formula, The position of the target in the inertial coordinate system. Linear velocity, Yaw angle Angular velocity; Design the velocity-dependent noise covariance matrix, where the lateral velocity noise figure is twice that of the longitudinal velocity. The outer loop proportional-integral-derivative controller is configured with a proportional coefficient of 1.2, an integral coefficient of 0.05, and a derivative coefficient of 0.

3. The inner-loop model predicts the control objective function, simultaneously optimizing trajectory tracking accuracy and control stability.

7. The method as described in claim 6, characterized in that, The inner-loop model predicts the control objective function, which also includes: The prediction time domain is set to 5 control periods; Configure a trajectory tracking weight matrix to enhance position tracking permissions; Configure the control input weight matrix to limit sudden changes in control commands; The objective function expression is: ) ； It is the noise covariance matrix of the time-varying process; () is a constructor for diagonal matrices; , These represent the target linear velocity and the target angular velocity, respectively.

8. The method as described in claim 1, characterized in that, The real-time tracking in a multi-obstacle environment on the ROS / Gazebo simulation platform described in S400 also includes: When the distance to an obstacle is detected to be less than the safety threshold, the trajectory replanning mechanism is triggered; Obstacles and target point clouds are separated in real time using an adaptive Euclidean clustering algorithm; When response latency is reduced, continuous turning tracking of dynamic targets is supported.

9. An autonomous tracking system for unmanned aerial vehicles based on millimeter-wave radar, characterized in that, The system is used to implement the UAV autonomous tracking method based on millimeter-wave radar as described in any one of claims 1-9, and the system comprises: Millimeter-wave radar acquisition module, used to acquire point cloud data of target drones and obstacles; The dynamic noise compensation module is used to construct a bivariate noise model that is related to speed and distance. An adaptive clustering module is used to perform distance-dependent neighborhood radius adjustments; The composite control module is used to realize the cascade optimization of extended Kalman filtering and model predictive control. The trajectory prediction unit of this module uses the extended Kalman filtering algorithm to solve the target's six-degree-of-freedom pose state, and performs local linearization of the nonlinear motion equations through the Jacobian matrix to estimate the target's position, velocity and turning angle in three-dimensional space in real time. The attitude optimization unit deploys a model predictive controller and generates the optimal attitude control command through a look-ahead rolling optimization mechanism: a multi-objective function containing trajectory tracking error and actuator energy consumption is constructed in the prediction time domain, and the optimal attitude angular rate command set is solved through quadratic programming; The outer loop proportional-integral-derivative controller generates basic propulsion commands based on trajectory prediction results. Its output is coupled with the torque compensation commands of the inner loop model predictive controller to achieve coordinated optimization of heading accuracy and attitude stability. The simulation verification module is used to test tracking performance on the ROS / Gazebo platform. The tracking and decision-making module includes a strategy update module and a closed-loop tracking module. The strategy update module monitors the dynamic changes of environmental variables and target movement characteristics in real time, and dynamically adjusts system parameters based on the monitoring results. The closed-loop tracking module drives the brushless motor actuator according to optimized control commands and adopts a feedforward-feedback composite control structure to eliminate the influence of aerodynamic disturbances. The feedforward branch predicts the inertial compensation amount based on the target kinematic model, and the feedback branch extracts the acceleration signal through the differential tracker for phase lead correction, so as to realize the smooth transition of control commands under high-speed turning conditions. Combined with the target position feedback collected in real time by millimeter-wave radar, a closed-loop tracking control system is formed to ensure that the dynamic target is always in the effective detection field of the sensor.

10. The unmanned aerial vehicle (UAV) autonomous tracking system as described in claim 9, characterized in that: The adaptive clustering module includes a target separation module, which calculates the local density distribution characteristics of point cloud data based on the kernel density estimation algorithm, identifies the initial cluster core region through peak detection, and selects effective cluster center points by combining the relative distance judgment criterion. A multi-scale clustering strategy is adopted to dynamically configure the neighborhood search radius. In the near-field region where the target distance is less than 5 meters, a compact radius parameter is used to prevent the target point cloud from sticking together. In the far-field region where the distance is greater than 20 meters, the radius parameter is expanded to compensate for the point cloud density attenuation and avoid over-segmentation. It also introduces density similarity constraint rules to automatically merge discrete clusters that are spatially adjacent and have matching density distribution characteristics, which significantly improves the complete recognition rate of broken targets behind occluders.

Images